Research On Early Warning Of Crowd Gathering Risk In Urban Public Places Based On Deep Learning

With the rapid development of the economy and the high concentration of the urban population,crowd gathering accidents frequently occur in public places,causing heavy casualties and property losses.Therefore,the risk of crowd gathering has attracted more and more attention all over the world.Reasonable urban environmental planning can reduce the possibility of crowd gathering accidents and ensure the safety of people’s lives and property.In order to monitor crowd dynamics in real time and reduce the risk of crowd gathering,urban public places have been equipped with video monitoring systems,but the monitoring mode is still manual,which is time-consuming and laborious and difficult to achieve satisfactory results.In recent years,the development of science and technology has entered the era of artificial intelligence,and national policies have accelerated the cross-application of artificial intelligence and other disciplines.Among them,the integration of deep learning and environmental science improves the intelligent level of urban environmental planning,and provides a new research idea for the early warning of the crowd gathering risk in urban public places.Based on deep learning theories and methods,this paper studies the key technologies for early warning of crowd gathering risks in urban public places.Specifically,we first quantify the three evaluation indicators including crowd density,crowd flow and crowd behavior,respectively.Then,we analyze and judge the degree of regional congestion,exit evacuation capacity and dynamic bottleneck status.Finally,we build a crowd gathering risk early warning system to monitor the key areas of densely populated places in real time,so as to provide intelligent technical support for real-time and accurate on-site handling of emergencies and crowd evacuation.The main research contents of this paper are as follows:(1)A crowd counting method based on multi-stage temporal inference network is proposed.Crowd density is an important evaluation indicator to describe the degree of regional crowding,and the intelligent analysis of crowd density can prevent regional crowding.In recent years,crowd counting methods have become the mainstream methods to estimate the crowd density.However,existing crowd counting methods mainly employ convolutional neural networks to regress density maps,which are usually difficult to accurately estimate pedestrian locations.This is mainly because the density map generated by the Gaussian kernel has a serious problem of overlap in dense areas,which directly affects the accuracy of the crowd counting method.To solve this problem,this paper proposes a crowd counting method based on multi-stage temporal inference network.First,the focal inverse distance transform is used to generate the density map with accurate pedestrian positions,which is used as the real label for network training.Then,a lightweight network is introduced to reduce the amount of computation and improve the overall operating efficiency of the network.Finally,a multi-stage temporal inference network is constructed to learn video-level semantic features and improve crowd counting accuracy.The multi-stage temporal inference network contains multiple sets of adaptive residual connection modules,in which each module is composed of a set of stacked temporal convolution layers.Experiments show that the proposed method has excellent performance,and it can quickly judge the degree of regional congestion,and meet the needs of real-time video crowd counting.(2)A crowd flow statistics method based on multi-head tracking is proposed.The crowd flow is an important evaluation indicator to measure the evacuation capacity of exits,and the intelligent analysis of crowd flow can improve the evacuation efficiency of exits.Due to the overlap and occlusion of people in dense crowd,existing pedestrian detection methods are difficult to accurately distinguish each individual in the crowd,which greatly increases the difficulty of crowd flow statistics.To solve this problem,this paper proposes a crowd flow statistics method based on multi-head tracking.Firstly,the local maximum detection strategy and k-nearest neighbor algorithm are used to transform the density map into the corresponding dense head bounding boxes.Then,a convolutional attention module is introduced to enhance the feature representation ability of the re-identification network by fusing cross-channel information and spatial information.Finally,Kalman filtering and Hungarian matching algorithms are used to predict and update the head trajectory,and construct the correlation matrix between the detection results and the trajectory to achieve efficient and accurate pedestrian head continuous tracking.Experiments verify the effectiveness and robustness of the proposed method,and the proposed method can be combined with the single-line method to accurately count the crowd flow in real-time.(3)A crowd behavior analysis method based on gaze estimation is proposed.Crowd behavior is an important evaluation indicator that affects the dynamic bottleneck state,and the intelligent analysis of crowd behavior can improve the crowd distribution capacity of the whole public place.However,because the crowd behavior is inconvenient to measure and evaluate,the existing bottleneck analysis methods usually ignore it.To solve this problem,this paper proposes a crowd behavior analysis method based on gaze estimation.This method constructs a three-pathway network to estimate the people’s gaze target in images/videos.The network can effectively integrate the relationship information between people/objects in the scene,scene saliency information and head information to predict the human gaze target in the image,and effectively capture the multi-scale features learned in the network training process through the micropyramid module.Experimental results demonstrate that the performance of the proposed method is superior to other comparison methods,and the statistical analysis of the gaze estimation results can accurately identify the crowd behavior type at the bottleneck.(4)An intelligent early warning system for crowd gathering risk is built.This paper takes a teaching building of R University in C City of J Province as the practical application scene to build an intelligent early warning system for crowd gathering risk.Based on deep learning technologies and crowd gathering risk theories,the designed system integrates the methods proposed in this paper into the original intelligent video monitoring system to comprehensively evaluate of multiple risk indicators,so as to achieve intelligent hierarchical early warning of crowd gathering risk.Our designed system can provide important technical support for monitoring staff to ensure crowd safety in emergencies.