lighttrack

Why we included this project

LightTrack pairs a top-down pose estimator with a graph-based matching module, so detected skeletons get linked into consistent tracks across frames instead of being re-identified every frame. That makes it a practical reference for anyone building real-time multi-person pose tracking: the repo includes working demos for live webcam and arbitrary video, scripts to reproduce PoseTrack benchmark results, and training code for both the pose estimation and matching modules. Because the pieces are modular, you can swap in a different pose estimator or tinker with the data association logic, which is what makes it attractive for researchers and engineers working on surveillance, sports analytics, or other human-centric video applications. One honest caveat: this is research code from 2019-2020, so expect to adapt dependencies before it runs cleanly in a modern stack.