An Online Approach and Evaluation Method for Tracking People Across Cameras in Extremely Long Video Sequence

Abstract

Multi-camera Multi-Object Tracking has drawn significant attention in recent years due to its critical role in surveillance, analytics, and related fields. Various challenges, including non-overlapping regions, varying occlusion conditions, and the need for cross-domain generalization in multi-camera tracking systems, remain unsolved in the field. We propose a novel online tracking framework that capitalizes on real-time camera calibration to achieve consistent multi-object tracking across camera networks. Our approach seamlessly integrates spatial and temporal association techniques, ensuring robust tracking even in long-duration videos. However, standard tracking evaluation metrics like CLEAR or HOTA fall short of accurately interpreting the performance of tracking over extended video sequences. Another contribution of this study is the proposal of a new evaluation metric, mHOTA, which provides a better assessment of tracking performance over prolonged periods. Our comprehensive experiments on the AIC24 Multi-Camera People Tracking dataset demonstrate the effectiveness and scalability of our method, along with the capability of the proposed evaluation metric. The code will be available at https://github.com/ipl-uw/mHOTA.