SC-Track: State Transition and Constrained Non-Negative Matrix Factorization for Multi-Camera Multi-Target Tracking

Abstract

With the rapid development of intelligent transportation systems and video surveillance technology, multi-camera multi-target tracking (MCMT) plays a significant role in public safety and traffic management. However, challenges such as target occlusion, abrupt changes in target motion states, and maintaining accurate global tracking of the same target across different cameras remain major issues in the current MCMT field. To address challenges like nonlinear motion, severe occlusion, and imperfect detection, this paper proposes a novel tracking strategy by incorporating state transition strategies into the tracking framework to enhance tracking performance. Specifically, it includes three state transition tracking strategies: bounding box selection before and after state changes, trajectory reconnection during the transition, and bidirectional tracking in both forward and backward directions. Additionally, this paper utilizes constrained non-negative matrix factorization to achieve global trajectory consistency matching at the trajectory level, ensuring consistent target IDs across cameras. Experimental validation on the CityFlowV2 dataset demonstrates the proposed method’s effectiveness in addressing occlusion, state changes, and cross-camera trajectory matching.