Research On Crowd Counting Algorithm Based On Multi Branch Convolutional Neural Network

In the development of the field of image processing,convolutional neural networks have a significant historical position because they can directly convolution multidimensional images.The model can obtain a large number of efficient features from images through convolutional kernels,avoiding the complex artificial feature extraction process.In the field of crowd counting,convolutional neural networks are more widely used.In order to effectively solve the pain points in crowd images,such as sharp changes in head size,interference from different backgrounds,mutual occlusion in the crowd,and the same number of people but different population distribution,current mainstream research schemes have adopted a deep neural network or multiple columns using convolutional kernels of different sizes.With the deepening of the convolutional layer,other issues have also been introduced.With the increase in training time,the characteristics learned in each column of the multi column convolutional layer are extremely similar.This article proposes some solutions to solve the above problems,and the specific content is as follows:1.Aiming at the problem of pedestrian coordinate deviation caused by crowd density estimation,this paper designs a multi column convolutional neural network structure to more accurately predict the location of pedestrians,rather than simply predicting the number of pedestrians.2.Aiming at the problems of long training cycle and large computational complexity of multi column deep convolution neural network models,this paper adopts a front-end single column architecture.In response to the problem caused by the gradual reduction of the feature map generated by the image as the convolutional neural network deepens and the pooling layer increases,this paper adopts a method of hollow convolution to equivalently replace the convolutional layer and the pooling layer.This is an upsampling method,which to some extent avoids the problems caused by the distortion of the feature map.