Research On Clustered Crowd Monitoring Technology And Application Based On Deep Learning

The safety of people’s life and property is the first priority of national development,but the current global population is large,and social public safety problems such as trampling and crowding are still unsolved.In this thesis,we launch a targeted research based on deep learning on the analysis of aggregated crowd count and crowd group abnormal behavior,and the main research contents are as follows:(1)In terms of aggregated crowd counting,a multi-layer connected crowd counting algorithm based on an attention mechanism is proposed for solving the scale change problem.The algorithm consists of a front-end network and a back-end network.The front-end network acts as the backbone feature extraction network to generate high-resolution feature maps and retain finer head features.The back-end network is stepped and interconnected across columns to obtain high-resolution features at different network levels,incorporating an attention mechanism before each column to aid in enhancing the perception of the human head region.The experimental results show that the above approach can solve the scale variation problem well,and the generated high quality crowd density map can effectively improve the counting accuracy.(2)In crowd group anomalous behaviour detection,an improved 3D convolutional neural network(CNN)based on crowd density flow maps is proposed for understanding anomalous escape behaviour generated by crowds in surveillance videos.The crowd density flow map reflects the variation in the density distribution of the crowd,and the improved 3DCNN is used to learn the spatio-temporal features between video sequences.Due to the unbalanced sample of anomalous behaviour of the crowd,a loss function is designed for the improved 3DCNN in this paper.This paper creates a new dataset based on the UMN dataset with manual annotation.The crowd density flow map is used as the input of the improved 3DCNN,and migration learning is used to achieve the detection of group abnormal escape behaviour.The experimental results show that the above method has good recognition effect for understanding group abnormal escape behaviour.(3)Based on the research content of this paper,a prototype crowd monitoring system is designed.The prototype system combines a crowd counting model to achieve accurate crowd density estimates within a scene,and dynamic crowd density changes to achieve early warning of unusual crowd escape behaviour.The test results show that the prototype system can dynamically count the number of crowds and detect abnormal crowd behaviour in real time,which can assist managers to make timely and effective emergency response plans in response to emergencies.