In recent decades, more and more cameras have been installed in public and private spaces for monitoring activities and behaviours. While some computer vision systems have been deployed to perform automated analysis and video surveillance, such deployment has been scarce, as few systems allow for automatic large scale wide-area monitoring.  Wide-area tracking can be defined as the capability to track every subject of interest through a camera network, where the cameras have non-overlapping fields of view and are distributed over an unknown and arbitrary layout. This problem entails a number of difficult problems that must be overcome in order to extend video surveillance from a single camera to a large network of cameras, including: different camera configuration, different appearance of targets between cameras, unknown camera layout and unknown building topology, to name but a few.

In this project, we propose a full wide-area tracking for person monitoring, composed of a within camera person detection and tracking, and a between camera video-reidentification system based on deep learning. By reformulating the wide area tracking problem as a re-identification problem, association between the tracks of people moving between cameras can be established regardless of changes of in the person’s appearance caused by viewpoint and camera configurations.  Our proposed architecture is shown in Figure 1.

The following video shows the capability of the system to tack subjects across an entire camera network using the DukeMTMC dataset.