3D pedestrian tracking based on overhead cameras
Zhognchuan Zhang, Fernand S. Cohen · 2013
This paper proposes a method to track pedestrians in crowded scenes based on the detection of the 3D head position of a person using two overhead cameras. A possible head area in one frame acquired from one of the overhead cameras is determined by evaluating a head area existence probability based on the integral polar mapped image, where a foreground pixel is assigned a probability to belong to the head area. A segment passing through the head top is estimated for each clustered head area. The disparities along each segment are calculated using the synchronized frame from the other overhead camera. The center of the points with the largest disparity on the segment is determined as the head point and its 3D position is computed using triangulation. It is then tracked using common assumptions on motion direction and velocity. This is efficiently done notwithstanding the fact that several segments may exist in a single foreground blob with each segment corresponding to a different person. The approach is tested using a publicly available visual surveillance simulation test bed. The experiments show that the 3D tracking errors are around 5 cm. The method allows for the capture of high quality close-up facial images.