Research On Crowd Congestion Detection Method In Public Area Based On Deep Learning

A large number of individuals gathering in public areas may lead to the spread of the epidemics,rigger stampedes and other vicious events that affect public security.Therefore,it is necessary to detect and warn the degree of crowd gathering or crowding in public areas,so as to provide a reference for relevant departments to make decisions and take actions.Traditional human observation and detection is not only time-consuming and laborious,but also limited by subjective and objective factors of accuracy.Therefore,it is of great value and significance to realize the congestion detection of crowd in public areas through artificial intelligence and computer technology to improve the efficiency of detection and maintain public security and social stability.As the public scenes are often more than complex,there are a lot of background interference in crowd videos or images,which seriously affects the accuracy of crowd congestion detection.Additionally,the density of the crowd in the actual scene becomes larger.Subsequently,the crowd image changes significantly in scale,and the phenomenon of mutual occlusion between targets is more serious.In view of the existing problems and challenges in congestion detection of crowd,this thesis proposes solutions from the perspectives of scene application and algorithm research.The main research of this thesis is as follows:(1)Aiming at the issue of crowd congestion detection in complex small-scale closed areas,a bus congestion classification algorithm based on occlusion removal was proposed,and the interior of the bus was taken as an application scenario.Due to the complex scene and interference factors in the bus,the question of passenger occlusion is prone to occur.The existing deep learning and object detection methods have poor performance in the application of the interior scene of the bus.To this end,this thesis proposes an effective method of bus congestion classification.Based on the idea of occlusion object detection and removal,the semantic information of the occlusion area is repaired,so as to detect the number of people that relative exact in the bus and realize the classification and analysis of bus congestion degree.The research was generated datasets from data of real bus and annotated them.The effectiveness of this algorithm is proved by the validation experiments on the dataset.(2)Aiming at the issue of crowd detection in large-scale open scenarios,a method for crowd detection in open areas based on multi-task enhancement is proposed.By counting the crowd in the public area and combining with the area of the region,the degree of crowding in the public area can be judged.Therefore,the number of people counting is one of the core indicators for estimating the degree of crowding.Most of the existing methods of crowd counting on the basis of the density map only focuses on feature extraction when training the counting model,and uses mean square error to calculate the error between the groundtruth and the density map of estimated,without further studying the influence of foreground and background on the task crowd detection.In order to solve the above problem,this thesis proposes a crowd counting algorithm based on multi-task enhancement combined with deep learning.By jointly optimizing the counting task and the foreground and background segmentation task,the ability to distinguish the foreground and background of the model was enhanced,and after that a higher-quality crowd density map was predicted.The results of comparative experimental illustrating that the algorithm can reduce the counting error,improve the robustness,and advance the quality of the estimated density map.