Research On Abnormal Behavior Detection Method In Video Surveillance

With the gradual improvement of people’s safety awareness and the widespread application of monitoring equipment,it has become common to observe people’s abnormal behavior through video surveillance.The fatigue caused by video surveillance personnel observing surveillance video for a long time will reduce their labor efficiency.In the current context,abnormal behavior detection in video surveillance has long received attention from academia and industry,and has become an important research direction in the field of computer vision.The detection of abnormal behavior in video surveillance is extremely challenging,because abnormal events rarely occur in the real world,and the types of abnormalities are ever-changing.It is almost impossible for us to collect all abnormal events and use classification methods to solve the problem.Therefore,abnormal behavior detection in video surveillance is generally considered as an unsupervised learning problem.With the development of deep learning,breakthroughs have been made in unsupervised abnormal behavior detection.However,this method has a common problem,that is,due to the strong generalization of deep neural networks,it is very likely that the trained model will also describe abnormal samples well,which will lead to a large detection error.In view of the problem of "over-generalization",this paper proposes two solutions to improve the accuracy and robustness of the abnormal behavior detection algorithm in video surveillance.After completing these two solutions,a system platform is designed to realize the Simple application of anomalous behavior detection algorithms.The main research contents are:(1)From the perspective of model description,this paper proposes a video anomalous behavior detection method based on Multi-Task Learning constraints,which uses an improved model based on U-Net.Previous unsupervised learning-based anomalous behavior detection methods only use one task to model normal patterns.On this basis,we propose to use multiple tasks to model normal patterns from different perspectives,thereby increasing the binding force of the model.To a certain extent,the "over-generalization" problem of the model is solved.The experimental results fully demonstrate the effectiveness of the anomalous behavior detection method based on multi-task learning constraints.(2)From the perspective of encoding information of the constraint model,a video anomalous behavior detection method based on spatiotemporal attention constraints is proposed.The attention mechanism can make the content learned by the model only focus on the part that is effective for the optimization goal.Using the attention mechanism in the abnormal behavior detection method in unsupervised video surveillance can make the model pay more attention to the content describing the normal pattern,so that the boundary of the space where the description of the normal pattern learned by the model is located is more compact,which in turn solves the "overgeneralization" problem of the model to a certain extent.The model using the spatiotemporal attention mechanism has achieved better results in abnormal behavior detection.(3)An abnormal behavior detection system in video surveillance is implemented,and the model is improved according to the problems in the actual abnormal behavior detection.The system adopts B/S structure and is implemented using the lightweight Flask framework.The user can submit the video to be detected in the system,and the system will return the video segment of abnormal behavior in the user’s video.We tested the system using the test set from the UCSD Ped2 dataset,and the test results show that the system’s functionality has met the needs of detecting anomalies.