A New Multi-scale Pedestrian Detection Algorithm in Traffic Environment

Abstract

Pedestrian detection in traffic environment requires high accuracy and speed of the algorithm. Traditional methods can meet the speed requirement, but there is a long gap in accuracy. Traditional methods based on convolution neural network have higher accuracy, but the amount of calculation is enormous. Aiming at the problems of background confusion, there are pedestrian ambiguity and pedestrian multi-scale in pedestrian detection, this paper constructs a pedestrian detection method with higher accuracy and faster speed based on Faster Regions Convolutional Neural Network (Faster RCNN) method, which is the leading method in the field of target detection. We have mainly improved three aspects: (1) The design criteria based on the summary and the scale characteristics of pedestrians; (2) The setting of anchor windows and the way of creating regional networks have been adjusted, and the pooling layer of environmental regions has been added; (3) Fusion of different levels of features acquires more comprehensive features. Then, based on the open-source deep learning framework, the network is implemented on the California Institute of Technology pedestrian data set. The experimental results show that our method improves the detection accuracy by 2.9% and the detection speed by 10.1 frames/second compared with the original Faster RCNN on the same data set.