Real-Time Drone Tracking (Learning-Free)

This project tracks a drone in real time using any learning-free approach. The tracker receives a bounding box in the first frame and must predict the box for each next frame using only visual information from previous frames.

General-purpose pre-trained models are allowed, but no drone-specific training or smoothing methods should be used. Any programming language.

Input/Output:

The script takes as input:

• a video file
• a bounding box for the first frame

and outputs a new video with predicted bounding boxes for all frames.

Performance goal: about 100ms per frame on a laptop, >50 FPS on Jetson Thor.

Example usage:

uv run track.py –video test_track_dron1.mp4 –bbox 120,80,60,40

Test file

The sample video: https://drive.google.com/file/d/1HJCNYOLodnfECc_sIJw9kIgjwGkzGrRi/view?usp=sharing

Communication

• It's totally fine to double-check with Roman after doing some research and choosing an approach (e.g., "I'm leaning toward X because Y - does that sound reasonable?").
• If anything is unclear (constraints, allowed tools/models, smoothing assumptions, metrics, etc.), please ask.