Real-time pedestrian detection in urban scenarios

A real-time pedestrian detection system is presented that runs at 24 fps on standard VGA resolution input images (640×480px) using only CPU processing. The detection algorithm uses a variable sized sliding window and intelligent simplifications such as a sparse scale space and fast candidate selection to obtain the desired speed. Details are provided about the initial version of the system ported on a mobile device. We also present a new labeled pedestrian dataset that was captured from a moving car that is suitable for training and testing pedestrian detection methods in urban scenarios.