What it does FastMOT is a Python multi-object tracker that runs YOLO or SSD detection, Deep SORT identity matching, and KLT optical flow tracking together. It targets real-time performance on NVIDIA Jetson boards, with TensorRT-accelerated inference and Numba-optimized algorithms throughout.

The interesting bit The speed trick is refreshingly blunt: run the expensive detector and ReID feature extractor only every N frames, and let a Lucas-Kanade tracker coast through the gaps. The README shows MOTA barely drops (66.8% → 65.1%) when skipping from N=1 to N=5 on MOT20, while frame rates on Jetson Xavier NX range from 18 FPS in dense 80-object scenes to 42 FPS in lighter ones. Camera motion compensation is also included, which the author notes helps where plain Deep SORT fails.

Key highlights

• Supports YOLOv3/v4 variants (including Scaled-YOLOv4) and SSD via TensorRT, with asynchronous inference
• Re-identification using OSNet, with multi-class tracking supported since August 2021
• Most algorithms (KLT, Kalman filter, data association) compiled with Numba for speed
• Docker setup for x86 Ubuntu; install script for Jetson Nano/TX2/Xavier family
• Extensible to custom classes, though you must train your own YOLO + ReID models and wrestle with ONNX 1.4.1 and Darknet-specific TensorRT plugins

Caveats

• Dependency stack is pinned and brittle: Numba == 0.48, CuPy == 9.2, TensorFlow < 2.0 for SSD, and specific NVIDIA driver versions for Docker
• First run is slow due to Numba JIT compilation
• SSD INT8 calibration requires VOC dataset and is unsupported on Ubuntu 20.04

Verdict Worth a look if you need MOT on actual edge hardware and can tolerate NVIDIA lock-in. Skip it if you want a plug-and-play CPU solution or a dependency stack that ages gracefully.