Research On Pedestrian Flow Detection Technology For Video Surveillance

Pedestrian flow detection technology for video surveillance is mainly for automatic statistics of the number of pedestrians in video and images,and can reflect the density distribution of pedestrians in video images.By using pedestrian flow detection technology,it is possible to optimize road design,avoid trampling accidents and ensure social safety.This dissertation proposes two detection techniques for different pedestrian density characteristics in crowd scenes.The main tasks completed in this article are as follows:(1)For sparse target scenes,it relies too much on traditional pedestrian target detection and tracking algorithms,and cannot fully utilize the information in the video and the complexity of the detection environment.Therefore,this dissertation will propose a pedestrian flow detection method based on convolutional neural network and video motion information.The Faster R-CNN network is combined with a hybrid Gaussian background modeling method to fuse all its pedestrian detection result calibration frames and pass The non-maximum suppression method optimizes the counting of the pedestrian detection calibration frame and eliminates the result of redundant pedestrian numbers.Compared with other convolutional neural network algorithms in the Caltech dataset of video data,the accuracy of the algorithm model is increased by 20%.(2)In a dense target scene,a series of problems such as overlapping,occlusion,and merging of pedestrians caused by excessive pedestrian density.This dissertation will introduce a calibration density map,and at the same time propose a pedestrian flow statistical method based on convolutional neural network.This neural network is divided into a front end and a back end.The front end uses VGG-16 with the fully connected layer removed,and the back end uses expansion convolution with different expansion rates.Compared with other algorithms,the overall performance of this algorithm in Shanghai Tech dataset and UCF?CC?50 dataset has been improved to different degrees.