Research On Video Human Abnormal Behavior Detection And Crowd Density Estimation Method

In recent years,with the increasing construction of "safe city" and "smart city" in China,the demand for video surveillance systems based on intelligent video analysis technology is also increasing.Processing and analysis of massive surveillance videos with graphics processing and computer vision technology,thereby achieving the rapid detection of abnormal events,the extraction of structured information from the crowd,and timely safety management and crowd control,etc.This is of great significance for maintaining public safety,improving the level of urban and rural management,and innovating the social governance system.Therefore,in-depth research has been carried out on methods such as human abnormal behavior detection and crowd number and density estimation in public place video surveillance.The main research work completed is as follows:1.We propose an abnormal behavior detection method based on YOLOv3 improved network model.Based on the classic YOLOv3 network model,it is improved from three aspects.First,build a multi-scale feature extraction network by adding residual modules,so as to achieve effective detection of large targets;then,add attention mechanism to the residual block and output feature detection layer,which can enhance the expressive ability of the network model and achieve feature importance weighting;finally,according to the specific detection scenarios and targets,we use kmeans++mean clustering to regenerate suitable target frames to improve network parameters.And the experimental results show that the average accuracy rate of YOLOv3-MSSE reaches 82%,which is 3% higher than the classic YOLOv3.At the same time,because of the residual module and attention mechanism,YOLOv3-MSSE greatly reduces the missed detection rate of the target.At last,we use this proposed YOLOv3-MSSE for abnormal behavior detection of targets in public places,and the experimental results show that the YOLOv3-MSSE can effectively detect the specific abnormal behavior of the human body in the monitoring scene,and can better deal with the unbalanced relationship between the detection accuracy rate and recall rate.2.We propose a density estimation algorithm that combines the attention mechanism and dilated convolution.First,combine part of the network layer of VGG-16 with the attention mechanism as the basic framework for feature extraction;then,replace the original network partial pooling layer and fully connected layer with zigzag dilation convolution module,so as to effectively compensate for the information loss caused by the pooling layer;finally,by fusing the feature information of the high and low layers to improve the ability of the network model for extract features,thereby improving the counting performance of the model.The experimental results show that our method performs better than most of the mainstream algorithms in accuracy.Compared with CSRNet,the MAE obtained by our method on the dataset UCF?CC?50is reduced by 7.2%,and MSE is reduced by 9.0%;the MAE obtained on the dataset Shanghai Tech are reduced by 9.3% and 24.0%,and the MSE are reduced by 7.2% and21.3%.This shows that our method can be well adapted to the detection of people with different densities,and has strong adaptability and robustness.