Video Based Real-Time Pedestrian Detection on Zebra Cross

This paper proposed a novel algorithm for real-time pedestrian detection based on the road user trajectory recognition. Frame series captured from the video stream are used to compose the object trajectory presenting basic features including width, speed, and length. The triangular fuzzy numbers are used for the feature description and decision reference. The multi-variable decision tree is applied for the decision-making. The proposed framework and algorithm are implemented at the junction of Eshan, Dongfang Road, Shanghai for pedestrian alerting system in red light time. The accuracy over 93% is achieved with single decision iteration time less than 100 ms.