Crowd Intelligent Analysis Based On Surveillance Video

In recent years,accidents caused by high-density crowds and abnormal events in public places have occurred frequently,and the safety problems brought about by this have attracted great attention from society and the government.Therefore,there is an urgent need to use intelligent video surveillance technology to improve the level of real-time monitoring.The thesis aims at crowds in public places in surveillance video,researches on optical flow estimation,obtains the movement characteristics of the crowd,and improves the accuracy of group detection and abnormal event detection in crowd intelligent analysis.The main work of the thesis is as follows:Firstly,an iterative residual refinement optical flow network based on unsupervised learning is proposed.Based on Irr-Pwc,an attention mechanism is introduced to optimize occlusion detection,and a corresponding unsupervised loss function is designed to avoid the ground truth of optical flow.The experiments in the Sintel dataset and Kitti dataset show that the algorithm is competitive among existing algorithms.This algorithm is used to estimate the optical flow of image pairs,and provides optical flow for group detection and abnormal event detection.Secondly,a group detection method assisted by spatio-temporal information is proposed.In order to enhance the extraction and utilization of features,density maps and optical flow maps are introduced to assist the detection task;multi-view semantic information is used to construct graphs for multi-view clustering;correlation calculation of graphs in multi-view clustering is improved.The experimental results show that the algorithm can accurately detect the groups in the scene,and achieve better results compared with comparison algorithm crowd on the CUHK dataset.Finally,a weakly supervised abnormal event detection method based on two-stream model is proposed.In order to reduce the dependence on anomalous event annotation in video,a spatio-temporal feature extractor that processes RGB image sequences and optical flow sequences based on a two-stream model is designed.The extracted feature maps are weighted with attention weights for event classification.Attention weight and time gradient activation mapping(Tgrad-CAM)are used to locate the event.The experimental results of THUMOS14 dataset and UCF-Crime dataset verify the effectiveness of this algorithm in event classification and temporal action localization.