Pedestrian Tracking Through Coordinated Mining of Multiple Moving Cameras

Yanting Zhang, Qingxiang Wang; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 252-261

Multiple object tracking has attracted great interest in the computer vision community. Most researchers focus on the applications under a single static or moving camera. In intelligent cities, tracking across multiple static cameras is also investigated due to the need for surveillance purposes. With the growing development of autonomous driving, it is critical to correlate all the vehicles' vision systems on the road to achieve a global perception. However, tracking across multiple moving cameras has not been well studied yet. We observe a lack of such a publicly available dataset for coordinated mining of multiple moving cameras. In this paper, we aim to bridge the gap and propose a new dataset of multiple moving cameras, called "DHU-MTMMC", in which the videos are collected from several cameras mounted on the moving cars. The dataset contains fourteen sequences in different scenarios with annotated pedestrians. We propose a baseline MTMMC workflow to deal with tracking pedestrians across cameras. When the joint detection and embedding are performed, the association algorithm can run online under single-camera settings. We treat multi-camera tracking as a linear assignment problem that can be solved efficiently. The overall IDF1 of the proposed MTMMC tracking on the dataset is 57.8%.