Real-time pedestrian detection system with novel thermal features at night

Abstract

Since the need of pedestrian detection at night has gained more and more interest, real-time pedestrian detection system at night has begun to attract significant attention. However, the detection performance of traditional pedestrian detection system at night is still insufficient since region of interests(ROI) generation and feature extraction are designed separately and less relative. This study presents a novel pedestrian detection algorithm and the corresponding system architecture for the real-time pedestrian detection system at night. The presented pedestrian detection algorithm uses novel outline features which have strong representative for the forward direction of pedestrian. A ROI generation method which is closely related to the outline features is presented to improve the detection accuracy. A three-layer back-propagation feed-forward neural network is applied as the classifier. Since the system architecture is efficiently and the calculation of the outline feature values is simple, the presented pedestrian detection system can operate in real-time. Experimental results show that the presented outline features are significantly effective and the detection performance of pedestrian detection system at night is improved.