A pedestrian detection algorithm based on deep learning

Abstract

In order to solve the problem of crowd counting and crowd density statistics in the security intelligent video surveillance system, this paper adopts the method based on deep learning to optimize the algorithm. This method mainly uses the VGG16 backbone network with 1*1 and 3*3 small convolution Classification and feature extraction of people information in a crowd. In order to reduce the sharp reduction in the number of positive samples after increasing the threshold, and to avoid the situation where using different thresholds during training and testing will cause the performance of the detector to degrade, this paper draws on the cascade structure of the Cascade R-CNN network for input video. The frame images are analyzed and processed, and different IoU thresholds are set at different stages to obtain enough positive samples to reduce over-fitting, and use the multi-task loss function and the Hadamard product to obtain the pedestrian detection network, and output the final number of people. The improved pedestrian counting algorithm in this paper is tested in the public dataset WorldExpo'10 Crowd Counting, and compared with other algorithms to verify the feasibility and effectiveness of this method.