GitHub - crowd-dataset/crowd: This repository contains code that extracts YouTube videos based on a mapping.csv file and performs object detection using YOLOv11. The primary objective of this work is to evaluate pedestrian behaviour in a cross-country or cross-cultural context using freely available YouTube videos.

City Road Observations With Dashcams (CROWD)

This repository contains code that extracts YouTube videos based on a mapping.csv file and performs object detection using YOLOv11. The primary objective of this work is to evaluate pedestrian behaviour in a cross-country or cross-cultural context using freely available YouTube videos.

This study presents a comprehensive cross-cultural evaluation of pedestrian behaviour during road crossings, examining variations between developed and developing states worldwide. As urban landscapes evolve and autonomous vehicles (AVs) become integral to future transportation, understanding pedestrian behaviour becomes paramount for ensuring safe interactions between humans and AVs. Through an extensive review of global pedestrian studies, we analyse key factors influencing crossing behaviour, such as cultural norms, socioeconomic factors, infrastructure development, and regulatory frameworks. Our findings reveal distinct patterns in pedestrian conduct across different regions. Developed states generally exhibit more structured and rule-oriented crossing behaviours, influenced by established traffic regulations and advanced infrastructure. In contrast, developing states often witness a higher degree of informal and adaptive behaviour due to limited infrastructure and diverse cultural influences. These insights underscore the necessity for AVs to adapt to diverse pedestrian behaviour on a global scale, emphasising the importance of incorporating cultural nuances into AV programming and decision-making algorithms. As the integration of AVs into urban environments accelerates, this research contributes valuable insights for enhancing the safety and efficiency of autonomous transportation systems. By recognising and accommodating diverse pedestrian behaviours, AVs can navigate complex and dynamic urban settings, ensuring a harmonious coexistence with human road users across the globe.

The dataset is available on kaggle. The dataset shall soon be made available on a permanent FAIR storage.

Dataset overview

Footage: 35,569.5 hours (1,482.1 days)
Videos: 80,308 unique videos split into 104,165 segments
Localities (cities): 10,237
Countries and territories: 238
Continents: 6
YOLO detections: 86,123 processed segments (one detection CSV each)
Detected objects (unique tracked objects, confidence ≥ 0.7): 31,050,010 persons, 81,888,691 cars, 2,434,215 bicycles, 10,320,404 motorcycles, 3,371,096 buses, 13,176,493 trucks
Last updated: 2026-10-21 17:43 UTC

All 238 countries and territories

#	Country	Cities	Videos	Footage (h)	Analysed (h)	Persons	Cars	Bicycles	Motorcycles	Buses	Trucks
1	🇺🇸 United States	2,299	20,883	10,373.8	7,805.0	4,272,701	23,685,912	271,118	142,810	463,888	3,766,859
2	🇦🇺 Australia	212	4,112	1,725.5	1,161.3	575,432	3,841,888	38,915	39,827	87,817	631,717
3	🇬🇧 United Kingdom	358	3,184	1,691.6	1,496.8	2,669,605	3,286,323	258,876	153,027	259,935	557,811
4	🇷🇺 Russia	212	3,396	1,515.1	1,272.0	739,160	3,488,118	15,963	6,876	153,569	448,842
5	🇵🇭 Philippines	399	3,583	1,311.0	1,302.0	2,280,616	2,263,741	71,888	1,423,298	303,676	910,602
6	🇨🇦 Canada	229	3,545	1,234.7	988.3	746,694	2,850,553	68,576	19,519	55,088	371,391
7	🇻🇳 Vietnam	79	3,079	1,196.4	1,110.8	2,841,773	1,918,728	89,613	2,791,899	124,640	295,741
8	🇮🇩 Indonesia	176	3,159	1,019.6	969.2	1,868,154	1,870,317	24,313	1,816,459	108,349	471,997
9	🇰🇷 South Korea	84	853	822.4	792.3	350,283	1,627,830	20,330	52,974	103,482	176,255
10	🇫🇷 France	360	1,893	746.5	715.5	811,503	1,741,161	100,263	122,825	46,589	193,698
11	🇨🇳 China	149	1,229	693.0	623.6	477,193	1,757,527	50,356	190,159	101,437	205,625
12	🇩🇪 Germany	241	1,385	654.8	602.4	529,924	1,684,445	183,442	34,832	59,311	204,147
13	🇯🇵 Japan	233	1,753	610.8	552.0	344,686	1,055,513	49,372	21,845	77,232	280,424
14	🇧🇷 Brazil	130	1,453	604.4	531.9	519,071	1,513,479	33,527	149,938	54,774	192,895
15	🇳🇱 Netherlands	232	1,448	548.0	432.9	487,941	991,001	339,799	54,752	31,801	113,489
16	🇮🇳 India	300	1,509	484.0	450.1	1,188,926	1,005,608	63,801	686,912	124,460	335,983
17	🇺🇦 Ukraine	94	1,083	442.8	360.6	264,908	1,196,004	7,847	4,225	59,354	177,315
18	🇹🇭 Thailand	50	794	386.4	368.0	446,399	969,153	8,569	340,642	48,655	219,253
19	🇲🇲 Myanmar	19	536	371.5	368.9	377,334	1,481,396	43,638	40,982	98,845	251,629
20	🇲🇾 Malaysia	66	631	368.2	279.2	139,060	821,081	1,749	83,469	17,702	77,039
21	🇦🇪 United Arab Emirates	10	865	360.3	303.8	101,554	780,794	3,950	15,679	42,356	109,840
22	🇵🇱 Poland	283	892	350.9	322.9	211,161	860,875	20,381	7,004	32,348	82,692
23	🇮🇷 Iran	108	983	318.4	243.3	157,668	1,165,662	3,260	29,859	13,082	104,765
24	🇸🇪 Sweden	167	988	273.2	264.8	126,745	430,696	21,512	2,783	18,496	50,656
25	🇹🇷 Türkiye	87	570	272.1	227.0	268,694	842,597	4,209	36,123	55,842	130,713
26	🇮🇹 Italy	168	673	263.1	261.1	265,315	817,830	23,685	100,055	17,599	68,722
27	🇦🇷 Argentina	64	412	257.3	251.6	339,200	819,300	33,589	41,484	29,416	97,763
28	🇩🇰 Denmark	22	565	256.3	250.3	369,220	719,152	194,235	8,296	20,176	60,376
29	🇷🇴 Romania	91	633	242.6	239.9	192,777	1,091,080	6,707	5,071	27,553	98,506
30	🇫🇮 Finland	42	383	239.3	231.8	126,585	442,436	14,166	2,367	21,442	54,634
31	🇪🇸 Spain	251	659	232.6	220.0	265,869	579,582	15,145	61,646	22,167	63,447
32	🇰🇭 Cambodia	7	427	222.3	222.3	581,388	540,095	13,515	597,642	38,795	179,386
33	🇨🇿 Czechia	84	378	216.0	214.6	211,202	529,498	5,602	5,071	23,140	55,113
34	🇦🇹 Austria	45	850	202.9	199.8	125,373	502,771	17,385	14,242	22,985	82,568
35	🇺🇬 Uganda	19	516	193.1	40.1	210,110	97,690	5,333	116,814	9,743	31,876
36	🇭🇺 Hungary	34	473	173.3	171.4	302,411	1,021,085	14,872	8,332	50,805	124,089
37	🇵🇰 Pakistan	38	425	167.5	141.9	441,809	466,633	16,051	232,921	27,715	116,824
38	🇧🇬 Bulgaria	89	643	162.6	162.6	83,794	508,664	1,792	1,203	18,043	48,284
39	🇸🇬 Singapore	1	414	162.0	152.6	81,730	342,138	5,203	20,569	33,441	60,162
40	🇳🇿 New Zealand	160	605	161.2	141.3	76,612	365,330	4,575	3,368	13,541	53,467
41	🇻🇪 Venezuela	26	620	154.1	129.6	92,844	331,841	4,908	35,984	7,127	68,507
42	🇲🇽 Mexico	73	310	145.3	122.0	125,401	377,330	6,052	21,125	13,579	60,564
43	🇮🇪 Ireland	55	499	143.5	138.5	108,118	310,653	10,483	2,209	13,984	41,030
44	🇧🇾 Belarus	83	281	139.2	78.5	44,011	171,135	1,579	398	9,951	24,933
45	🇳🇴 Norway	68	421	139.1	136.9	85,184	283,525	7,453	1,997	13,853	38,237
46	🇿🇦 South Africa	59	481	138.7	129.4	90,590	325,224	1,701	4,647	12,966	60,022
47	🇸🇾 Syria	13	316	124.6	122.3	216,348	507,868	6,463	16,996	17,203	83,736
48	🇪🇬 Egypt	20	270	107.5	52.0	71,736	235,831	634	8,241	11,975	31,172
49	🇧🇩 Bangladesh	29	298	107.3	107.2	390,478	113,971	74,318	72,325	40,702	85,738
50	🇭🇰 Hong Kong	11	228	107.3	59.1	100,646	99,158	1,443	2,991	19,375	19,435
51	🇵🇹 Portugal	73	226	104.4	104.4	87,578	287,432	2,143	7,460	7,731	31,681
52	🇨🇭 Switzerland	94	302	100.2	92.5	103,933	188,046	21,150	13,360	9,502	24,085
53	🇬🇲 Gambia	29	361	95.4	95.4	226,529	285,640	14,671	20,858	8,740	64,562
54	🇳🇬 Nigeria	35	230	92.0	56.9	216,399	176,420	1,810	32,108	13,929	45,474
55	🇸🇱 Sierra Leone	47	184	79.3	79.1	170,227	48,806	1,877	49,358	4,591	23,204
56	🇯🇴 Jordan	13	250	78.3	78.3	46,155	311,691	181	983	7,319	39,417
57	🇨🇱 Chile	43	160	71.9	23.5	22,163	61,334	775	1,113	2,706	13,174
58	🇲🇳 Mongolia	4	130	71.4	71.4	88,118	304,976	1,881	1,654	9,124	25,710
59	🇭🇷 Croatia	88	281	71.3	71.2	52,249	228,297	4,591	4,309	5,509	19,685
60	🇬🇷 Greece	40	196	66.9	66.0	73,714	265,121	2,806	25,708	6,316	25,638
61	🇨🇾 Cyprus	17	252	66.4	66.2	26,967	202,441	1,081	5,007	4,165	28,218
62	🇪🇪 Estonia	6	302	65.3	65.0	33,674	167,480	1,206	418	8,464	16,981
63	🇧🇪 Belgium	61	191	63.0	62.3	65,468	167,638	14,859	3,856	7,204	23,728
64	🇧🇦 Bosnia and Herzegovina	60	291	62.0	48.8	50,620	157,116	865	576	4,488	13,869
65	🇨🇴 Colombia	36	131	61.7	41.4	86,062	89,103	5,878	39,551	6,101	17,256
66	🇺🇿 Uzbekistan	12	178	53.6	34.6	26,975	209,918	434	194	4,187	17,563
67	🇵🇪 Peru	29	94	53.1	53.1	84,003	147,775	5,452	12,966	9,566	30,312
68	🇹🇼 Taiwan	38	138	50.7	43.7	49,007	115,619	2,523	44,935	4,493	12,517
69	🇷🇸 Serbia	37	146	47.9	46.5	42,935	171,668	1,152	1,224	5,929	14,641
70	🇱🇰 Sri Lanka	15	128	47.7	47.3	70,859	99,910	1,065	25,398	17,602	40,207
71	🇲🇦 Morocco	35	135	47.6	46.9	85,390	175,883	3,788	17,447	5,038	21,670
72	🇦🇿 Azerbaijan	8	112	45.5	45.5	44,791	206,195	690	1,127	6,072	21,374
73	🇶🇦 Qatar	5	98	44.3	44.3	10,603	109,498	434	1,281	5,308	16,307
74	🇪🇹 Ethiopia	8	77	44.1	42.2	104,306	104,629	540	1,978	9,657	31,967
75	🇸🇦 Saudi Arabia	11	110	43.6	32.4	5,107	142,192	118	248	2,833	17,976
76	🇬🇪 Georgia	11	117	42.4	40.3	31,146	170,922	246	1,164	6,716	21,141
77	🇳🇵 Nepal	4	120	38.2	38.1	212,044	71,559	6,035	126,705	10,857	19,965
78	🇱🇻 Latvia	20	123	38.0	34.5	24,270	91,567	1,611	716	3,964	11,579
79	🇱🇧 Lebanon	60	154	35.2	35.2	22,732	146,170	182	7,506	2,695	17,820
80	🇨🇺 Cuba	25	168	34.6	9.9	24,271	5,125	2,110	3,773	719	2,701
81	🇱🇹 Lithuania	10	84	33.5	33.5	31,866	98,334	2,608	715	3,317	7,486
82	🇵🇷 Puerto Rico	60	159	33.1	10.1	3,778	30,191	82	125	280	4,859
83	🇲🇼 Malawi	51	113	32.4	29.9	67,538	30,370	9,781	5,080	1,788	8,925
84	🇰🇿 Kazakhstan	26	113	30.9	30.4	21,562	115,902	260	201	4,020	12,666
85	🇦🇲 Armenia	10	105	30.8	8.2	6,209	40,203	35	64	1,393	5,562
86	🇰🇪 Kenya	58	115	27.0	19.9	62,585	46,029	1,155	12,133	4,453	12,793
87	🇪🇨 Ecuador	15	56	25.4	22.8	27,801	68,164	685	4,730	3,522	12,731
88	🇺🇾 Uruguay	23	90	25.3	25.3	31,932	90,573	2,748	5,257	1,626	12,618
89	🇲🇰 North Macedonia	21	72	24.4	24.4	25,462	86,433	1,764	1,066	2,506	7,806
90	🇲🇩 Moldova	14	107	24.2	24.2	25,864	102,739	302	217	5,075	12,436
91	🇱🇦 Laos	10	60	22.9	22.9	23,497	49,112	526	22,212	1,271	19,464
92	🇧🇴 Bolivia	20	55	21.5	19.3	26,213	42,478	628	4,934	2,728	13,037
93	🇲🇨 Monaco	1	45	20.3	19.5	35,821	30,718	962	12,489	1,542	4,295
94	🇮🇱 Israel	18	68	19.9	19.0	14,470	65,564	2,335	3,088	2,706	5,634
95	🇽🇰 Kosovo	31	93	15.4	13.5	12,872	51,774	508	510	1,264	7,371
96	🇬🇹 Guatemala	9	23	15.3	12.6	13,328	32,754	527	7,132	1,411	9,393
97	🇦🇱 Albania	23	62	14.8	14.8	17,949	45,917	1,604	1,215	1,389	5,350
98	🇩🇿 Algeria	8	35	14.5	14.5	24,662	61,115	133	1,301	1,363	6,805
99	🇨🇬 Congo (Congo-Brazzaville)	5	38	13.9	11.8	26,483	22,471	229	3,047	1,058	5,202
100	🇮🇸 Iceland	14	40	13.2	12.8	7,211	28,191	474	68	952	4,491
101	🇱🇾 Libya	3	39	13.2	13.2	4,891	54,787	79	115	299	6,981
102	🇹🇲 Turkmenistan	1	85	11.9	11.9	3,731	22,595	47	6	812	2,774
103	🇬🇭 Ghana	10	39	11.4	11.4	22,184	27,168	670	4,453	2,213	8,526
104	🇮🇶 Iraq	10	40	11.4	9.6	4,150	33,761	120	477	523	4,834
105	🇴🇲 Oman	10	32	10.9	10.9	2,967	35,947	156	60	742	3,513
106	🇸🇻 El Salvador	6	13	10.8	5.3	7,666	13,960	246	2,318	687	3,575
107	🇭🇳 Honduras	8	23	10.3	7.2	7,734	20,112	380	3,007	1,061	4,816
108	🇲🇪 Montenegro	22	65	10.2	10.2	8,812	38,620	172	751	963	4,063
109	🇹🇿 Tanzania	14	51	9.5	5.2	12,790	7,858	657	3,490	1,331	3,023
110	🇵🇦 Panama	7	22	9.4	3.0	1,682	10,262	55	117	183	2,349
111	🇬🇾 Guyana	6	20	9.0	8.2	8,363	22,593	662	826	1,087	6,297
112	🇸🇮 Slovenia	26	49	8.9	8.5	9,937	25,347	1,570	356	897	2,050
113	🇧🇭 Bahrain	11	42	8.8	8.8	2,007	24,546	182	228	746	3,611
114	🇰🇬 Kyrgyzstan	7	40	8.4	8.4	6,645	30,471	161	69	1,493	4,913
115	🇯🇲 Jamaica	10	30	8.3	8.3	8,835	19,261	220	346	1,458	3,469
116	🇵🇸 Palestine	10	31	8.3	8.3	5,348	26,384	95	244	807	3,880
117	🇸🇰 Slovakia	13	27	8.0	7.4	5,559	22,126	292	193	879	2,484
118	🇬🇮 Gibraltar	1	21	7.9	7.9	7,384	14,720	384	3,883	630	2,258
119	🇲🇹 Malta	31	56	7.6	7.6	3,461	23,323	49	533	757	3,788
120	🇨🇮 Côte d'Ivoire	2	19	7.6	7.6	13,309	29,310	176	1,474	924	4,167
121	🇬🇳 Guinea	6	28	7.5	6.3	18,434	15,992	146	6,364	451	3,611
122	🇱🇺 Luxembourg	12	25	7.0	7.0	8,121	15,294	468	316	914	1,638
123	🇰🇼 Kuwait	2	21	6.9	6.9	2,664	23,894	27	74	1,067	2,654
124	🇱🇷 Liberia	11	43	6.8	6.8	26,168	16,747	120	9,633	970	5,277
125	🇸🇷 Suriname	3	19	6.7	5.4	2,262	16,386	130	506	521	2,645
126	🇿🇼 Zimbabwe	4	26	6.3	6.3	16,477	25,223	110	148	1,292	6,010
127	🇦🇬 Antigua and Barbuda	9	21	6.1	5.7	4,292	10,555	43	25	270	1,908
128	🇵🇾 Paraguay	8	18	5.9	5.9	6,004	19,678	299	3,007	470	4,024
129	🇻🇨 Saint Vincent and the Grenadines	11	25	5.7	5.7	3,913	7,296	23	30	546	1,835
130	🇧🇯 Benin	10	16	5.6	5.6	19,911	10,339	325	12,657	291	2,243
131	🇦🇫 Afghanistan	8	26	5.3	5.3	10,698	15,305	996	694	858	3,172
132	🇨🇫 Central African Republic	3	18	5.3	5.3	33,260	7,691	373	14,433	187	2,431
133	🇱🇮 Liechtenstein	16	59	5.2	5.2	1,156	8,408	176	210	434	1,474
134	🇹🇳 Tunisia	5	15	5.1	5.1	9,163	23,032	173	672	506	3,303
135	🇹🇹 Trinidad and Tobago	3	5	4.6	4.6	1,657	8,274	3	11	228	1,770
136	🇪🇷 Eritrea	11	27	4.6	4.6	14,201	7,216	3,305	138	286	1,719
137	🇦🇩 Andorra	8	45	4.6	4.6	4,008	12,742	84	1,183	644	2,748
138	🇳🇮 Nicaragua	8	17	4.6	4.3	5,973	9,241	589	2,642	460	3,022
139	🇬🇺 Guam	10	41	4.5	4.5	510	14,313	14	33	154	3,039
140	🇲🇴 Macau	2	15	4.5	4.5	3,030	5,239	47	653	889	951
141	🇰🇵 North Korea	1	19	4.4	4.4	14,070	3,924	1,823	102	932	771
142	🇨🇷 Costa Rica	10	19	4.3	4.3	4,316	10,203	253	765	448	2,157
143	🇲🇬 Madagascar	3	11	4.3	4.3	19,088	6,412	2,000	2,699	1,016	3,360
144	🇵🇫 French Polynesia	11	17	4.3	4.3	2,002	10,477	210	975	196	2,538
145	🇸🇳 Senegal	6	16	4.3	4.3	11,200	12,515	159	2,982	463	2,408
146	🇰🇾 Cayman Islands	5	17	4.2	4.2	759	10,150	84	36	319	2,019
147	🇬🇵 Guadeloupe	11	30	4.1	3.9	1,311	12,588	38	117	259	1,792
148	🇦🇴 Angola	2	17	4.1	4.1	11,170	13,849	70	951	1,041	3,558
149	🇿🇲 Zambia	4	13	3.7	3.6	7,055	6,968	408	96	527	2,244
150	🇷🇼 Rwanda	7	14	3.5	2.9	10,556	4,987	638	4,322	326	1,543
151	🇬🇱 Greenland	3	11	3.5	2.8	1,349	2,888	39	2	162	772
152	🇸🇲 San Marino	4	13	3.5	3.5	1,032	8,866	11	245	155	890
153	🇨🇩 Congo (Democratic Republic)	3	16	3.3	3.2	9,462	7,610	66	2,604	289	1,534
154	🇨🇲 Cameroon	9	16	3.2	3.2	9,836	11,413	42	2,935	200	1,714
155	🇫🇯 Fiji	3	9	3.2	3.0	2,017	6,661	6	23	428	1,497
156	🇬🇬 Guernsey	8	20	3.1	3.1	1,510	6,086	113	166	186	974
157	🇱🇨 Saint Lucia	7	20	3.1	2.8	2,447	6,450	16	42	532	1,784
158	🇬🇦 Gabon	6	17	3.0	3.0	4,690	6,478	33	81	182	1,924
159	🇭🇹 Haiti	1	6	2.9	2.9	8,084	4,643	45	1,580	425	2,678
160	🇩🇴 Dominican Republic	7	9	2.9	2.9	3,782	6,949	172	1,872	298	1,352
161	🇸🇴 Somalia	4	15	2.8	2.8	7,097	4,498	21	170	864	2,487
162	🇲🇺 Mauritius	7	13	2.8	2.8	3,494	7,076	76	1,004	455	1,505
163	🇧🇲 Bermuda	4	5	2.7	2.7	658	3,199	24	737	189	700
164	🇳🇨 New Caledonia	1	5	2.6	2.6	675	5,424	49	42	103	1,377
165	🇾🇪 Yemen	3	12	2.5	2.5	7,388	9,522	96	2,496	347	2,781
166	🇰🇲 Comoros	3	5	2.5	2.5	3,578	4,854	55	1,564	87	770
167	🇧🇼 Botswana	4	9	2.3	2.3	1,589	4,779	5	10	174	1,077
168	🇯🇪 Jersey	10	18	2.2	2.2	889	4,543	101	67	148	956
169	🇻🇬 British Virgin Islands	7	17	2.2	2.2	680	5,237	3	49	101	1,193
170	🇰🇳 Saint Kitts and Nevis	3	9	2.1	2.1	1,762	3,245	26	25	174	743
171	🇼🇸 Samoa	1	7	2.0	1.2	392	2,715	0	0	224	588
172	🇷🇪 Réunion	4	6	2.0	2.0	1,205	4,903	31	89	107	632
173	🇹🇯 Tajikistan	2	9	1.9	1.9	2,537	8,973	30	72	355	855
174	🇸🇸 South Sudan	1	5	1.9	1.9	6,672	3,931	76	2,770	166	1,453
175	🇳🇦 Namibia	4	10	1.8	1.8	1,457	3,452	18	11	95	876
176	🇫🇴 Faroe Islands	21	36	1.8	1.8	287	2,560	12	4	124	456
177	🇬🇫 French Guiana	1	3	1.7	1.7	831	5,927	59	387	71	1,292
178	🇮🇲 Isle of Man	8	19	1.6	1.6	1,351	6,939	24	645	209	1,150
179	🇧🇹 Bhutan	2	6	1.6	1.2	1,646	3,041	14	40	91	671
180	🇩🇯 Djibouti	3	5	1.5	1.5	746	1,537	5	116	243	1,306
181	🇱🇸 Lesotho	10	15	1.5	1.5	1,730	2,222	5	11	95	619
182	🇬🇩 Grenada	4	8	1.4	1.4	577	1,762	8	43	97	611
183	🇲🇫 Saint Martin	6	9	1.3	1.3	461	4,156	13	130	100	1,155
184	🇧🇮 Burundi	6	9	1.3	1.3	5,541	2,270	433	313	173	492
185	🇧🇧 Barbados	3	4	1.3	1.3	1,506	2,706	41	7	177	500
186	🇲🇿 Mozambique	5	8	1.3	1.3	1,953	1,850	34	149	119	462
187	🇲🇷 Mauritania	1	3	1.3	1.3	1,854	3,898	74	116	38	937
188	🇪🇭 Western Sahara	3	4	1.3	1.3	726	2,450	34	100	56	564
189	🇹🇬 Togo	6	11	1.2	1.2	2,940	1,458	59	1,696	55	370
190	🇵🇬 Papua New Guinea	2	6	1.2	1.2	1,095	1,656	0	0	216	443
191	🇲🇸 Montserrat	1	3	1.2	1.2	4,465	1,127	267	4,014	383	835
192	🇦🇸 American Samoa	9	13	1.1	1.1	266	2,385	5	1	127	1,119
193	🇧🇸 Bahamas	2	5	1.1	0.9	257	1,383	21	22	55	248
194	🇹🇨 Turks and Caicos Islands	2	4	1.0	1.0	102	1,469	0	2	23	212
195	🇻🇺 Vanuatu	1	5	1.0	1.0	1,038	1,438	14	3	156	601
196	🇹🇩 Chad	1	4	1.0	1.0	3,975	2,890	74	1,962	52	748
197	🇻🇮 United States Virgin Islands	3	5	1.0	1.0	399	2,765	0	7	72	779
198	🇲🇱 Mali	1	2	0.9	0.9	4,487	2,330	29	3,053	173	696
199	🇳🇷 Nauru	11	18	0.9	0.9	329	1,067	4	341	49	571
200	🇧🇫 Burkina Faso	1	5	0.8	0.8	2,932	2,972	321	2,752	55	460
201	🇸🇽 Sint Maarten	4	5	0.8	0.8	570	2,892	7	42	67	613
202	🇦🇼 Aruba	2	3	0.8	0.8	312	2,342	6	9	44	272
203	🇨🇼 Curaçao	1	3	0.8	0.8	32	1,635	0	0	15	195
204	🇦🇽 Åland Islands	1	1	0.8	0.8	110	617	26	4	11	98
205	🇲🇭 Marshall Islands	3	3	0.8	0.8	114	333	0	0	13	95
206	🇧🇳 Brunei	3	4	0.8	0.8	74	1,391	1	4	27	145
207	🇸🇯 Svalbard and Jan Mayen	1	3	0.7	0.7	60	497	5	5	39	169
208	🇧🇿 Belize	3	5	0.7	0.7	404	1,571	61	91	49	519
209	🇧🇶 Bonaire, Sint Eustatius and Saba	1	1	0.7	0.7	236	1,737	21	82	27	695
210	🇬🇶 Equatorial Guinea	1	5	0.7	0.7	805	2,063	12	9	51	373
211	🇫🇲 Micronesia	3	3	0.6	0.6	164	661	0	0	6	148
212	🇳🇪 Niger	1	2	0.6	0.6	2,224	1,023	66	266	7	126
213	🇵🇼 Palau	4	5	0.6	0.6	23	521	3	3	9	126
214	🇦🇮 Anguilla	4	5	0.6	0.6	57	769	0	1	15	178
215	🇸🇩 Sudan	2	6	0.6	0.6	1,080	3,294	8	181	140	654
216	🇧🇱 Saint Barthélemy	2	3	0.5	0.5	214	1,343	0	222	22	510
217	🇲🇵 Northern Mariana Islands	2	5	0.5	0.5	109	1,729	3	0	20	300
218	🇲🇻 Maldives	2	4	0.5	0.5	1,597	751	57	1,488	36	187
219	🇵🇲 Saint Pierre and Miquelon	1	1	0.4	0.4	76	1,119	2	5	18	372
220	🇹🇴 Tonga	1	1	0.4	0.4	92	810	1	1	41	256
221	🇸🇿 Eswatini	3	3	0.4	0.4	595	784	1	1	37	156
222	🇹🇻 Tuvalu	1	2	0.4	0.4	109	65	3	191	6	38
223	🇫🇰 Falkland Islands	1	4	0.3	0.3	37	82	0	0	18	93
224	🇨🇰 Cook Islands	2	5	0.3	0.3	88	329	7	125	10	73
225	🇸🇨 Seychelles	3	3	0.3	0.3	275	603	10	12	42	175
226	🇸🇭 Saint Helena	2	3	0.3	0.3	59	333	0	6	28	107
227	🇩🇲 Dominica	1	2	0.3	0.3	192	446	1	8	19	141
228	🇹🇱 Timor-Leste	5	5	0.3	0.3	274	124	0	83	31	59
229	🇳🇫 Norfolk Island	1	2	0.3	0.3	4	16	0	0	3	8
230	🇨🇻 Cabo Verde	2	2	0.2	0.2	301	546	17	16	29	144
231	🇲🇶 Martinique	3	4	0.2	0.2	99	545	1	6	7	69
232	🇸🇹 Sao Tome and Principe	1	1	0.2	0.2	491	196	20	193	6	91
233	🇬🇼 Guinea-Bissau	1	1	0.2	0.2	721	463	4	40	23	93
234	🇾🇹 Mayotte	2	2	0.2	0.2	125	253	0	25	20	50
235	🇸🇧 Solomon Islands	1	2	0.1	0.1	166	379	0	0	37	135
236	🇨🇽 Christmas Island	1	1	0.1	0.1	56	127	0	1	7	32
237	🇳🇺 Niue	3	3	0.1	0.1	0	39	0	0	3	9
238	🇰🇮 Kiribati	2	2	0.1	0.1	71	96	5	17	8	30

Top 200 cities by footage

#	City	Country	Videos	Footage (h)	Analysed (h)	Persons	Cars	Bicycles	Motorcycles	Buses	Trucks
1	New York, NY	🇺🇸 United States	4,154	2,268.9	1,444.8	1,706,401	5,052,176	129,512	64,473	183,174	787,499
2	Chicago, IL	🇺🇸 United States	1,896	1,713.9	1,361.7	557,028	4,660,659	30,176	9,676	71,355	609,334
3	London	🇬🇧 United Kingdom	2,029	1,358.5	1,178.4	2,453,710	2,498,403	247,264	144,524	234,765	459,569
4	Los Angeles, CA	🇺🇸 United States	2,043	1,162.5	894.0	537,321	3,352,836	28,868	13,836	53,046	473,977
5	Hanoi	🇻🇳 Vietnam	2,153	876.1	837.6	2,271,937	1,607,428	63,791	2,164,743	100,239	215,239
6	Manila, Metro Manila	🇵🇭 Philippines	2,148	781.6	778.8	1,145,457	1,371,466	36,235	601,253	162,779	342,650
7	Sydney, NSW	🇦🇺 Australia	1,432	682.1	501.9	271,359	1,892,879	20,266	23,030	48,875	330,708
8	Saint Petersburg, Saint Petersburg	🇷🇺 Russia	689	566.0	566.0	290,101	1,405,043	6,951	1,923	68,078	166,481
9	Seoul	🇰🇷 South Korea	293	542.9	516.6	229,316	902,725	13,254	38,006	74,720	86,184
10	Jakarta	🇮🇩 Indonesia	1,644	463.3	447.8	746,405	985,480	6,244	663,413	59,104	232,593
11	Paris	🇫🇷 France	975	436.1	409.2	552,903	1,036,028	67,843	87,489	31,039	121,109
12	Toronto, ON	🇨🇦 Canada	924	399.7	391.4	463,377	1,279,686	42,527	12,535	26,332	146,625
13	Yangon	🇲🇲 Myanmar	488	357.1	354.5	350,278	1,460,982	40,640	20,580	95,807	243,678
14	Melbourne, VIC	🇦🇺 Australia	670	329.6	274.2	199,141	994,924	12,797	10,296	18,731	115,711
15	Moscow, Moscow	🇷🇺 Russia	505	327.5	305.9	198,432	972,537	4,619	3,304	33,888	137,535
16	Philadelphia, PA	🇺🇸 United States	562	315.7	297.7	142,077	960,721	8,349	4,166	14,431	144,502
17	Dubai	🇦🇪 United Arab Emirates	745	311.6	260.1	81,498	623,309	2,874	13,966	36,942	89,583
18	Bangkok	🇹🇭 Thailand	607	303.2	292.8	378,040	805,865	6,907	266,751	40,184	164,837
19	Cotabato, Maguindanao del Norte	🇵🇭 Philippines	693	287.1	287.1	729,903	576,485	15,873	519,204	96,566	403,247
20	Sao Paulo, SP	🇧🇷 Brazil	567	251.5	248.9	236,579	682,482	11,189	61,981	21,273	88,328
21	København	🇩🇰 Denmark	529	244.7	238.7	363,421	698,970	192,474	8,042	19,601	58,402
22	Kyiv, Kyiv	🇺🇦 Ukraine	569	240.6	234.5	163,823	890,032	4,840	2,486	39,375	123,287
23	Kuala Lumpur, KUL	🇲🇾 Malaysia	346	237.3	156.3	95,099	460,754	718	49,102	10,679	38,508
24	Warsaw	🇵🇱 Poland	371	227.2	218.0	151,696	647,305	14,195	5,468	23,763	54,370
25	Berlin	🇩🇪 Germany	459	219.4	210.4	252,279	781,310	81,816	20,448	27,597	102,307
26	Buenos Aires, CABA	🇦🇷 Argentina	286	216.7	212.5	303,287	708,316	30,223	34,209	27,282	78,056
27	Townsville, QLD	🇦🇺 Australia	436	207.5	6.2	959	15,490	61	115	259	3,646
28	Detroit, MI	🇺🇸 United States	558	205.9	198.8	25,570	548,493	1,192	808	5,802	92,523
29	Guangzhou, Guangdong	🇨🇳 China	245	205.8	205.8	129,697	565,446	13,030	48,679	36,144	74,742
30	Montreal, QC	🇨🇦 Canada	470	202.5	202.5	114,906	570,680	13,299	3,268	10,961	50,197
31	Phnom Penh	🇰🇭 Cambodia	387	201.9	201.9	547,864	502,464	12,588	564,323	35,663	166,564
32	Helsinki	🇫🇮 Finland	282	198.1	195.5	107,865	388,813	10,614	1,719	19,427	48,452
33	Rio de Janeiro, RJ	🇧🇷 Brazil	448	195.7	145.2	157,425	441,215	11,282	27,791	23,171	47,546
34	Prague	🇨🇿 Czechia	241	188.2	187.1	198,447	462,120	5,078	4,823	20,212	45,241
35	Miami, FL	🇺🇸 United States	467	186.8	144.8	96,891	433,052	5,809	7,981	6,482	59,999
36	Kampala	🇺🇬 Uganda	466	176.4	40.1	210,110	97,690	5,333	116,814	9,743	31,876
37	Vienna	🇦🇹 Austria	714	173.7	172.5	103,254	436,287	11,849	12,577	20,222	74,380
38	Istanbul	🇹🇷 Türkiye	317	172.5	128.9	160,023	494,905	1,408	20,086	34,516	76,076
39	Eindhoven	🇳🇱 Netherlands	329	171.2	171.1	157,940	402,576	109,956	15,504	10,035	39,405
40	Yogyakarta	🇮🇩 Indonesia	298	171.0	168.9	505,006	332,763	7,633	519,508	13,346	70,159
41	Perth, WA	🇦🇺 Australia	418	170.9	99.6	18,366	242,204	1,209	1,154	5,082	44,734
42	Tokyo, Tokyo	🇯🇵 Japan	253	170.2	139.2	176,774	250,287	19,589	9,299	26,281	76,096
43	Stockholm	🇸🇪 Sweden	549	170.0	167.9	84,535	272,625	13,009	1,853	12,814	33,713
44	Nashville, TN	🇺🇸 United States	854	168.2	24.6	29,278	58,625	293	243	1,472	9,800
45	Bucharest	🇷🇴 Romania	325	166.7	164.2	126,986	768,567	4,781	4,178	19,656	68,103
46	Budapest	🇭🇺 Hungary	432	166.0	164.3	297,919	1,006,590	14,328	8,210	50,346	122,520
47	Singapore	🇸🇬 Singapore	414	162.0	152.6	81,730	342,138	5,203	20,569	33,441	60,162
48	Vancouver, BC	🇨🇦 Canada	669	160.7	40.3	37,238	127,551	4,288	745	2,760	18,383
49	Sacramento, CA	🇺🇸 United States	413	159.5	158.0	17,385	548,148	1,273	972	6,082	97,520
50	Atlanta, GA	🇺🇸 United States	372	159.4	154.4	31,624	380,748	1,267	941	4,942	79,393
51	Houston, TX	🇺🇸 United States	238	158.9	36.3	9,135	87,402	303	110	1,334	15,737
52	Oldenburg	🇩🇪 Germany	172	152.7	150.0	80,109	221,155	43,435	3,352	7,404	22,413
53	Tyumen, Tyumen oblast	🇷🇺 Russia	306	147.5	147.3	92,426	428,881	1,999	537	19,867	49,241
54	Winnipeg, MB	🇨🇦 Canada	504	137.4	49.7	18,868	158,186	1,061	373	2,979	25,269
55	San Jose, CA	🇺🇸 United States	288	134.9	133.8	25,398	580,106	1,714	1,093	6,796	82,755
56	Sofia	🇧🇬 Bulgaria	429	132.4	132.4	68,315	419,963	1,316	910	15,132	36,911
57	San Francisco, CA	🇺🇸 United States	194	128.0	119.3	151,893	375,262	8,128	6,349	10,175	49,048
58	Den Haag	🇳🇱 Netherlands	259	126.8	24.4	17,237	73,901	11,683	2,891	2,087	7,860
59	Anchorage, AK	🇺🇸 United States	347	125.1	121.6	5,832	174,761	276	177	2,765	43,598
60	Dallas, TX	🇺🇸 United States	295	124.9	120.6	13,586	211,948	365	345	2,841	37,571
61	Washington, DC	🇺🇸 United States	217	123.9	118.0	80,735	353,897	6,526	2,341	8,202	47,413
62	Tehran	🇮🇷 Iran	266	122.4	55.3	40,265	267,292	1,103	15,061	4,567	24,450
63	Busan	🇰🇷 South Korea	143	121.7	121.7	66,773	304,102	2,522	8,513	12,317	38,578
64	Las Vegas, NV	🇺🇸 United States	172	120.0	69.4	39,565	173,761	326	591	3,586	22,099
65	Maracaibo, ZU	🇻🇪 Venezuela	505	119.6	119.2	80,947	307,185	4,467	30,463	6,089	63,208
66	Dublin	🇮🇪 Ireland	392	119.5	114.8	94,809	261,693	9,941	2,021	12,669	34,841
67	Oslo	🇳🇴 Norway	312	116.5	115.0	70,501	249,450	6,475	1,640	11,863	33,043
68	Portland, OR	🇺🇸 United States	183	113.1	90.5	55,768	215,451	3,746	872	4,358	27,877
69	Volsk, Saratov oblast	🇷🇺 Russia	502	111.2	0.0	–	–	–	–	–	–
70	Phoenix, AZ	🇺🇸 United States	235	110.1	39.8	9,745	96,185	457	292	1,412	19,898
71	Amsterdam	🇳🇱 Netherlands	205	109.6	108.3	238,339	214,152	173,948	28,103	10,986	30,853
72	Hong Kong	🇭🇰 Hong Kong	199	99.3	56.2	95,898	94,454	1,185	2,867	18,406	18,096
73	Mumbai, MH	🇮🇳 India	263	98.5	93.0	232,809	317,101	6,219	94,627	18,328	71,425
74	Seattle, WA	🇺🇸 United States	204	97.4	0.0	–	–	–	–	–	–
75	Darwin, NT	🇦🇺 Australia	310	93.9	44.6	12,507	95,060	454	930	1,832	19,780
76	Cairo	🇪🇬 Egypt	206	91.6	39.0	59,798	204,076	503	6,928	9,885	25,205
77	Kermanshah	🇮🇷 Iran	247	89.6	88.5	58,357	510,160	573	3,502	3,744	42,344
78	Cape Town	🇿🇦 South Africa	229	86.4	82.8	55,930	233,672	1,332	3,815	7,796	39,276
79	Minsk	🇧🇾 Belarus	115	84.8	24.1	18,245	94,211	457	216	4,464	11,568
80	München	🇩🇪 Germany	169	81.8	81.8	80,877	260,636	26,811	4,455	10,029	33,084
81	Baltimore, MD	🇺🇸 United States	210	80.6	79.2	21,870	240,536	474	424	3,940	36,626
82	Serrekunda	🇬🇲 Gambia	267	77.0	77.0	179,757	252,094	10,645	17,150	6,831	52,811
83	Sylhet	🇧🇩 Bangladesh	194	74.5	74.5	302,421	78,883	58,803	56,760	26,739	63,853
84	Ho Chi Minh City	🇻🇳 Vietnam	140	73.8	55.0	168,431	72,961	3,064	173,005	6,858	17,849
85	Denver, CO	🇺🇸 United States	204	71.6	58.4	16,772	113,558	1,325	582	2,031	20,464
86	Gamagori, Aichi	🇯🇵 Japan	266	70.3	68.8	12,856	144,912	3,164	1,480	6,700	38,038
87	Ulaanbaatar	🇲🇳 Mongolia	127	70.2	70.2	87,943	304,331	1,879	1,639	9,117	25,617
88	San Diego, CA	🇺🇸 United States	94	68.3	65.1	50,800	209,892	2,225	1,785	3,279	32,191
89	Brisbane, QLD	🇦🇺 Australia	179	68.2	67.5	31,363	187,883	1,308	1,918	4,481	34,722
90	Karachi, SD	🇵🇰 Pakistan	147	67.1	58.4	186,698	238,894	6,397	95,433	16,302	55,090
91	Damascus	🇸🇾 Syria	148	65.6	65.6	100,239	279,804	4,492	11,104	10,387	43,459
92	Amman	🇯🇴 Jordan	199	61.9	61.9	35,564	263,325	107	881	5,889	32,157
93	Semarang	🇮🇩 Indonesia	322	61.7	61.5	85,584	91,843	1,616	85,609	6,698	31,419
94	Auckland	🇳🇿 New Zealand	142	61.5	58.7	43,978	159,198	2,536	1,631	6,578	19,959
95	Denpasar	🇮🇩 Indonesia	28	60.6	36.5	125,556	57,117	2,188	142,913	2,004	15,020
96	Tallinn	🇪🇪 Estonia	281	60.4	60.4	31,328	158,890	1,075	374	8,114	15,862
97	Calgary, AB	🇨🇦 Canada	170	60.0	54.5	15,237	130,767	699	195	1,970	22,194
98	Ninh Binh	🇻🇳 Vietnam	197	58.2	55.8	61,841	55,887	6,806	71,260	3,687	15,457
99	Beijing, Beijing	🇨🇳 China	130	57.2	57.2	45,355	162,551	11,268	11,586	10,206	8,844
100	Catania	🇮🇹 Italy	144	56.5	56.5	17,124	125,837	199	8,707	1,660	6,666
101	Delhi, DL	🇮🇳 India	163	55.9	37.8	51,595	94,506	2,173	20,504	11,917	17,356
102	Boston, MA	🇺🇸 United States	108	54.0	35.0	53,953	96,954	2,750	769	2,391	14,297
103	Milan	🇮🇹 Italy	101	52.9	52.9	73,742	184,386	10,189	16,478	3,902	15,866
104	Orlando, FL	🇺🇸 United States	113	52.1	40.8	13,575	112,764	1,122	583	1,983	16,372
105	Aleppo	🇸🇾 Syria	137	51.0	50.3	109,909	209,868	1,733	4,556	6,149	36,402
106	Shenzhen, Guangdong	🇨🇳 China	141	50.8	46.2	37,599	113,673	4,293	13,064	9,837	13,954
107	Honolulu, HI	🇺🇸 United States	118	49.7	24.0	15,235	58,589	959	667	1,444	8,390
108	Shanghai, Shanghai	🇨🇳 China	65	49.1	40.0	53,160	86,157	6,478	14,750	6,013	6,637
109	Ben Tre	🇻🇳 Vietnam	99	49.0	44.6	140,156	31,560	4,697	138,142	3,292	11,874
110	Barcelona	🇪🇸 Spain	100	48.3	41.3	66,762	92,102	7,090	28,619	5,776	12,996
111	Santiago	🇨🇱 Chile	85	48.1	2.4	3,844	8,094	171	358	608	919
112	Mexico City, CDMX	🇲🇽 Mexico	75	47.2	42.4	52,019	120,584	2,579	4,669	4,066	12,100
113	Kolkata, WB	🇮🇳 India	150	46.6	45.6	119,686	74,046	21,248	34,029	18,923	35,086
114	Ninohe, Iwate	🇯🇵 Japan	247	44.2	44.2	7,095	74,697	390	333	5,606	24,760
115	Lviv, Lviv oblast	🇺🇦 Ukraine	145	44.2	16.3	25,400	44,522	386	202	2,937	6,041
116	Sarajevo	🇧🇦 Bosnia and Herzegovina	158	44.0	30.8	37,053	107,527	389	316	3,220	8,656
117	Toyohashi, Aichi	🇯🇵 Japan	158	43.3	42.6	10,689	87,598	2,765	875	4,699	24,684
118	Colombo	🇱🇰 Sri Lanka	106	43.0	42.8	63,356	88,764	843	21,477	15,464	34,393
119	Adelaide, SA	🇦🇺 Australia	123	42.5	42.1	9,398	127,687	911	567	2,790	17,139
120	Addis Ababa	🇪🇹 Ethiopia	68	42.3	40.4	97,122	102,712	472	1,800	8,940	30,449
121	Mashhad	🇮🇷 Iran	224	41.4	41.4	16,596	134,177	402	3,860	2,156	6,318
122	Surabaya	🇮🇩 Indonesia	106	41.1	37.9	71,753	59,969	1,193	62,451	1,969	10,740
123	Nam Dinh	🇻🇳 Vietnam	128	40.9	35.6	63,624	44,482	5,773	74,649	1,685	8,048
124	Zaozhuang, Shandong	🇨🇳 China	122	39.9	39.9	23,493	126,774	1,344	13,859	5,983	19,760
125	Doha	🇶🇦 Qatar	78	39.5	39.5	10,088	100,428	407	1,230	4,807	14,696
126	Zürich, ZH	🇨🇭 Switzerland	76	39.4	39.4	53,676	80,646	9,908	5,253	4,172	10,945
127	Austin, TX	🇺🇸 United States	69	39.0	33.8	18,344	88,770	1,449	789	1,443	18,723
128	Jersey City, NJ	🇺🇸 United States	104	38.6	31.5	13,242	96,385	1,040	613	1,785	12,427
129	Baku	🇦🇿 Azerbaijan	91	38.3	38.3	37,213	189,258	524	1,048	5,479	17,773
130	Madrid	🇪🇸 Spain	75	38.2	37.4	49,386	92,021	1,480	8,534	5,202	9,817
131	Rome	🇮🇹 Italy	73	37.7	36.8	57,411	145,068	2,234	12,687	3,480	10,865
132	Lisbon	🇵🇹 Portugal	70	37.3	37.3	30,103	111,823	1,071	3,689	3,349	11,815
133	Brussels	🇧🇪 Belgium	82	37.1	37.1	43,354	104,356	6,663	2,438	4,921	14,586
134	Nicosia	🇨🇾 Cyprus	126	37.0	37.0	10,741	121,456	466	2,889	2,331	16,521
135	Freetown	🇸🇱 Sierra Leone	90	36.7	36.7	93,685	39,617	676	21,317	3,852	18,564
136	Kathmandu	🇳🇵 Nepal	112	35.8	35.6	207,593	69,547	5,489	123,918	10,451	18,978
137	Wellington	🇳🇿 New Zealand	118	35.7	34.8	17,968	92,977	986	1,071	3,114	12,604
138	Zagreb	🇭🇷 Croatia	94	35.7	35.7	27,055	131,851	2,828	958	3,139	9,776
139	Johor Bahru, JHR	🇲🇾 Malaysia	30	35.1	29.1	10,728	110,190	179	7,636	2,193	9,741
140	Kharkiv, Kharkiv oblast	🇺🇦 Ukraine	75	35.1	4.2	1,736	13,439	43	22	377	1,815
141	Lagos, Lagos	🇳🇬 Nigeria	73	35.1	34.3	150,463	103,955	1,369	17,007	9,246	27,667
142	Genoa	🇮🇹 Italy	73	35.1	34.3	46,192	126,323	928	38,965	3,751	12,946
143	Asaba, Delta	🇳🇬 Nigeria	77	34.3	0.8	1,748	2,602	8	88	354	986
144	Chennai, TN	🇮🇳 India	46	34.0	34.0	86,371	43,707	1,507	39,207	14,592	23,969
145	Riga	🇱🇻 Latvia	97	33.9	31.0	22,370	85,284	1,482	673	3,787	10,979
146	Thessaloniki	🇬🇷 Greece	81	33.8	33.4	42,570	158,855	1,217	12,761	3,361	13,565
147	Belgrade	🇷🇸 Serbia	96	32.7	32.1	31,196	132,090	446	949	4,988	10,467
148	Bengaluru, KA	🇮🇳 India	92	32.6	30.2	68,846	64,323	1,172	44,449	10,857	23,174
149	Tampa, FL	🇺🇸 United States	52	32.4	30.2	9,708	77,906	599	383	907	12,090
150	Birmingham	🇬🇧 United Kingdom	115	31.4	31.4	15,647	99,371	452	331	3,131	12,724
151	Goiânia, GO	🇧🇷 Brazil	69	30.8	28.6	15,464	75,395	807	8,488	1,145	8,074
152	Rotterdam	🇳🇱 Netherlands	78	30.7	30.7	28,374	79,600	15,394	3,276	2,296	8,649
153	Edmonton, AB	🇨🇦 Canada	56	30.4	28.8	6,735	90,209	385	216	1,400	13,357
154	Riyadh	🇸🇦 Saudi Arabia	68	29.9	18.7	2,912	92,150	52	182	2,073	12,038
155	Tbilisi	🇬🇪 Georgia	75	29.6	28.3	21,518	126,225	124	1,027	4,603	14,487
156	Chengdu, Sichuan	🇨🇳 China	54	29.1	27.9	17,124	108,684	2,470	5,470	4,274	8,164
157	Kansas City, KS	🇺🇸 United States	53	29.1	26.4	3,353	60,130	91	68	826	10,550
158	Toshkent	🇺🇿 Uzbekistan	85	28.2	26.2	19,360	169,476	198	76	3,115	12,382
159	Milton Keynes	🇬🇧 United Kingdom	50	27.7	27.7	1,573	56,289	173	185	751	5,888
160	Yerevan	🇦🇲 Armenia	89	27.1	5.3	4,343	32,103	30	60	1,184	3,979
161	Lahore, PB	🇵🇰 Pakistan	89	26.9	11.7	37,406	45,089	1,435	23,507	1,479	8,244
162	Omaha, NE	🇺🇸 United States	82	26.7	7.4	1,386	18,252	56	39	212	3,181
163	Manchester	🇬🇧 United Kingdom	61	26.4	26.2	18,914	68,504	978	268	2,393	8,171
164	Havana	🇨🇺 Cuba	132	26.3	1.8	6,648	1,806	219	1,028	238	768
165	Pittsburgh, PA	🇺🇸 United States	54	25.8	23.9	13,181	49,044	580	242	1,098	7,794
166	Cherepovets, Vologda oblast	🇷🇺 Russia	176	25.0	24.5	25,909	71,495	65	25	4,084	10,354
167	Thiruvananthapuram, KL	🇮🇳 India	74	24.7	24.5	62,049	56,188	1,007	59,061	6,669	21,952
168	Lima, LI	🇵🇪 Peru	36	24.3	24.3	42,349	99,765	4,224	6,467	4,832	12,191
169	Frankfurt	🇩🇪 Germany	20	24.3	0.0	–	–	–	–	–	–
170	Vilnius	🇱🇹 Lithuania	60	24.2	24.2	26,719	69,153	2,302	624	2,342	4,938
171	Daejeon	🇰🇷 South Korea	60	24.0	23.9	9,482	70,925	1,046	1,140	2,785	5,925
172	Islamabad, ICT	🇵🇰 Pakistan	57	23.9	23.6	29,256	69,216	1,253	13,868	2,143	9,481
173	Stavropol, Stavropol krai	🇷🇺 Russia	79	23.9	23.9	13,437	84,585	262	83	3,340	8,224
174	New Orleans, LA	🇺🇸 United States	34	23.7	21.8	23,581	52,522	1,217	356	1,248	11,118
175	Samarkand	🇺🇿 Uzbekistan	81	23.6	6.5	5,900	34,462	121	104	871	4,017
176	Charleston, SC	🇺🇸 United States	38	23.0	23.0	24,457	54,823	932	308	853	12,763
177	Sendai, Miyagi	🇯🇵 Japan	58	22.7	22.5	6,181	53,941	896	511	3,157	12,835
178	Ōsaka, Ōsaka	🇯🇵 Japan	52	22.5	21.2	29,007	41,703	7,570	1,635	3,054	9,231
179	Porto	🇵🇹 Portugal	27	22.4	22.4	29,125	56,586	360	1,671	1,825	6,486
180	Hangzhou, Zhejiang	🇨🇳 China	28	22.3	8.8	6,651	23,852	411	2,726	1,181	1,997
181	San Antonio, TX	🇺🇸 United States	26	22.2	11.1	8,981	21,950	350	146	423	4,316
182	Kyoto, Kyōto	🇯🇵 Japan	45	21.9	19.7	26,428	38,391	3,395	2,054	4,409	10,184
183	Taipei	🇹🇼 Taiwan	49	21.8	19.6	25,569	49,857	1,741	20,972	3,186	6,147
184	Cheongju	🇰🇷 South Korea	65	21.7	21.1	6,168	57,371	316	676	2,049	7,607
185	Ottawa, ON	🇨🇦 Canada	37	21.3	21.3	25,772	53,063	2,916	395	1,347	5,874
186	Chiang Mai	🇹🇭 Thailand	53	21.0	21.0	15,632	41,695	507	15,811	1,510	12,962
187	Sapporo, Hokkaidō	🇯🇵 Japan	61	20.9	20.8	7,760	35,309	523	92	2,059	4,800
188	Halifax, NS	🇨🇦 Canada	82	20.8	20.0	8,200	48,062	297	138	898	5,361
189	Medellín, ANT	🇨🇴 Colombia	53	20.8	2.9	5,392	7,277	141	1,864	478	962
190	Beirut	🇱🇧 Lebanon	68	20.5	20.5	16,855	101,182	106	6,178	1,850	10,127
191	Monaco	🇲🇨 Monaco	45	20.3	19.5	35,821	30,718	962	12,489	1,542	4,295
192	Manaus, AM	🇧🇷 Brazil	75	19.8	19.3	18,745	67,239	201	9,711	2,259	10,500
193	Chișinău	🇲🇩 Moldova	70	19.7	19.7	22,351	92,843	170	173	4,535	10,741
194	Ankara	🇹🇷 Türkiye	47	19.3	19.3	16,485	89,827	48	855	4,883	9,579
195	Malibu, CA	🇺🇸 United States	34	19.2	16.7	4,065	67,359	276	452	785	14,381
196	Nice	🇫🇷 France	32	19.0	18.2	31,342	47,380	3,513	7,303	1,300	5,498
197	Pattaya	🇹🇭 Thailand	35	19.0	13.0	15,622	32,422	200	13,824	1,828	10,178
198	Skopje	🇲🇰 North Macedonia	39	18.9	18.9	21,132	71,350	1,425	755	2,142	6,384
199	Johannesburg	🇿🇦 South Africa	76	18.6	18.2	14,194	42,414	66	428	2,293	8,813
200	Roseville, CA	🇺🇸 United States	46	18.4	18.4	1,251	69,292	107	123	559	12,479

Citation and usage of code

If you use this work for academic work please cite the following paper:

Alam, M. S., Bazilinska, O., & Bazilinskyy, P. (2026). A global dataset of continuous urban dashcam driving. arXiv preprint arXiv:2604.01044. Under review. Available at https://arxiv.org/abs/2604.01044

What the dataset contains, how it was collected and its known biases are described in the datasheet. Citation metadata is in CITATION.cff.

The code is open-source and free to use. It is aimed for, but not limited to, academic research. We welcome forking of this repository, pull requests, and any contributions in the spirit of open science and open-source code. For inquiries about collaboration, you may contact Md Shadab Alam (md_shadab_alam@outlook.com) or Pavlo Bazilinskyy (p.bazilinskyy@tue.nl).

Repository layout

Top folder: processing the videos and the analysis

main.py: downloads the videos of the mapping (from the FTP server, else YouTube) and runs YOLO tracking on their segments, writing one detection CSV per segment to the data folders; runs in a loop. helper_script.py holds its download, trimming and tracking code.
analysis.py: the tables, maps and figures of the dataset and the statistics in this README, from the mapping and, when available, the YOLO detections. count_detections.py counts the detections where the CSVs are and saves the counts (analysis_data/detection_counts.csv.gz), so that analysis.py and the GitHub Action can use them without the CSVs.
add_video.py: the page for adding videos and segments to the mapping by hand (templates in templates/).
common.py (configs, secrets, country codes, git, email), custom_logger.py and logmod.py (logging): shared by all scripts.

Top folder: proposing segments from YouTube channels (see Proposing segments)

propose_segments.py: proposes segments, localities, time of day and vehicle of a channel's videos, and serves the review page of a batch.
run_channels.py: works through the channels sheet, batch by batch; refresh_queue.py builds the queue of channels to look at again for new uploads.
review_hub.py: the page listing the batches to review, and applying and pushing reviewed segments.
sync_workers.py: brings the batches of the worker machines to the reviewing machine and the mapping to them.
review_merges.py: merging towns into the city whose metropolitan area they lie in; review_lessons.py: what the reviews say about the proposals.

Folders

utils/: the analysis and figure code used by analysis.py:
- core/parsing.py (the mapping's list cells, the processed length of a segment), core/dataset_stats.py (totals);
- analytics/contributors.py (who added each video), analytics/metrics_cache.py (YOLO class IDs, unique keys);
- plotting/dataset_figures.py (the README figures), plotting/area_figures.py (county, region, smoothed and spike maps), plotting/maps.py (world and continent maps), plotting/map_labels.py (label placement), plotting/hover.py (popups shared by all figures), plotting/io.py (saving figures), plotting/constants.py.
release/: the weekly metadata update and the release files (update_info.py, make_crowd_jsonl.py, split_large_files.py, validation_release.py, weekly_metadata.sh); make_splits.py (the train, validation and test splits and the balancing weights in crowd_splits.csv: grouped by channel or by locality, so no source is in two splits); and the manual audits: sample_audit.py draws the stratified samples, audit_stats.py scores the label audit and evaluate_detections.py the detections from the manual review; annotate.py is the local browser tool for those audits (labels, review of the detector's boxes in frames, tracking errors in clips), which finds the videos and detection files in any number of folders.
tools/: maintenance scripts for the mapping and the data, e.g. download_videos.py, video_inventory.py, update_traffic_index.py, update_params.py and weekly_detection_counts.sh; the review of cuts is detect_cuts.py (candidates), review_cuts.py (the review page) and apply_cuts.py (splits the mapping and the detection files at the confirmed cuts).
dashboards/: pages and widgets that follow the pipeline (its own README).
analysis_data/: data files the analysis reads and writes: country tables (countries.csv, height_data.csv), the YOLO detection counts (detection_counts.csv.gz) and who added which video (video_contributors.csv). The mapping and metadata files (mapping.csv, mapping-NAME.csv, mapping_metadata.csv, crowd_index.jsonl, ...) and cut_times.csv stay in the top folder.
figures/: the figures of this README (PNG, HTML and, when made on purpose, EPS), and figure_hashes.json, which records what each file was made from, so that only changed figures are saved again (see Step 11); readme/ and screenshots/: images of this README and of the figures.
logs/: records kept by the scripts (traffic_index_log.csv, appended by tools/update_traffic_index.py).
archive/: old files no code uses any more (mappingsubmitted.csv).
tests/: tests of the scripts.
.github/workflows/: the weekly refresh of the README statistics and figures (dataset-figures.yml) and the linter.

Run the scripts with the virtual environment active (see Step 11).

Getting started

Tested with Python 3.12.13 and the uv package manager.

The project is configured to automatically select the appropriate PyTorch backend. On systems with a supported NVIDIA GPU and CUDA driver, uv installs a CUDA enabled version of PyTorch. If no supported GPU is detected, it automatically falls back to the CPU version.

Step 1: Install `uv`

uv is a fast Python package and environment manager.

macOS / Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows PowerShell:

irm https://astral.sh/uv/install.ps1 | iex

Alternative, if Python and pip are already installed:

pip install uv

Step 2: Fix permissions if needed

In some environments, uv may need to create directories under:

macOS / Linux:

~/.local/share/uv/python

Windows:

%LOCALAPPDATA%\uv\python

If these directories were previously created by another user or with elevated privileges, you may encounter an error such as:

error: failed to create directory ... Permission denied (os error 13)

macOS / Linux

mkdir -p ~/.local/share/uv
chown -R "$(id -un)":"$(id -gn)" ~/.local/share/uv
chmod -R u+rwX ~/.local/share/uv

Windows PowerShell

New-Item -ItemType Directory -Force "$env:LOCALAPPDATA\uv"

icacls "$env:LOCALAPPDATA\uv" /grant "$($env:UserName):(OI)(CI)F"

Step 3: Verify `uv`

Check that uv is available:

uv --version

Step 4: Clone the repository

git clone https://github.com/crowd-dataset/crowd.git
cd crowd

Step 5: Install Python 3.12.13

The project is tested with Python 3.12.13.

Install it using uv:

uv python install 3.12.13

The repository contains a .python-version file so that uv can automatically select the correct Python version.

Step 6: Create the virtual environment

Create a virtual environment in .venv using Python 3.12.13:

uv venv --python 3.12.13

Step 7: Activate the virtual environment

macOS / Linux:

source .venv/bin/activate

Windows PowerShell:

.\.venv\Scripts\Activate.ps1

Windows Command Prompt:

.\.venv\Scripts\activate.bat

Step 8: Install the dependencies

Install the project dependencies from pyproject.toml:

uv pip install -r pyproject.toml

The project uses:

[tool.uv.pip]
torch-backend = "auto"

This allows uv to automatically select the appropriate PyTorch backend for the current machine.

For example:

NVIDIA GPU with a compatible CUDA driver → CUDA enabled PyTorch
No supported GPU → CPU PyTorch

No manual selection of a CUDA version should normally be required.

Note: Do not use uv sync for the initial installation if automatic PyTorch backend detection is required. Automatic torch-backend selection currently applies to the uv pip interface.

Step 9: Verify the PyTorch backend

Check which PyTorch backend was installed:

python -c "import torch; print('PyTorch:', torch.__version__); print('CUDA available:', torch.cuda.is_available()); print('CUDA version:', torch.version.cuda); print('Device:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'CPU')"

On an NVIDIA CUDA system, the output should look similar to:

PyTorch: 2.8.0+cu129
CUDA available: True
CUDA version: 12.9
Device: NVIDIA GeForce RTX 5090

On a machine without a supported GPU, CUDA available will be False and the CPU version of PyTorch will be used.

Step 10: Create `config` and `secret`

mapping.csv (the videos, segments and localities) is in the top folder of the repository. Copy the two templates in the top folder:

cp default.config config
cp default.secret secret

In config, set data to the folders that hold (or will hold) the YOLO output (bbox/*.csv in each), videos to the folders with the videos, and mapping to the mapping file (mapping.csv); the other parameters are described in Configuration of project. The scripts stop when there is no config. Fill in only the keys of secret you need (The secret file); the analysis needs none.

Step 11: Run the scripts

Run the scripts from the top folder of the repository with the virtual environment active:

python main.py                     # download the videos and run YOLO on their segments (runs in a loop)
python analysis.py                 # tables, figures and README statistics (YOLO counts from the data folders
                                   # or, without them, from analysis_data/detection_counts.csv.gz)
python count_detections.py --data /path/to/data   # count the YOLO detections and save the snapshot

analysis.py saves a figure (each of its HTML, PNG and EPS files) only when what it shows changed since it was last saved, as recorded in figures/figure_hashes.json, and renders the PNG and EPS files in parallel; an unchanged run takes about a minute. To save every figure again, delete figures/figure_hashes.json first.

The scripts in tools/ and release/ can be run as a file or as a module, and from any folder (they switch to the top folder, where mapping.csv and the other files are):

python tools/check_lists.py
python -m tools.check_lists

The tests run with python -m pytest tests (after uv pip install pytest); tests/test_reject_segments.py is a script run with python tests/test_reject_segments.py on the reviewing machine, as it needs its review plans. Some modules have a self-check: python -m utils.core.parsing, python -m utils.plotting.map_labels.

Configuration of project

Configuration of the project needs to be defined in config; copy default.config to start. The scripts stop when config is missing or has fewer entries than default.config. Entries marked optional may be left out. The config file has the following parameters:

Data and mapping

data: folders with the YOLO output (a bbox/ folder of detection CSVs in each).
videos: folders with the downloaded videos.
mapping: the mapping file (mapping.csv).
countries_analyse: ISO3 codes of the countries to analyse; empty for all.
min_confidence: minimum YOLO confidence of a detection counted in the analysis.

Video processing (main.py)

machine_name: name of the machine in the emails sent by main.py.
tracking_mode: run YOLO tracking and write the detection CSVs.
tracking_model: the YOLO model file (e.g. yolo11x.pt).
bbox_tracker: the tracker settings (bbox_custom_tracker.yaml, its track buffer set from track_buffer_sec).
track_buffer_sec: how long (s) a lost track is kept before it ends.
max_workers: segments processed at the same time.
download_workers: videos downloaded or prepared at the same time.
max_active_segments_per_video: segments of the same video processed at the same time: with 1, the workers take segments of different videos (with max_workers=3, 3 videos at once); with 2+, several workers may take segments of one video, which helps when a video has many segments.
runs_root: folder for YOLO's working files.
external_ssd: the videos are on an external drive: segments are trimmed on the internal drive (the last folder of videos).
ftp_server: server from which videos are fetched before trying YouTube (ftp_username and ftp_password, and optionally ftp_token, in secret).
delete_youtube_video, delete_runs_files, delete_labels: delete downloaded videos, YOLO's working files and its label files after processing.
save_annotated_video, save_annoted_img, save_tracked_img, display_frame_tracking: save an annotated video, save annotated or tracked frames, show the tracking while it runs.
snellius_mode, snellius_shard_mode: running as several tasks on the Snellius cluster, each taking its share of the segments.
sleep_sec, git_pull: pause (s) at the end of each loop of main.py, and pulling the repository then.
email_send, email_sender, email_recipients: email when a job finishes or crashes (see the secret file below).

Downloading from YouTube

client: pytubefix client ("WEB", "ANDROID" or "IOS").
need_authentication: sign in to YouTube for downloads with pytubefix.
update_package: update pytubefix and yt-dlp (on Mondays) before downloading.
Optional: use_pytubefix (false: yt-dlp only), cookies_file (a Netscape cookies file for yt-dlp), cookies_browser (yt-dlp's --cookies-from-browser, e.g. chrome) and js_runtime (yt-dlp's JavaScript runtime, e.g. node:/path/to/node).

Updating the mapping

update_pop_country, update_gini_value: update the countries' population and Gini index in the mapping from the World Bank when main.py starts.

Figures and output

font_family, font_size, plotly_template: look of the figures.
logger_level: level of console output: debug, info, warning or error.
Optional: open_html_figures (true: open every saved HTML figure in the browser) and save_eps (false: no EPS figures, as in the GitHub Action; EPS needs poppler's pdftops).

The `secret` file

API keys, passwords and usernames live in a file named secret (no extension) in the root of the project. It is ignored by git, so it is never committed. To create it, copy the template default.secret to secret and fill in the values you need. Any value can stay empty ("") if you do not use the feature it belongs to. Running the analysis on the released dataset needs none of them; running main.py needs at least an empty secret copied from the template.

The file is a single JSON object:

{
  "geonames_username": "",
  "bea_api_key": "",
  "tomtom_api_key": "",
  "restcountries_api_key": "",
  "email_smtp": "",
  "email_account": "",
  "email_password": "",
  "ftp_username": "",
  "ftp_password": "",
  "carto_api_key": "",
  "youtube_api_key": "",
  "locationiq_api_key": ""
}

Which keys you need depends on what you run:

analysis of the released dataset (analysis.py): none; carto_api_key is optional.
adding videos (add_video.py): geonames_username, restcountries_api_key, tomtom_api_key and youtube_api_key; bea_api_key for US localities.
proposing segments from channels (propose_segments.py): youtube_api_key; geonames_username and locationiq_api_key help.
updating the traffic index (tools/update_traffic_index.py): tomtom_api_key.
processing videos (main.py): ftp_username and ftp_password for the VideoFiles server; the email keys only if email_send is on in config.

Key	Used by	What for	Where to get it
`geonames_username`	`add_video.py`, `propose_segments.py`	locality population, a fallback for coordinates, and place lookups	free account at GeoNames; see below
`bea_api_key`	`add_video.py`	GDP of US localities	free key from BEA; see below
`tomtom_api_key`	`add_video.py`, `tools/update_traffic_index.py`	traffic index from TomTom traffic flow	free account at TomTom; see below
`restcountries_api_key`	`add_video.py`	country population, continent and Gini	free account at REST Countries; see below
`email_smtp`, `email_account`, `email_password`	`main.py`	email when a job finishes (`email_send` in `config`)	your email provider; see below
`ftp_username`, `ftp_password`	`main.py`	downloading videos from the VideoFiles server	the maintainers (see Contact)
`carto_api_key`	`utils/plotting/maps.py`	optional: basemap tiles for the Mapbox-style maps without CARTO's watermark; without a key, CARTO's free basemap is used	CARTO; see below
`youtube_api_key`	`propose_segments.py`, `add_video.py`	listing a channel's videos and their length, upload date and channel, without downloading them	see below
`locationiq_api_key`	`propose_segments.py`	optional: faster place lookups (localities from titles, descriptions and chapters, and the towns along a drive "from A to B")	see below

Getting a YouTube Data API key

The key is free; the default quota of 10,000 units a day covers several large channels (listing a channel's search results costs 100 units per 50 videos, video details 1 unit per 50 videos).

Go to the Google Cloud console and sign in with the Google account you want the key to belong to.
Create a project: project selector at the top → New project → give it a name (e.g. crowd) → Create, then select it.
Open APIs & Services → Library, search for YouTube Data API v3 and click Enable.
Open APIs & Services → Credentials → Create credentials → API key. Copy the key that appears.
Recommended: click the new key → under API restrictions choose Restrict key → select YouTube Data API v3 → Save, so the key cannot be used for anything else.
Put the key in secret as "youtube_api_key": "...".

Getting a LocationIQ key (optional)

propose_segments.py looks places up on OpenStreetMap: to guess a video's locality from a named place ("Al Khan Beach" is in Sharjah) and to split a drive "from A to B" into the towns along its route (about 25 lookups per such drive). Without a key it uses the public Nominatim server, which allows one request a second and refuses requests for a while when a busy day goes over that. LocationIQ runs the same OpenStreetMap search with a free key: 5,000 requests a day at two a second, with no payment details. With a key in secret, lookups go to LocationIQ; once its daily requests are used up, they fall back to the public server.

Sign up for the free plan at locationiq.com and confirm your email address.
Open the dashboard's Access Tokens page and copy the token there.
Put it in secret as "locationiq_api_key": "...". Nothing else needs changing; running channel batches pick it up from their next chunk.

Getting a GeoNames username

GeoNames is free; an account allows 10,000 requests a day and 1,000 an hour.

Create an account at geonames.org/login and confirm your email address with the link GeoNames sends.
Sign in and open your account page. At the bottom, click Click here to enable the free web services. Requests are refused until this is done.
Put your username (not your email address) in secret as "geonames_username": "...".

Getting a BEA API key

The U.S. Bureau of Economic Analysis key is free and is used only for localities in the USA.

Fill in your name and email address at apps.bea.gov/API/signup.
BEA emails you the key (a 36-character code); click the activation link in that email.
Put the key in secret as "bea_api_key": "...".

Getting a TomTom API key

The traffic index uses TomTom's Traffic Flow API (segment data). Its free tier allows 20,000 requests a month, with no payment details needed.

Create an account at developer.tomtom.com and confirm your email address.
Open the Dashboard: a key is created with the account (Keys section). Copy it.
Put it in secret as "tomtom_api_key": "...".

tools/update_traffic_index.py uses one request per locality and only queries localities where it is local daytime; requests beyond the free tier are refused unless payment details are added to the TomTom account.

Getting a REST Countries API key

add_video.py uses version 5 of the REST Countries API, which needs a key. The free plan allows 1,000 requests a month, with no payment details.

Sign up at restcountries.com/sign-up and confirm your email address.
Create an API key in the dashboard and copy it.
Put it in secret as "restcountries_api_key": "...".

Email settings

main.py can email you when a job finishes (email_send, email_sender and email_recipients in config). It connects to the SMTP server over SSL on port 465.

email_smtp: the SMTP server of your provider, e.g. smtp.gmail.com for Gmail.
email_account: the full email address you send from.
email_password: for Gmail, an app password, not your normal password. Turn on 2-Step Verification for the Google account, then create an app password at myaccount.google.com/apppasswords and paste the 16 characters. Other providers have similar app passwords or use the account password.

Getting a CARTO key (optional)

The Mapbox-style maps (mapbox_map_*) use CARTO's free basemap, which shows a CARTO watermark. With a key, maps.py loads CARTO's basemap tiles with the key instead.

Create an account at carto.com.
In the CARTO Workspace, create an API access token under Developers → Credentials and copy it.
Put it in secret as "carto_api_key": "...".

Example of YOLO output for a video with dashcam footage

Video: https://www.youtube.com/watch?v=_Wyg213IZDI.

Selection procedure

Verification of segments

After segments were added, they were checked again by hand. The checks were done by Md Shadab Alam, on the main branch, for the segments of mapping.csv, including the videos that other contributors of this repository added (see Who added the videos).

Why a manual step. A retained segment should be one continuous, unedited stretch of driving. Watching all the footage again is not feasible, but an automatic detector of abrupt visual change can point to the few seconds where a cut is likely. The detector alone is not reliable enough to cut the data: a sudden change can be an edit, but also the entrance of a tunnel, a flash of sunlight or a passing lorry, and telling them apart needs context. So the detector proposes candidates, a person decides, and only the decisions of the person are applied. Among the candidates that could be judged so far, most were not cuts (about 59%), so splitting at every candidate would have broken many continuous segments.

1. Candidate cuts. python -m tools.detect_cuts runs FFmpeg scene-change detection (threshold 0.45) on the original video of each segment, over the segment only, merges times closer than 2 s, and writes the candidates to cut_times.csv with the video, the segment start, the frame rate and the time of the candidate. The threshold favours abrupt cuts: gradual transitions, and cuts between similar scenes, can be missed, and the detector does not find speed changes or advertisements.

2. Reviewing. python -m tools.review_cuts serves a page (port 5050) that plays the YouTube video from two seconds before to two seconds after each candidate. The reviewer marks the candidate as yes (a real cut), no (not a cut) or unavailable (YouTube no longer plays or embeds the video), and can add a note. The decision and the time of the review are saved in cut_times.csv.

In total, 52,838 cuts were analysed.

3. What the reviewer looked for.

Cuts in videos added by other contributors. The segments of the videos added by the other authors of the repository were checked for cuts, in the same way as the others.
Videos that are not at normal speed. Videos that play faster or slower than 1x (for example fast-forwarded footage or time-lapses) are not suitable, because speeds and crossing times in them are wrong. Such cases are noted, for example as fast forward.
Advertisements in the footage. An advertisement that was cut into the middle of a video is not footage of the road. Where one was found, it was removed from the segments by the reviewer.

4. Applying the result. python -m tools.apply_cuts applies the candidates marked yes. A cut at time t splits its segment in mapping.csv into a part that ends at floor(t) and a part that starts at ceil(t), so one to two seconds around the cut are dropped, and parts shorter than one second are discarded. The detection CSV of the segment is split at the same frames, with frame numbers restarted at 1 and the true frame rate read from the video (not the rounded one in the file name), so the split files stay aligned. Segments that become too short are removed, and detection CSVs that no longer match a segment in mapping.csv are removed from the data folder.

Description and analysis of dataset

Note: The interactive figures are displayed using htmlpreview.github.io, which may be blocked on some institutional or education networks. If an interactive figure does not open, try another network or download the corresponding .html file from the figures directory and open it locally in a web browser.

Description of dataset

Locations of cities with footage in dataset. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Locations of cities with footage in dataset with a density overlay of population. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Locations of cities with footage in dataset with a density overlay of the number of videos in the dataset. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Locations of cities with footage in dataset with a density overlay of the total number of seconds of footage in the dataset. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Amount of footage (hours) per country in Africa. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in Asia. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in Europe. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in North America. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in Oceania. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in South America. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per US state. Note: the colour scale is logarithmic.

Footage per US county (log colour scale): localities are assigned to the county they lie in. Hover over a county for its localities.

Hours of footage per million people per US county (population: US Census Bureau estimates, 2024).

Share of footage recorded at night per US county (counties with at least an hour of footage).

Footage per European region: Eurostat NUTS 2 regions for the EU, EFTA, the United Kingdom and candidate countries, and geoBoundaries regions for European Russia (west of the Urals), Ukraine, Belarus and Bosnia and Herzegovina; Moldova, Kosovo and small countries as a whole.

Hours of footage per million people per European region (population: Eurostat, each region's latest year; for Russia, Ukraine, Belarus and Bosnia and Herzegovina, Wikidata's latest figure).

Share of footage recorded at night per European region (regions with at least an hour of footage).

One spike per locality in the contiguous United States (Alaska and Hawaii are not in the view); height and colour show its hours of footage (log scale).

One spike per locality in Russia; height and colour show its hours of footage (log scale).

One spike per locality in Japan.

Where the footage is in Europe: hours of footage per 10,000 km², smoothed over about 50 km (log colour scale); grey: no footage nearby. Callouts mark the densest places.

Where the footage is in North America.

Where the footage is in Asia.

Blind spots of the dataset in Europe: distance from every place to the nearest locality with at least an hour of footage. Callouts mark the places furthest from any footage.

Blind spots of the dataset in North America.

Blind spots of the dataset in Asia.

Share of footage recorded at night in Europe, smoothed over about 50 km and weighted by footage; grey where there is less than an hour of footage within 100 km.

Average year of upload of the footage in Europe (weighted by footage): where the footage is recent and where it is older.

Number of videos against the amount of footage per locality (log scales), coloured by continent, with every country and its flag in the legend on the right, in alphabetical order (scroll it in the HTML version; click a country to hide or show it); hover over a point for its details. The 40 localities with the most footage are labelled where there is room, the top 12 in larger black text; in the HTML version, zoom in to label every point.

The same on linear scales, which show how far the localities with the most footage lead, with the same country legend. The 40 localities with the most footage are labelled where there is room, the top 12 in larger black text; in the HTML version, zoom in to label every point.

Number of videos against the amount of footage per country (log scales), coloured by continent; hover over a point for its details. Countries are labelled with their flag and ISO3 code, the 30 with the most footage in larger black text; in the HTML version, zoom in to label every point.

The same on linear scales, which show how far the countries with the most footage lead. The 30 countries with the most footage are labelled with their flag and ISO3 code where there is room; in the HTML version, zoom in to label every point.

Footage (hours) by continent, by day and by night. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Number of videos by quarter of upload and continent. The current quarter (hatched) is not complete yet.

Footage by the type of vehicle it is filmed from (log scale), with each type's share of the footage and of its footage at night.

Amount of footage per locality against its population. Large localities with little footage are candidates for adding videos.

Footage per locality against its population (log scales), coloured by whether a locality has more than three times, about, or less than a third of the footage its size predicts (the dotted least-squares line). Labelled: the most populous localities, those with the most footage, and the most under-covered cities of over a million people, which are the next ones to collect.

Amount of footage per locality against its gross metropolitan product (GMP of the metropolitan area: localities in one metropolitan area share its GMP), with the Spearman rank correlation: whether economically large cities dominate the data.

Amount of footage per locality against its traffic index (how much slower than free flow traffic is, from one TomTom reading per locality), for localities with a reading above 0, with the Spearman rank correlation.

Amount of footage per country against country indicators, with the Spearman rank correlation and a trend line per panel, showing whether some kinds of countries are over- or under-represented.

Share of footage recorded at night per country. Countries with less than 10 hours of footage are grey, as their share is not reliable.

Share of footage recorded at night per European country, printed on each (countries with at least 10 hours of footage; Russia and Türkiye whole).

Hours of footage per million inhabitants per country: coverage relative to country size (log scale; countries with less than 10 hours of footage are grey).

Hours of footage per million inhabitants per European country, printed on each (log scale; countries with at least 10 hours of footage; Russia and Türkiye whole).

Hours of footage per million people against GDP per person (both log scales; circle size: population), coloured by whether a country has more than twice, about, or less than half the footage its wealth predicts (the dotted least-squares line). GDP per person at purchasing-power parity is the World Bank's latest value.

Share of each country's footage that comes from its largest YouTube channel. Where it is high, one uploader's routes and camera dominate that country's data. Countries with less than 10 hours of footage are grey.

Type of vehicle the footage is recorded from, as a share of footage per continent.

Share of footage coming from the largest YouTube channels.

Distribution of the length of segments (parts of videos included in the dataset).

Every locality in the dataset as a dot coloured by continent; the dot area is proportional to the hours of footage.

The same localities on a globe. The HTML version spins until you grab it; drag to turn it.

The localities on a 3D satellite globe (NASA Blue Marble), with a spike for each locality whose height shows its hours of footage on a log scale, so localities with little footage are visible too. In the HTML version, a button merges neighbouring localities into one spike per area of about 200 km, which is easier to read where localities are dense. Each spike is orange for the share of its footage recorded by day and blue for the share at night. The HTML version spins until you grab it; drag to turn it, scroll to zoom.

The spinning globe of the CROWD dashboards: every locality as a square that grows with its hours of footage, and the places with segments added in the last week (red) and month (amber). The HTML version loads the latest data when it opens (from the globe on bazilinskyy.github.io, rebuilt every night) and turns until you grab it; drag to turn it, hover over a place for its footage and new segments, and click it to watch its longest segment on YouTube.

Footage by continent, country and locality, coloured by the share recorded at night. A country with localities on two continents (Russia, Türkiye) has a box in each. In the HTML version, click a box to zoom in down to the localities.

Share of footage against share of population (of all countries in the dataset) for the countries among the 25 largest by footage or the 25 largest by population (log scale). Where the orange dot is left of the blue one, the country is under-represented.

Share of footage by population of the locality, per continent and overall: what "urban" means in each part of the dataset.

Effective number of YouTube channels per country (inverse Herfindahl index of the channels' shares of footage): 1 means a single uploader; n means as diverse as n channels with equal shares. Countries with less than 10 hours of footage are grey.

Effective number of YouTube channels against hours of footage per country (log scales; circle size: population), coloured by whether a country has more than twice, about, or less than half the channels its amount of footage predicts. Countries below the line rest on few uploaders, so their footage reflects few channels' routes, cameras and styles.

Where the 20 YouTube channels with the most footage film, numbered by footage, with the number of countries and hours of each (dot area proportional to hours). Travel channels span many countries; most large channels film in one country or city. In the HTML version, each channel's number links to the channel on YouTube.

Length of segments by continent and by type of vehicle (boxes: quartiles; whiskers: 5th and 95th percentiles).

Detections in the footage (YOLO)

Objects detected in the footage by YOLO (yolo11x with tracking): unique tracked persons, cars, bicycles, motorcycles, buses and trucks, counted with a confidence of at least 0.7 and expressed per minute of analysed footage. The counts per segment are kept in analysis_data/detection_counts.csv.gz (video, start of the segment and the unique tracked objects of each class), which python count_detections.py --data <folder with bbox/> writes where the YOLO detection CSV files are (the first run reads about 900 GB of CSVs, in parallel with --workers; later runs read only new or changed files, cached in _output/). analysis.py, also in the GitHub Action, makes the tables and figures below from it, matched to the current mapping, so they stay up to date without the CSV files. Every Sunday at 20:00 (Amsterdam time), tools/weekly_detection_counts.sh on the machine with the data drive counts newly analysed footage and pushes analysis_data/detection_counts.csv.gz if it changed, ahead of the weekly refresh of the figures.

The detections are automatic outputs, not ground truth. What the detector sees depends on the camera, resolution and viewpoint, so the figures by vehicle type and by year of upload show these effects next to the comparisons between countries.

Based on the YOLO detections in 86,123 processed segments, last updated 2026-10-21 17:43 UTC.

Tables

Detections per continent

#	Continent	Analysed (h)	Of footage (%)	Localities	Persons / min	Cars / min	Bicycles / min	Motorcycles / min	Buses / min	Trucks / min
1	Africa	756.5	73.5	540	37.23	43.51	1.16	7.93	2.24	9.40
2	Asia	9,277.5	94.2	2,141	24.28	40.58	1.03	15.66	2.73	8.11
3	Europe	8,440.7	88.5	3,651	17.13	44.75	2.61	1.35	1.94	5.54
4	North America	9,030.0	75.7	2,285	9.71	50.13	0.65	0.39	1.00	7.86
5	Oceania	1,326.6	69.4	436	8.31	53.50	0.55	0.56	1.30	8.77
6	South America	1,127.0	86.3	398	18.47	47.95	1.32	4.45	1.78	7.03
7	All	29,958.2	84.2	9,451	17.27	45.56	1.35	5.74	1.88	7.33

Top 20 countries by pedestrians per minute (at least 10 analysed hours in at least 3 localities)

#	Country	Pedestrians / min	95% interval	Analysed (h)	Localities
1	🇳🇵 Nepal	92.85	25.71–96.55	38.1	4
2	🇳🇬 Nigeria	63.33	38.81–70.29	56.9	34
3	🇧🇩 Bangladesh	60.73	41.95–66.20	107.2	29
4	🇰🇪 Kenya	52.53	49.16–59.94	19.9	55
5	🇵🇰 Pakistan	51.89	38.88–68.35	141.9	37
6	🇮🇳 India	44.03	40.32–48.90	450.1	293
7	🇰🇭 Cambodia	43.59	24.48–45.17	222.3	7
8	🇻🇳 Vietnam	42.64	25.71–44.98	1,110.8	68
9	🇪🇹 Ethiopia	41.18	40.06–74.06	42.2	8
10	🇬🇲 Gambia	39.59	33.06–50.93	95.4	29
11	🇲🇼 Malawi	37.69	29.70–50.05	29.9	48
12	🇨🇬 Congo (Congo-Brazzaville)	37.40	5.83–38.06	11.8	3
13	🇸🇱 Sierra Leone	35.88	24.30–41.23	79.1	46
14	🇨🇴 Colombia	34.63	31.51–38.40	41.4	35
15	🇬🇭 Ghana	32.50	27.05–59.00	11.4	10
16	🇮🇩 Indonesia	32.12	26.05–41.46	969.2	175
17	🇲🇦 Morocco	30.32	26.78–36.30	46.9	34
18	🇬🇧 United Kingdom	29.73	9.51–32.97	1,496.8	353
19	🇸🇾 Syria	29.49	15.46–35.75	122.3	11
20	🇭🇺 Hungary	29.41	8.35–30.01	171.4	34

Top 20 localities by pedestrians per minute (at least 1 analysed hour)

#	Locality	Pedestrians / min	Analysed (h)
1	🇮🇳 Varanasi	126.64	1.0
2	🇨🇫 Bangui	105.51	5.2
3	🇮🇳 Anantapur	103.57	1.2
4	🇵🇰 Dera Ismail Khan	98.21	1.2
5	🇵🇰 Sialkot	98.18	3.4
6	🇳🇵 Kathmandu	97.07	35.6
7	🇲🇬 Toamasina	89.05	1.3
8	🇵🇰 Multan	88.86	9.8
9	🇺🇬 Kampala	87.35	40.1
10	🇸🇱 Kenema	86.35	2.4
11	🇸🇱 Makeni	80.86	1.2
12	🇵🇰 Mansehra	76.74	2.1
13	🇮🇳 Gaya	76.16	3.0
14	🇳🇬 Lagos	73.20	34.3
15	🇳🇬 Ibadan	71.18	5.2
16	🇮🇳 Dehradun	71.17	1.3
17	🇵🇰 Abbottabad	70.47	1.9
18	🇵🇰 Faisalabad	70.15	1.8
19	🇸🇱 Bo	68.89	2.6
20	🇮🇳 Vadodara	68.30	1.3

Figures

Share of each country's footage that has been analysed with YOLO (countries with less than 10 hours of footage are grey).

Pedestrians per minute of footage per country (unique tracked persons; countries with less than 10 analysed hours are grey).

Pedestrians per minute of footage per European country (countries with at least 10 analysed hours; Russia and Türkiye whole).

Pedestrians per minute of footage per country, with a 95% bootstrap interval over its localities (countries with at least 10 analysed hours in at least 3 localities).

How much footage a stable estimate of a locality needs: the width of its 95% bootstrap interval of pedestrians per minute over its segments (as a share of the rate) against its analysed hours (localities with at least 5 segments; segments of one drive are alike, so the true intervals are somewhat wider).

Pedestrians against cars per minute per locality (at least an analysed hour), split at the medians into walkable, busy, car-dominated and quiet streets.

Pedestrians per minute by population of the locality, per continent, with a 95% bootstrap interval over localities (groups with at least 3 localities and an analysed hour).

Road-user mix of the 30 localities with the most analysed footage, each as 100 squares (ordered by the share of pedestrians).

Detected road users per minute of footage, per class, in the 40 countries with the most analysed footage (log colour scale).

Pedestrians per minute of footage per locality, by continent.

Distribution of pedestrians per minute over segments of at least a minute (log scale).

Mix of detected road users per continent.

Mix of detected road users in the 30 countries with the most analysed footage.

Share of two-wheelers (motorcycles and bicycles) among the detected vehicles per country (countries with less than 10 analysed hours are grey).

Detected road users per minute by day and by night, per class and continent (log scale).

Pedestrians per minute of footage by day and by night per country (countries with at least an hour of each).

Pedestrians per minute by the type of vehicle the footage is filmed from: the viewpoint changes what the detector sees.

Detected road users per minute by year of upload (years with at least 10 analysed hours): newer cameras and higher resolutions can change what the detector sees.

Pedestrians per minute of footage per country (at least 10 analysed hours) against country indicators: population, road deaths, Gini index, median age and literacy (from the mapping), GDP per person, urban population, density and jobs in agriculture (World Bank), pedestrians among road deaths, registered vehicles, physical inactivity and obesity (WHO), mean temperature and rainfall (ERA5, World Bank climate portal) and the footage-weighted TomTom traffic index. Dashed lines: medians of equal-sized groups of countries along each indicator.

Pedestrians per minute of footage against GDP per person (countries with at least 10 analysed hours; circle size: population), coloured by whether a country has more than 1.5 times, about, or less than two-thirds of the pedestrians its wealth predicts.

Motorcycles as a share of detected vehicles against GDP per person, coloured by whether a country has more than twice, about, or less than half the share its wealth predicts.

Bicycles per minute against GDP per person: cycling cultures stand out above the line.

Cars per minute against GDP per person: motorisation on the street, relative to wealth.

Road traffic deaths per 100,000 people against pedestrians per minute: countries with more or fewer deaths than their number of pedestrians on the street predicts.

Pedestrian road deaths per 100,000 people (road deaths times the WHO's pedestrian share of them, 2016) against pedestrians per minute: where walking is more or less deadly than the number of pedestrians on the street predicts (not countries where most road deaths are of an unknown road user, e.g. the Philippines).

Pedestrians per minute per US county (counties with at least an hour of analysed footage).

Pedestrians per minute per European region (NUTS 2, and first-level regions of European Russia, Ukraine, Belarus and Bosnia and Herzegovina; regions with at least an hour of analysed footage).

Bicycles per minute per European region: cycling regions stand out.

Pedestrians per minute in Europe, smoothed over about 50 km; grey where there is less than an analysed hour within 100 km.

Walkable street minutes in Europe: the share of analysed minutes with at least 2 pedestrians and at most 10 cars per minute, smoothed over about 50 km.

Walkable street minutes per European country: the share of analysed minutes with at least 2 pedestrians and at most 10 cars per minute (countries with at least 10 analysed hours; Russia and Türkiye whole; the colours end at the 95th percentile, higher shares get the darkest).

Motorcycles as a share of detected vehicles in Europe, smoothed over about 50 km.

Bicycles per minute in Europe, smoothed over about 50 km.

Pedestrians per minute per locality against its population (localities with at least an hour of analysed footage): busy and quiet streets for a city's size.

Pedestrians per minute of footage per locality against its population.

Adding videos to dataset

To add videos to the mapping file by hand, run python add_video.py. It opens a web form (a Flask app on a local port) for one video at a time. Fill in the locality, the state (optional, recommended for the USA 🇺🇸 and Canada 🇨🇦), the country (mandatory, chosen from a list) and the YouTube link, and click Fetch data. While you type the locality, the form suggests localities that are already in the mapping file.

Adding a video to Delft, Netherlands 🇳🇱 (with no state), before fetching its data. While the locality is typed, the form suggests the matching locality already in the mapping file.

Fetch data checks whether the locality is already in the mapping file and looks up the rest. For the video: its upload date and channel from YouTube. For a new locality: its coordinates, population and gross metropolitan product (GDP), the country's population, road deaths, continent, literacy rate, average height, median age and Gini index, and the traffic index (see The secret file for the services used). The video plays next to the form, with the current second of the player, the last second of the video and a map of the locality.

The same video after fetching its data, from before Delft was in the mapping file: with no entry for Delft yet, the form adds it as a new locality. Its coordinates are shown on the map under the video; dragging the marker adjusts them.

Kyiv, Ukraine 💙💛 is in the mapping file (with 569 videos), so the form adds the video to it. This video is in the mapping file already: the form shows its existing segments in red (here from second 1 to 1408) next to the player, so a new segment is added after them or the existing ones can be checked.

For each segment (part of a video), add the Time of day (day or night), the Vehicle the footage is filmed from, the Start time (seconds) and the End time (seconds) (larger than the start), and click Submit. A video can have several segments, each starting at or after the end of the previous one; the values of the video (upload date, channel) and of the locality are kept, so for each further segment only its start, end, time of day and vehicle need to be filled in. All values can be corrected in the form before submitting, and by hand in the mapping file later.

Shortcuts and click events

The form accepts the following shortcuts (when the cursor is not in a text field; they also work with non-Latin keyboard layouts):

A: the current second of the video into "Start time (seconds)".
S: the current second of the video into "End time (seconds)".
D: the whole video: "Start time (seconds)" 0 and "End time (seconds)" the last second of the video. Clicking on "Last second" under the video does the same (and copies it).
Q: "Day" for "Time of day".
W: "Night" for "Time of day".

Proposing segments from YouTube channels

Instead of finding segments by hand in add_video.py, whole YouTube channels can be processed automatically: propose_segments.py downloads each video of a channel at low resolution, proposes segments, localities, time of day and vehicle type, and a curator reviews the proposals on a web page and applies the approved ones to the mapping file. Nothing is written to the mapping before a curator applies it.

1. Channels to process. run_channels.py takes channels from a Google Sheet (the sheet URL is set in run_channels.py): column A is the channel URL, column B its status and column C its country. It works from the bottom of the sheet up, through channels with an empty status, and proposes segments for each in chunks of videos (--chunk, default 15), each chunk in a new process, so changes to the code apply from the next chunk:

python run_channels.py

It runs until stopped (--max-minutes N stops after N minutes) and continues where it stopped when restarted. It cannot write to the sheet: it logs lines starting with SHEET: (set row N to Processing, Processed or Rejected) for the curator to copy into the sheet. Without a country in column C, it uses the channel's country on YouTube. A single channel can also be processed directly:

python propose_segments.py --channel https://www.youtube.com/@CHANNEL --country "Netherlands" --limit 20

The country must be written as in the add_video.py form. --single-city treats all of a channel's videos as one city (no drives from A to B), and --no-download proposes whole videos from YouTube metadata only. Downloading needs yt-dlp, ffmpeg and node; when YouTube asks to sign in, put a cookies.txt exported from your browser in the root of the project. The youtube_api_key in the secret file is needed to list channels (see The secret file).

2. What is proposed automatically. Videos are left out from their metadata when they are shorter than 5 minutes, live, portrait, or walking tours, time-lapses, compilations, crash videos, highway drives or road trips. For the others, a 160x90 copy is analysed at 2 frames per second:

segments: stationary stretches at the start and end are trimmed and stops longer than 3 minutes are cut out; hard cuts (edits), blank frames and title cards end a segment; footage with the camera facing the driver, not facing forward or shaking, and rural or highway stretches (from YOLO detections) are left out; segments shorter than 30 seconds are dropped. Whole videos are left out for picture-in-picture or skipped footage, and a channel where ten videos in a row have skips is rejected;
time of day: night when the sky or the street is dark for at least a minute, with a hint at dusk;
vehicle: the most common vehicle type of the channel's videos already in the mapping (a car, with a note, for a new channel);
locality: from the title, then the description, then the chapters (a video with chapters for several places is split), and otherwise a named place looked up on OpenStreetMap; drives "from A to B" are split into the towns along the route; a town within 30 km of a city with over a million people counts as that city unless it has 250,000 people itself;
re-uploads of a video already in the mapping are recognised and left out.

The proposals of a channel are kept in _output/proposals/<channel>/plan.json, with frames and signals of each video. Videos left out are listed on the review page with the reason:

Videos of the same channel as the Kyiv example of add_video.py, left out automatically: a video under 5 minutes, a walk, and a drive that skips footage (an edit cut).

3. Reviewing. Run python review_hub.py and open http://127.0.0.1:8770: it lists the channels with videos to review, and clicking one opens its review page (also available directly with python propose_segments.py --review _output/proposals/<channel>/plan.json).

The review hub: channels with videos still to decide. For each video, the page shows the player, the title and description, a contact sheet of frames, a route map, and the proposed segments with the reasons for each cut. Per segment, the curator can change the start and end, switch day and night, approve, reject, split it at a time, move it to another locality, or add a new segment; per video, change the vehicle, locality, state and country (with a map for a new locality), approve the rest of the video or reject it. Shortcuts: A sets the start and S the end of a segment to the current time of the player, D takes the whole video, Q and W set day and night, and F puts the current time in the split box.

The proposal for the Kyiv video used in the add_video.py example above: one segment from 0:01 to 23:30 in Kyiv, Ukraine (the locality guessed from the title, the vehicle from the channel's other videos in the mapping), with a contact sheet of frames and the video's description; it matches the segment added by hand (seconds 1 to 1408).

4. Applying. Apply adds the approved segments to the mapping file through the same code as the add_video.py form: the video, its segments, time of day, vehicle type, upload date and channel, and for a new locality a new row with its coordinates, population and other values. Applied segments are marked in plan.json and cannot be changed on the page any more. Overlapping segments are refused, and writes to the mapping file are locked, so several review pages and a running channel batch can be used at once. Do not stop a review page while it is applying.

5. Merging towns into cities. python review_merges.py serves a page at http://127.0.0.1:8772 with towns of under 250,000 people within 30 km of a city of over a million in the same state, side by side on maps: Y merges a town into the city, N keeps it separate. python review_merges.py --apply merges the accepted towns: their videos move to the city and their name is added to the city's other names.

Weekly update of video metadata

mapping_metadata.csv holds the YouTube metadata of every video in the mapping file (title, upload date, channel, views, description, chapters), and crowd_index.jsonl is the same dataset as JSON Lines, one object per segment, built from mapping.csv and mapping_metadata.csv without the free-text fields (titles, descriptions, chapters). Both are refreshed every Monday at 02:00 (Amsterdam time) by release/weekly_metadata.sh, run by cron on a Mac mini at TU/e in a fresh shallow clone made for each run and removed after it:

it resets the clone to the latest main on GitHub;
python -m release.update_info fetches the metadata of videos added since the last run and retries the videos with missing values (offline videos in the mapping keep being retried). Each video is written to the file as soon as it is fetched (a new video is appended as a row, so an interrupted run keeps what it has; a failed retry leaves its row as it was); rows of videos no longer in the mapping are dropped;
python -m release.make_crowd_jsonl rebuilds crowd_index.jsonl;
python -m release.split_large_files split splits a file over 99 MB into NAME_part1.EXT, NAME_part2.EXT, ... (each under 99 MB; CSV parts repeat the header), as GitHub refuses files over 100 MB;
the two files (or their parts) are committed as "Weekly update of video metadata" and pushed. A file over 99 MB is never pushed: the job stops instead. When someone pushed in the meantime, the commit is replayed on top of theirs.

When the files are split, the whole file is not in the repository. release/update_info.py, release/make_crowd_jsonl.py and release/validation_release.py rebuild it from its parts when needed, or run python -m release.split_large_files join mapping_metadata.csv crowd_index.jsonl. The log of the job is in ~/Library/Logs/crowd-metadata.log on the mini. To run it elsewhere, clone the repository with push access and add 0 2 * * 1 /path/to/clone/release/weekly_metadata.sh >> /path/to/crowd-metadata.log 2>&1 to crontab -e (set PY to the Python of the environment if it is not ~/micromamba/envs/crowd/bin/python).

Who added the videos

Videos are added by several people, each in their own copy of the mapping file: mapping.csv is Pavlo's, mapping-olena.csv Olena's, mapping-shadab.csv and mapping-epfl.csv Shadab's and, in general, mapping-NAME.csv NAME's. Videos added to mapping.csv in commits from a contributor's own account (currently Shadab's and Faye's) are credited to that contributor. Each video is credited to the file it first appeared in. The credits are kept in analysis_data/video_contributors.csv: one row per video with the contributor, the time of the commit that added it (UTC) and the mapping file it first appeared in. It was built from the git history with python -m utils.analytics.contributors, and analysis.py (also in the GitHub Action) adds new videos from the commits of the last 60 days, so a video Olena added to mapping-olena.csv before it was reviewed into mapping.csv is hers; videos in no commit yet get the time of the last commit to their mapping file. It is the only record of who added what once the history of the mapping files is gone, so keep it when rewriting the repository's history. A video that appeared in mapping.csv and another file at the same time is credited to mapping.csv, as the other files are refreshed with copies of it. The table counts the videos in the dataset now.

#	Contributor	Videos	Footage (h)	Localities	Countries	First added	Last added
1	Pavlo	72,504	32,310.8	7,554	232	2025-01-19	2026-10-21
2	Olena	4,852	1,755.1	948	90	2025-05-03	2026-10-21
3	Shadab	2,818	1,445.9	499	96	2023-12-17	2026-07-22
4	Margarida	54	24.9	34	4	2025-05-02	2025-06-05
5	Aloysia	38	18.0	17	2	2025-04-29	2025-07-08
6	Faye	42	14.8	29	11	2026-06-30	2026-07-12

Footage in the dataset over time, by who added it (cumulative hours of the videos in the dataset now, by the date they were added).

Contact

If you have any questions or suggestions, feel free to reach out to Md Shadab Alam (md_shadab_alam@outlook.com) or Pavlo Bazilinskyy (p.bazilinskyy@tue.nl).

Licence

Code: MIT License
CROWD dataset: Creative Commons Attribution 4.0 International (CC BY 4.0)

The dataset licence covers CROWD generated structured data, annotations, and derived computer vision outputs. It does not cover the underlying YouTube videos or other third party content.

CROWD does not redistribute the original YouTube videos, extracted frames, or images. Source videos remain subject to the rights of their respective owners and YouTube's Terms of Service.

Name		Name	Last commit message	Last commit date
.github		.github
analysis_data		analysis_data
archive		archive
dashboards		dashboards
figures		figures
logs		logs
readme		readme
release		release
screenshots		screenshots
templates		templates
tests		tests
tools		tools
utils		utils
.gitignore		.gitignore
.jscpd.json		.jscpd.json
.python-version		.python-version
CITATION.cff		CITATION.cff
DATASHEET.md		DATASHEET.md
DATA_LICENSE.md		DATA_LICENSE.md
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
__init__.py		__init__.py
add_video.py		add_video.py
analysis.py		analysis.py
bbox_custom_tracker.yaml		bbox_custom_tracker.yaml
common.py		common.py
common_words.txt		common_words.txt
count_detections.py		count_detections.py
crowd_index.jsonl		crowd_index.jsonl
crowd_splits.csv		crowd_splits.csv
custom_logger.py		custom_logger.py
cut_times.csv		cut_times.csv
default.config		default.config
default.secret		default.secret
helper_script.py		helper_script.py
logmod.py		logmod.py
main.py		main.py
mapping-olena.csv		mapping-olena.csv
mapping.csv		mapping.csv
mapping_metadata.csv		mapping_metadata.csv
mapping_remaining.csv		mapping_remaining.csv
propose_segments.py		propose_segments.py
pyproject.toml		pyproject.toml
refresh_queue.py		refresh_queue.py
review_hub.py		review_hub.py
review_lessons.py		review_lessons.py
review_merges.py		review_merges.py
run_channels.py		run_channels.py
sync_workers.py		sync_workers.py
uv.lock		uv.lock

Latest commit

History

Folders and files

Repository files navigation

City Road Observations With Dashcams (CROWD)

Dataset overview

Citation and usage of code

Repository layout

Getting started

Step 1: Install uv

Step 2: Fix permissions if needed

macOS / Linux

Windows PowerShell

Step 3: Verify uv

Step 4: Clone the repository

Step 5: Install Python 3.12.13

Step 6: Create the virtual environment

Step 7: Activate the virtual environment

Step 8: Install the dependencies

Step 9: Verify the PyTorch backend

Step 10: Create config and secret

Step 11: Run the scripts

Configuration of project

The secret file

Getting a YouTube Data API key

Getting a LocationIQ key (optional)

Getting a GeoNames username

Getting a BEA API key

Getting a TomTom API key

Getting a REST Countries API key

Email settings

Getting a CARTO key (optional)

Example of YOLO output for a video with dashcam footage

Selection procedure

Segments of videos were not selected if frames are skipped

Snippets of videos are not analysed during the movement of camera

Videos are excluded from analysis if the camera is unstable or shaking

Snippets of videos captured in parking areas were excluded from analysis

Videos are excluded from analysis if another video is a part of main video

Verification of segments

Description and analysis of dataset

Description of dataset

Detections in the footage (YOLO)

Tables

Figures

Adding videos to dataset

Shortcuts and click events

Proposing segments from YouTube channels

Weekly update of video metadata

Who added the videos

Contact

Licence

About

Topics

Resources

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages

Step 1: Install `uv`

Step 3: Verify `uv`

Step 10: Create `config` and `secret`

The `secret` file