Silhouette coverage analysis for multi-modal video surveillance

In order to improve the accuracy in video-based object detection, the proposed multi-modal video surveillance system takes advantage of the difierent kinds of information rep- resented by visual, thermal and/or depth imaging sensors. The multi-modal object detector of the system can be split up in two consecutive parts: the registration and the coverage analysis. The multi-modal image registration is performed using a three step silhouette-mapping algorithm which detects the rotation, scale and translation between moving objects in the visual, (thermal) infrared and/or depth images. First, moving object silhouettes are extracted to separate the cal- ibration objects, i.e., the foreground, from the static background. Key components are dynamic background subtraction, foreground enhancement and automatic thresholding. Then, 1D con- tour vectors are generated from the resulting multi-modal silhouettes using silhouette boundary extraction, cartesian to polar transform and radial vector analysis. Next, to retrieve the rotation angle and the scale factor between the multi-sensor image, these contours are mapped on each other using circular cross correlation and contour scaling. Finally, the translation between the images is calculated using maximization of binary correlation. The silhouette coverage analysis also starts with moving object silhouette extraction. Then, it uses the registration information, i.e., rotation angle, scale factor and translation vector, to map the thermal, depth and visual silhouette images on each other. Finally, the coverage of the resulting multi-modal silhouette map is computed and is analyzed over time to reduce false alarms and to improve object detection. Prior experiments on real-world multi-sensor video sequences indicate that automated multi- modal video surveillance is promising. This paper shows that merging information from multi- modal video further increases the detection results.