Structure maps based pedestrian detection

Pedestrian detection is widely needed in many applications, such as people counting, surveillance and scene understanding. In this paper, we propose a robust stereo vision based approach using novel structure maps of static scene, which can be generated from the disparity image and represents the geometry profiles of 3D objects in pixels. Both of the background subtraction and pedestrian detection are taken on the maps only, this decreases the computational cost greatly compared with general reprojection based methods. Due to the maps retain the shape of pedestrian in color image, our detection can precisely locate the head of people and robust to any view angles. The experimental results show that the approach works well in challenging conditions, and can be used in a real time system for various applications.