Estimating human-flow speed for video surveillance by probabilistic stands

There are strong demands to extend existing surveillance systems with monocular cameras to create systems that can automatically detect unusual situations and issue alarms to the administrator. Most public surveillance functions pay attention to human traffic: the speed and direction of human flow. This is because these clues help to identify exceptional situations. In this paper, we propose a method for estimating people-flow speed in video sequences by allocating stands on CG (computer generated) scenes. To effectively handle crowded indoor scenes captured through a common monocular camera, we calculate an actual motion vector with optical flows and a background scene model. Because the location of objects cannot be specified by cross shots, it is not possible to relate a motion vector in a video sequence to that in the real world. To overcome this problem, we infer the motion vector in the real world by scaling the vector in the video sequence. For this purpose, we employ the "scaling factor" which is derived from a probability distribution of camera-distances of CG objects in CG images. We examine the effectiveness of our method using video sequences captured by two train station surveillance cameras.