Research On The Method Of Crowd Counting In Public Places Based On Deep Learning

With the development of the times,population statistics have important application requirements in many fields.When a large number of people gather in a public place,it may be crowded due to too many pedestrians,and safety accidents such as trampling may occur.In order to maintain the safety of public places,it is necessary to monitor the number of people in public places for timely warning.In addition,statistical analysis of the number of people in public places is also of great significance in commercial construction,urban planning and construction.The current population statistics subject is facing many challenges,especially the effect of crowd counting methods in sparse and dense cases is quite different.Aiming at sparse and dense crowds with two different densities,this paper proposes to use two crowd statistics methods based on pedestrian head detection and density map regression to realize the crowds counting of different densities in public places.(1)For the crowd statistics method based on pedestrian head detection,this method is mainly suitable for sparse crowd occasions or occasions where pedestrians need to be located to provide their location information.In view of the difficulty of detection caused by the serious occlusion of the torso of the crowd,we choose to detect the head of the pedestrian to reduce the impact of the torso occlusion;because the features extracted by the shallow network have better detail localization capabilities for small-sized targets,it is a new feature detection layer is added to the network to improve the ability to detect small-sized pedestrians;an efficient channel attention module is introduced into the feature extraction network to enhance the network's attention to pedestrians;in the face of the interference of complex backgrounds on crowd recognition,multi-scale feedback is used to improve network interaction pedestrian's ability to recognize non-pedestrian objects with similar characteristics.(2)For the population statistics method based on density map regression,this method is mainly suitable for dense populations.In order to improve the network's counting performance of dense crowds,multiple hollow convolution modules are introduced into the network to expand the network's receptive field;for the pedestrian scale changes drastically caused by the perspective effect,and the context perception module is used to learn the characteristics of the receptive field at multiple scales;use bayesian loss function instead of Euclidean loss function;finally,a high-quality crowd density map is generated to realize the crowd count of the image.Analysis of experimental results show that,for sparse population statistics,the proposed method of crowd statistics based on pedestrian head detection has improved the accuracy of the three datasets.The highest accuracy rate of 89.44% is obtained on the SCUT-HEAD-A dataset,which is 4.24% higher than the YOLOv3 algorithm.For dense population statistics,the proposed population statistics algorithm based on density map regression has achieved good counting results in multiple public population datasets.Compared with most mainstream methods,both MAE and MSE have declined.The project results can provide useful reference ideas for the statistics of other targets.