People Counting Based on Head Detection and Reidentification in Overlapping Cameras System

Abstract

People counting is one of the key tasks in video surveillance system. Usually, one camera cannot cover all the area for a big room or plaza, so multiple cameras will be used. For overlapping cameras system, the overlapped area should be marked for people counting. To get the area, an easy way is to calibrate the cameras. But in real practical system there are hundreds of cameras, it’s impracticable to calibrate all the cameras. A practicable way is to get the overlapping area is to applying image mosaic algorithm. Using local features such as SIFT cannot to find the overlapping area due to corresponding low quality and repeated similarity (such as seats and tables) in most surveillance environment. In this paper, we are committed to finding repetitive people to improve the accuracy of the population statistics. First, a person head detector is trained to detect the human head in the frame taken by each camera video and then cut it, and the images of the head pair taken by each camera are taken as a database which belongs to this camera. Then, we select one database as Gallery, image from another database as the Probe, we use a Siamese networks to match a probe with Gallery, repeating the above process until all the iterations of the probe in the database are completed, so that we find all the recurring people between the two cameras. Finally, the median of the total number of people in all video frame image counts is calculated, and the median result is taken as the final statistical result of the scene.