Research On Crowd Counting Algorithm Based On Multi-level Convolutional Neural Network

With the significant improvement of people's material living standards in China,the largescale aggregation of people in public places has become the norm,which has caused serious security risks.Nowadays,intelligent video monitoring equipment is widely used in real-time monitoring of crowd gathering occasions,which provides a lot of effective information for crowd behavior control and emergency warning.Among them,the number of people is an important decision-making information for public security management,so it is of great practical significance to research on the crowd counting algorithm.In the process of crowd counting in video surveillance image of public scene,the algorithm of crowd counting based on convolutional neural network(CNN)can automatically extract and combine the effective features in the image.However,in the practical application,the change of shooting angle,distortion of perspective distortion and individual difference of pedestrians make the characters in the image show multi-scale shape,which leads to the performance of the algorithm to decline.In the previous research work,it is proposed to use multi-column network structure with scale perception ability to realize the adaptive detection of different scale characters.However,the structure of multi column network is complex,the parameters are large and the training is difficult,which brings some limitations for the wide use of the algorithm.At the same time,setting too many down-sampling layers in the network structure reduces the image resolution and affects the accuracy of small target counting.Therefore,this paper proposes a population counting algorithm based on multi-level convolution neural network.This algorithm uses a single column deep convolution neural network structure with multi-level feature fusion.The front-end part draws on the design of VGG-16 structure,fully extracts the shallow semantic information in the image and combines and abstracts the high-level features.Through the fusion of multiple network layer output feature maps,it enriches the scale information and makes the model have multi-scale sensing ability.At the same time,it uses dilated convolution instead of pooling to increase the receptive field,which avoids the loss of detail features caused by over down-sampling in deep network,resulting in false detection and missed detection.Compared with the previous algorithms based on multi column network structure,the algorithm described in this paper has significantly reduced the counting error and faster model convergence speed in training.At last,a new public surveillance camera angle crowd data set is made to test the mobility of the training model,which verifies the effectiveness and practicability of the algorithm in different scenes.