A Research Of Crowd Behavior Analysis Based On Deep Learning

With the increase of urban population and the intensification of population migration,a large-scale crowd gathering phenomenon appears in many public places,which leads to a series of group safety accidents.Therefore,it has become a new research hotspot to use computer vision methods to analyze the crowd behavior in public surveillance video.And crowd behavior analysis is also of great significance in smart city construction,shopping mall planning,transportation planning and other fields.Crowd behavior analysis technology can be divided into low-level feature analysis and high-level feature analysis.The former is mainly to estimate the crowd density,and its goal is to generate a crowd density map containing crowd distribution information from a photo containing dense crowds,as well as to obtain the number of pedestrians.The difficulty of this problem lies in the missing and wrong detection caused by the occlusion between people and between people and objects,as well as the change of perspective makes the front and back scales of the same image change too much,which requires a high fitting ability of the algorithm.The latter is mainly about crowd behavior understanding,and its goal is to understand the crowd behavior of the video clips containing crowd behavior and detect crowd abnormal behavior.The difficulty of this problem is that it is difficult to clearly define the normal and abnormal behavior of the crowd,and it is difficult to extract the spatiotemporal information of the video clips,and the lack of the crowd abnormal behavior data makes the dataset extremely unbalanced,causing the algorithm to have learning bias.For crowd density estimation,this thesis proposes a crowd counting algorithm EFCCNN based on convolutional neural networks,which can generate high-quality crowd density map from dense crowd photos and achieve accurate crowd counting.For crowd behavior understanding,this thesis proposes a crowd behavior understanding algorithm based on 3D convolutional neural networksa,which can accurately detect the video with crowd abnormal behavior.The main work of this thesis is as follows:1.This thesis proposes a crowd count model based on convolutional neural network,which combines multiple receptive fields branch network and SENet enhanced channel structure.In multiple receptive fields branch network,the ability of information capture for different size of human head is imporved by setting the receptive fields of different columns.In SENet enhanced channel structure,by reweighting the channel,important feature channels are enhanced and unimportant channels are weakened.Then the residual connection is used to transfer the information,which is helpful to solve the problem of occlusion and scale change,and improve the accuracy of crowd counting.The ablation experiment results on the ShanghaiTech dataset part B show that using the SENet structure on the basic network reduces the average absolute error index of the crowd counting by12.7%.2.In this thesis,a new loss function is designed for crowd count algorithm.The loss function focuses on the quality of the generated crowd density map,which helps to solve the problem of inaccurate counting under different backgrounds,and improves the quality of the crowd density map and the accuracy of the crowd counting.The ablation experiment results on the ShanghaiTech dataset part B show that using the new loss has increased the average structure similarity index of crowd density map by 14.3% and reduced the average absolute error index of crowd counting by 14.2% compared to the old loss function on the basic network.3.This thesis uses EFCCNN to perform ablation experiments on the ShanghaiTech dataset to verify the effect of SENet structure and new loss function.EFCCNN is trained and tested on ShanghaiTech,Mall and UCF_CC₅0 public crowd counting dataset,and the results of EFCCNN are compared with that of other crowd counting methods to verify the effectiveness of EFCCNN.In order to further demonstrate the design idea of the algorithm,this thesis visualizes the EFCCNN and verifies that the structure conforms to the design idea.In addition,EFCCNN is tested in the actual scene to verify the generalization and practicability of EFCCNN.4.Based on the definition of crowd abnormal behavior,this thesis proposes a crowd behavior understanding algorithm based on fibrotic 3D convolutional neural network under unbalanced data.It takes crowd density flow map as input and uses fine-tuning for transfer learning to complete the detection of crowd abnormal behavior.In the implementation process of algorithm,based on the UMN dataset,this thesis manually labels 774 crowd density maps containing crowd density distribution and crowd behavior labels,and then uses EFCCNN to generate the crowd density flow map of UMN dataset through transfer learning.Furthermore,the algorithm of crowd behavior understanding is trained and tested,and the results is compared with several other methods to verify the effectiveness of this method.