People Counting Across Non-overlapping Camera Views by Flow Estimation Among Foreground Regions

Abstract

Counting the number of people traveling across nonoverlapping camera views generally requires every person who has exited a camera view to be reidentified when he/she reenters another camera view. A typical solution is to detect an individual person exiting or entering each camera view and establish their correspondence based on their visual appearances and the knowledge of the camera topology, transition time between cameras, etc. One of the main challenges is that the appearances of different people can be similar, while the appearance of the same person can vary in different camera views. On the other hand, a recent approach for counting people within a single camera view is “crowd-centric”, which is to extract foreground regions and estimate the crowd density of the regions. Considering that people often walk together with their acquaintances but not with strangers, the reidentification solution can be applied to the foreground regions to reidentify the groups of people. In this case, another problem arises, that is, people sometimes meet or part outside the field of views of the cameras. Thus, a foreground region can have correspondence with multiple foreground regions. Our proposed method handles both of these problems by estimating the flows from the foreground regions exiting the camera views to those entering other camera views based on the confidence levels of their correspondence and the constraints defined by the relationships among their areas. The estimated flows are then summed up to count the people traveling across each pair of cameras.