Multiple Camera-based Correspondences of People using Camera Networks and Context Information

This paper proposes a method to correspondence a same person from different views. To design a system with multi vision sensors, we should to solve the correspondence problem. For obtain a correspondence, a proposed method has three steps. The first step is to detect moving objects by background subtraction from multiple background model. The background model is generated from image sequence and it updates by itself in each process. The temporal difference is used jointly to compensate a limitation of multiple background model. The detected silhouettes are divided as individual by labeling when more than two people are detected in an image. The second step is to generate blobs from detected silhouette. The silhouette is segmented as blobs by a criterion with color range and context information of human body. A person is composed of a set of blobs. The blob set is represented as a model by estimation of Gaussian mixture model (GMM). In the final step, a GMM of detected person from a camera is matched with other GMMs from different cameras to identify a same person.