Tracking multiple persons under partial and global occlusions: Application to customers' behavior analysis

Multiple objects (targets) tracking plays an important role in computer vision. It is considered as the first step in many artificial intelligence applications that are developed to analyze people behavior for either security or statistical purposes. The most important challenge faced by algorithms designed for multiple objects tracking is the identity switches that occur between tracked objects due to occlusions and interactions between these same objects. This work falls within the scope of video-based behavioral marketing analysis and aims to better understand the purchasing behavior of customers by analyzing their movements in a densely-populated sales area. We propose to use a re-identification strategy to prevent these identity switches. This re-identification strategy is based on segmenting detected individuals into head, torso, and legs in addition to the classification of their appearances into front and back poses. This re-identification module is integrated within our tracking system to fuse tracklets obtained from a particle filter based tracking framework in a mono-camera tracking system. The combination of these tracking and re-identification modules allows the recovery of global trajectories for tracked individuals.