Pedestrian Detection by Multiple Decision-Based Neural Networks

This paper describes an approach of pedestrian detection for onboard application in night driving. Based on single-frame analysis, two-stage method is designed for detecting pedestrians in the cluttered scenes, which are obtained via a normal camera installed on moving vehicle. In the first stage, bright foreground objects are extracted from dim background as candidates. In the second one, we cascade different-feature-based classifiers, emphasizing shape-based classification. Novel contributions of this paper are, 1) developing the shape representation of candidates; 2) combining multiple Decision-Based Neural Networks for elaborate classification, and further reducing the false alarms. Experiments show that our approach is promising, while the system can achieve real-time detection.