Region based detection of occluded people for the tracking in video image sequences

Abstract

This paper presents a framework to deal with occlusions when detecting people for the tracking in the image sequences of a stationary surveillance video camera. Unlike the cases of most existing techniques, people are in low-resolution and the detected foreground images are noisy. As the small sizes of target people make it difficult to build statistical shape or motion models, techniques proposed use simple features of the bounding boxes of target people such as position and size. Each foreground region in a bounding box is identified in independent, partially occluded, or completely occluded state, and the state is updated during tracking. Proposed technique is tested with an experiment of counting the number of pedestrians in a scene. © Springer-Verlag Berlin Heidelberg 2005.