Drone Real-Time Rescue Detection

I want to upgrade my drone so it can automatically spot and keep track of people and animals in real time while flying over forested terrain during search-and-rescue missions. The live 4K video stream comes down from a gimbal-stabilised camera and is processed on an onboard NVIDIA Jetson Xavier NX. I need a robust deep-learning pipeline (e.g., YOLOv8 + DeepSORT/ByteTrack in PyTorch or TensorFlow) that can:

• detect people and common wildlife through partial canopy cover, variable light and motion blur
• maintain unique IDs as targets move in and out of view
• draw bounding boxes and labels on the feed at a steady 30 fps with end-to-end latency under 200 ms

The system has to fuse camera pose and GPS so each detection can be translated into lat-long coordinates, then publish that data over ROS 2 (MAVLink bridge already running) to my ground station software.

Please containerise the solution so I can deploy with a single Docker command on the Jetson. Along with the code, I’ll need:

1. A brief model-training notebook (or script) that shows how to fine-tune on additional forest footage I’ll provide.
2. A read-me covering setup, dependencies, and startup flags.
3. A five-minute demonstration video that proves the tracker works on a test flight or annotated replay.

I already have datasets and flight logs ready to share once the project starts, and I’m happy to run field tests while you iterate. If this sounds like your kind of challenge, let’s talk timelines and milestones right away.