Multi-camera Tracking Based on Spatio-Temporal Association in Small Overlapping Regions

Multi-camera tracking (MCT) aims to track people across multiple cameras. To match tracks across cameras, existing MCT solutions primarily rely on Person Re-Identification (Re-ID) that compares people's visual appearance. However, this approach fails to match people with very similar appearances, such as people wearing uniforms in workplaces. In this paper, we propose a method based on spatio-temporal association (STA) to overcome the limitations of visual-based Re-ID in the problem of similar-appearance MCT. Our proposed method operates effectively when there are (even small) overlaps between cameras and a moderate number (i.e., maximum from 4 to 7 individuals) of people moving closely to each other in each overlapping region. We evaluate our proposed method on our prepared private dataset and the PETS2009 public one. The experimental results show that our proposed method matches people appearing in multiple cameras correctly and outperforms the MCT based on visual Re-ID method in case people have similar appearances, and it works well even if the overlapping region is small. To further strengthen the proposed method, we perform error analysis and introduce three extensions to mitigate the problems of missing detections and inaccurate footpoint interpolation. These three extensions further improve our proposed baseline method accuracy of the matching at frame level.