Multi-Target Multi-Camera Tracking with Spatial-Temporal Network
Yi Gao, Wanneng Wu, Ao Liu, Qiaokang Liang, Jianwen Hu · 2023
Multi-Target Multi-Camera Tracking (MTMCT) aims to link objects of interest between frames and different camera views, which has many application potentials in the fields of surveillance and autonomous driving. Existing methods usually extract motion or appearance features from single-view trajectories to perform cross-view association, but struggle with camera view variations and occlusions. To this end, this paper proposes a novel method that integrates spatial-temporal cues for more robust cross-view matching. The proposed method firstly introduces a spatial-temporal trajectory selection module to enhancing feature saliency and reducing the impact of erroneous trajectories on feature matching and then extracts discriminative video-level appearance features for each trajectory with a temporal complementary learning network. Cross-view targets are finally associated by computing the cosine distances of the trajectory features. Experimental results on the public DIVOTrack dataset validate the effectiveness of the proposed method.