tello-depth

Real-time monocular depth estimation for the DJI Tello drone, using LibreYOLO depth models. The drone is piloted from the keyboard while the camera feed is displayed side-by-side with the depth map, and the center-region distance is measured.

Features

• Keyboard piloting of the Tello (takeoff, land, all axes, yaw rotation, emergency stop)

• Real-time depth estimation with LibreYOLO models, running in a separate thread so video and controls stay responsive

• Center-region depth measurement: a square at the center of the frame; the average depth of the area is displayed (higher relative value = closer)

• Metric distance via a saved scale factor, calibrated with a one-point reference (key C) or automatically with the drone's TOF + IMU during a short forward maneuver (key V)

• Camera calibration (config/tello_02.yaml): intrinsics are loaded and lens distortion is corrected before depth inference

• Available depth models (download automatically on first use):

  • LibreMiDaSs-depth.pt — MiDaS Small (EfficientNet-Lite3 encoder, 256 px input). Fastest, lightest quality. Uses extra dependency timm (pip install 'libreyolo[midas]')
  • LibreDepthAnythingV2s-depth.pt — Depth Anything V2 Small (ViT-S + DPT head, 518 px input). Better quality, moderate CPU cost (default)
  • LibreZipDepthb-depth.pt — ZipDepth Base (lightweight CNN distilled from Depth Anything V2 Large, 384 px input). Fast, good quality
  • LibreMiDaSl-depth.pt — MiDaS Large (DPT-Large, 384 px input). Much better, heavy on CPU (~1.3 GB weights)
  • Depth Anything V2 Base/Large/Giant and Depth Anything 3: not redistributed by LibreYOLO (CC-BY-NC license on Base/Large/Giant); they require manual conversion of the upstream checkpoints

  All models output relative inverse depth (no metric unit): higher values mean closer. Select the model in main.py via DepthEstimator("<checkpoint>").

Installation

python -m venv venv
source venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu   # CPU build (skip if GPU)
pip install "libreyolo[midas]" libtello-zune opencv-python numpy

Usage

1. Connect to the Tello Wi-Fi network.
2. Run:

python main.py

Keyboard controls

Key	Action
T	Takeoff
L	Land
W / S	Forward / Backward
A / D	Left / Right
R / F	Up / Down
Q / E	Rotate counter-clockwise / clockwise (yaw)
Space	Emergency stop (motors cut immediately!)
C	Instant scale calibration: point the center square at an object at KNOWN_DISTANCE_M meters (see main.py) and press
V	Automatic calibration: drone verifies TOF height, flies ~50 cm forward while IMU/odometry measures the displacement, then computes the metric scale
ESC	Exit

The metric scale is saved to config/depth_scale.json and reloaded on the next run, so calibration is only needed once per depth model.

Project structure

main.py                  # Main loop: video capture, overlays, keyboard control
keyboard_controller.py   # Keyboard mapping and RC commands to the drone
depth_estimator.py       # Depth inference thread (LibreYOLO) + center-region averaging
camera_calibration.py    # Camera intrinsics from config/tello_02.yaml + metric scale storage
motion_calibration.py    # Automatic scale calibration via TOF + IMU forward maneuver
config/tello_02.yaml     # Camera calibration (intrinsics + distortion, 960x720)

Notes

• The Tello's TOF sensor points downward (measures height above the ground); it cannot measure what the forward camera sees. It is used only as a safety check before the automatic calibration maneuver, and its reading is displayed on screen as Height.
• Monocular depth models have an inherent scale ambiguity; metric conversion requires a scale factor, which the C and V calibrations provide.
• Inference runs on CPU. Resolution can be changed with tello.set_image_size((width, height)) in main.py; lower resolutions improve display/keyboard FPS (the model letterboxes internally anyway, so inference time barely changes).