Automatic video surveillance of outdoor scenes using track before detect

This paper concerns automatic video surveillance of outdoor scenes using a single camera. The first step in automatic interpretation of the video stream is activity detection based on background subtraction. Usually, this process will generate a large number of false alarms in outdoor scenes due to e.g. movement of thicket and changes in illumination. To reduce the number of false alarms a Track Before Detect (TBD) approach is suggested. In this TBD implementation all objects detected in the background subtraction process are followed over a number of frames. An alarm is given only if a detected object shows a pattern of movement consistent with predefined rules. The method is tested on a number of video sequences and a substantial reduction in the number of false alarms is demonstrated.