FastMOT

Why we included this project

FastMOT earns a look from teams that need multi-object tracking to hold up in real time on constrained hardware, particularly NVIDIA Jetson devices. It pairs a YOLO or SSD detector with Deep SORT-style association and an OSNet re-identification model, but the interesting part is how it skips detection and feature extraction for several frames and lets a KLT tracker cover the gap, with camera motion compensation for moving footage. That sidesteps the usual bottleneck in two-stage trackers, where detection and feature extraction run in sequence and drag the pipeline down. If you are building a people counter, crowd analytics, or video surveillance feature that must keep stable IDs while running on edge GPUs, this is a practical reference worth studying or adapting. The code is also a clean example of wiring TensorRT with asynchronous inference and Numba-optimized tracking logic.