Research On Pedestrian Flow Statistics Algorithm Based On Deep Learning

In recent years,computer vision technology has matured,and its use in the field of intelligent monitoring has become more and more extensive.A lot of the work that was supposed to be done manually can be replaced by a visual algorithm,which greatly reduces the cost of labor.In the field of smart surveillance,there is a wide range of demand and applications for pedestrian traffic statistics,such as malls,schools,train stations,and smart buildings.Therefore,it is necessary to devise a smart computer vision based pedestrian traffic statistics algorithm.On the other hand,if you can apply the recent development of deep learning techniques to them,it will greatly improve the performance of the algorithm.This paper studies and designs a method of pedestrian flow statistics based on deep learning.This paper applies the method of target detection algorithm based on deep learning,single target tracking algorithm and data correlation algorithm,and a framework is designed for the detection-tracking-correlation algorithm,which is used to complete the statistics on the traffic to the monitoring video.This paper has done the following research work:Firstly,this paper explores the application background of pedestrian traffic statistics and expounds the significance of research.Then the development status of pedestrian flow statistics and the target detection algorithm based on deep learning are analyzed.Finally,the main research content and research scheme of this paper are expounded,and the overall algorithm and framework of the design are given in the research project.In this paper,the structure and optimization of convolution neural network are studied.Then we reviewed the traditional target detection algorithm and the target detection algorithm based on the convolution neural network in recent years.Finally,the SSD algorithm is used as the object detection algorithm in pedestrian flow statistics.Next,this article examines the framework and principles of SSD algorithms,including network structure,selection of default boxes,and training target functions.Then,we focused on the base network in SSD.Based on the SSD primitive network VGG and the popular CNN network structure zf-net and SqueezeNet,the two base networks are redesigned to compare with VGG.Finally,we determine to use VGG as the base network.After the completion of the study of detection algorithm based on convolution neural network,this paper also studied the need to use the tracking algorithm,data correlation algorithm and trajectory analysis algorithm.The KCF algorithm is used as the tracing algorithm,and the trace library in OpenCV is used directly.The correlation algorithm is a simple and fast based correlation algorithm.Finally,the algorithm of trajectory analysis was designed to realize the counting of the two sides.By the end of the work,the algorithm is complete.Finally,this paper expounds the actual operation,including the erection of cameras and the sample video acquisition,detection,image data sets of production,the training of the SSD detector,the realization of the tracking algorithm,correlation and path analysis design key points of the algorithm.Then the paper uses the design's pedestrian traffic statistics algorithm to analyze all the samples from video.Some performance indexes are used to evaluate the performance.The average recognition rate was 96.24% and the average error detection rate was 2.19%.The average leakage rate was 3.76% and the total video frame rate is 24.09.The proposed algorithm can meet the requirements of the project.