An Efficient Multi-Object Tracking Guided by Spatial Clustering on Vision Sensors

Detection-based tracking is a prominent multi-object tracking (MOT) paradigm that utilizes powerful detectors for high tracking performance. However, this method primarily relies on object prediction, leading to error accumulation when objects are occluded or lost. To address this issue, we propose a novel spatial clustering-guided MOT method (SCGTrack). First, we introduce a fast spatial clustering (FSC) algorithm to efficiently group the detections by their spatial information to form clusters. Second, we develop a spatial clustering matching module to reconnect lost trajectories and re-match trajectories within the clusters. By doing so, SCGTrack can continuously track the lost objects as long as the cluster exists, which effectively mitigates long-term occlusion problems. Extensive experiment results on the MOT17 and DanceTrack dataset demonstrate that our approach achieves state-of-the-art performance in terms of tracking accuracy, with an exceptional matching speed of over 800 FPS on a single CPU. Code is available at
https://github.com/Yamahhh/SCGTrack
.