Research And Application Of Crowd Counting In Static Images Based On Deep Convolution Features

With the increasing attention paid to pedestrian analysis on monitoring scenes,a variety of vision technologies,including pedestrian detection,human attribute recognition,crowd counting and other technologies,have been widely applied in many important fields such as home,security and new retail.Crowd counting technology,which is used for crowded places,has been widely used in shopping malls,stations,vehicles operations,and other scenes because of its higher accuracy and speed than counting with the naked eye.On the one hand,crowd counting can provide guidance for safety warning,for effectively preventing crowd trampling,overload and other potential dangers.On the other hand,it can also bring some economic benefits,such as improving service quality,analyzing customer behaviors,and optimizing advertising and resource allocation.Crowd counting refers to the process of estimating the total number of pedestrians in the video stream or static images through algorithm analysis.In this paper,crowd counting methods in static images are studies and a crowd counting system is designed.The main contents are as follows:(1)Aiming at the problem that the existing global average strategy carries out feature fusion while ignoring changes of local crowd density,a crowd counting method based on the scale adaptive network is proposed.This method is based on Residual Network and Multi-column Convolutional Neural Network,and through the joint learning to train,to solve the problem of insufficient counting performance on the scenes with scale changes on a large scale.Experiments and analyses on the ShanghaiTech and UCF_CC₅0 datasets show that this method has great modeling ability for large-scale scale changes and counting performance.(2)Aiming at the problem that the max pooling is commonly used in convolutional neural networks to regress density maps,which leads to the loss of a large number of important information,a crowd counting method based on the pooling strategy is proposed.In this method,the max pooling is replaced by the stochastic pooling based on index,and pooling layers are stacked to reduce the degree of network overfitting.Experiments and analyses on the ShanghaiTech and UCF_CC₅0 datasets indicate that this method can suppress the overfitting phenomenon caused by the insufficient sample size of public datasets,and shows great ability of model generalization.(3)In order to further enhance the practical application ability of the proposed methods,a crowd counting system suitable for the scenes with deeper depth of field is designed.The system uses the depth information of the scene to segment the corresponding crowd image.People in the near region are counted using an end-to-end target detection method,while people in the far region are counted through estimating the crowd density map with the model of the scale adaptive network combined with the pooling strategy.Finally,combined with the two counting results,the total number of people in the target image is obtained.