A microcontroller is too small to run object detection, so this system splits the work: a GPU machine runs the vision model, and an ESP32 acts on the results.
Highlights
- YOLOv8 inference accelerated with TensorRT
- Multi-threaded, zero-copy UDP pipeline between the inference host and the ESP32
- Under 10 ms latency at 60+ FPS end to end
Stack
Python, PyTorch, TensorRT, OpenCV, network sockets, ESP32 firmware.