Research On Crowd Counting Algorithm Based On Deep Convolutional Neural Network

The acceleration of the urbanization process has caused the urban population to continue to grow,and large-scale crowding can easily lead to trampling events,resulting in casualties and property losses.If the population can be accurately counted and corresponding early warning and security measures can be taken,such events can be effectively reduced or avoided.Therefore,the study of crowd counting has extremely important practical significance,and has attracted extensive attention from scholars at home and abroad.However,due to the diversity of crowd distribution in different scenes,uneven illumination,mutual occlusion and scale changes,the task of crowd counting is still challenging.In this paper,in order to improve the accuracy of crowd counting,based on the deep convolutional neural network,the crowd counting algorithm in different density scenarios is studied.The main work of this article is as follows:1.Analyze the detection accuracy of the current target detection algorithm,apply the current mainstream Yolo V3 network architecture to the crowd counting task,use the K-means clustering algorithm to classify the characteristics of different people,and the simulation verifies that the network architecture has high accuracy and robustness for the detection of low-density crowd.2.For the detection of medium and high-density crowd,a network architecture(DCNN)of VGG19 and dilated convolution fusion is designed.The front end uses VGG19's first 12-layer convolutional layer,based on the analysis of the influence of dilation rate on the receptive field,the back end adopts 6 layers of dilated convolution of serrated dilation rate.Verified on different data sets,the DCNN model has achieved good experimental results in the detection of medium and high-density crowd.3.For high-density crowd detection,a crowd counting model(Res DC)based on fusion of deep residual network and dilated convolution is proposed.15 residual units are continuously stacked as the front end,and 3 layers of dilated convolution are used as the back end,and improved it is a "short circuit connection" method.Verified by the data set with different density levels,the Res DC model has higher accuracy and robustness for the detection of high-density crowd.