Automatic cooperative camera system for real-time bag detection in visual surveillance

Visual surveillance systems use more and more cameras in order to cover wider areas and reduce blind spots. Cameras placement and configuration usually depends on the area to be monitored and the size of objects in the scene. Video analytics systems also require a minimal size to get detailed features of objects or people. Most vision-based surveillance systems focus on detection and tracking of people or objects in the scene. However, it is often more meaningful to describe people with high-level information such as hair style, carrying bag or other attributes. In order to perform this detection a close view is required. In this paper, a collaborative camera pair system tackles this problem and retrieves detailed features in a wide scene following the master-slave approach. A PTZ (Pan-Tilt-Zoom) camera is defined as slave and zooms on the targets detected by the master camera with wide coverage. We introduce an automatic method to estimate the internal camera parameters in order to have an efficient control of the camera pair combined with a novel real-time bag detection algorithm. Targets are first identified in the master camera and the slave camera will zoom in to the targets to detect different types of bags. Experimental results will be shown on real-data at each step of the approach.