Research On Abnormal Behavior Detection Methods For Video Crowds

With the widespread popularity of monitoring equipment and the rapid development of computer vision technology,video based anomaly detection technology has become an effective means in fields such as public safety management.Video based anomaly detection obtains real-time video streams through monitoring devices,and then uses computer vision technology to process and analyze the video streams,thereby achieving real-time monitoring and anomaly detection of crowd behavior.This technology can be widely applied in various scenarios,such as traffic control,intelligent security,accident warning,and other fields,playing an important role in ensuring public safety,improving management efficiency,and reducing accidents.The crowd abnormal behavior detection technology based on surveillance videos can identify abnormal behaviors that do not conform to normal behavior patterns by analyzing the behavior trajectory,posture,facial expressions,and other characteristics of the crowd,such as crowd gathering,running,pushing,crowding,and other behaviors.The existing research on detecting individual behavioral abnormalities is relatively complete,while the research on group behavioral abnormalities is still insufficient.Due to the large amount of surveillance video data,the cost of obtaining labels is high,and in most scenarios,the problem of imbalanced positive and negative samples is caused by the scarcity of crowd abnormal behavior events.Normal data is much larger than abnormal data,and the type of abnormal data is not clear,which can easily lead to the omission of abnormal events.In response to the above issues,this article first studied the application of methods based on population density and object detection in crowd abnormal behavior scenarios,and determined the specific types of anomalies that occurred through changes in population density and movement.Then,an unsupervised anomaly detection algorithm based on image reconstruction was studied.The reconstruction model was trained by learning from normal samples,resulting in significantly different reconstruction errors between normal and abnormal samples.This allowed for the identification of abnormal samples,avoiding the problem of imbalanced positive and negative samples and excessive dependence on abnormal data.The specific research content is as follows:1.Research on methods for identifying abnormal crowd aggregation and escape based on crowd density and target detection.Considering the abnormal crowd aggregation behavior,a fusion density feature map detection method is adopted.The DSNet network extracts the crowd density feature map and obtains the spatial distribution of crowd density in the density map.Based on the characteristics of the scene,density level intervals and abnormal thresholds are defined to determine whether abnormal crowd aggregation has occurred;For the abnormal escape behavior of the crowd,the object detection model YOLOv5 is used to track and detect the abnormal escape crowd in real time,obtain the position coordinates of the pedestrians,and estimate the movement speed and displacement vector of the crowd from the pixel displacement of multiple frames.Further achieve the detection of abnormal escape of crowds and event source localization through crowd velocity and displacement vectors.Comparative experiments were conducted on the public dataset pets2009 and the collected dataset to verify the effectiveness of the proposed method.2.For crowd anomaly recognition based on image reconstruction,on the basis of the variational self encoder,a prediction based space-time counter variational auto encoder(STAVAE)crowd anomaly aggregation and escape detection model is proposed.It uses the encoder of the variational self encoder to map the image to a low dimensional vector space,and then decodes and reconstructs a generated image,Finally,compare the differences between the generated image and the original image as a basis for determining anomalies.In order to improve the quality of image reconstruction and improve the accuracy of anomaly detection,combined with the convolution VAE model of the residual network,the long and short term memory network(LSTM)and the countermeasure network module are introduced to jointly represent and reconstruct the time and space dimensions of the video sequence,reducing the feature loss in the process of normal sample reconstruction and expanding the prediction loss of abnormal samples,Implemented the detection of abnormal behavior of crowd dispersal based on model reconstruction error.And the model was compared with Spatial Temporal Autoencoder(ST-AE)and Spatial Temporal Variational Autoencoder(ST-VAE)on public datasets UMN,Pets2009,and captured video datasets.The experiments showed that the STAVAE model has the best detection accuracy and recall rate in detecting abnormal escape behavior of crowds based on surveillance videos,The adversarial network module significantly improves the performance of anomaly detection.