Research On Abnormal Event Detection Method Based On Video Context Semantics

Video surveillance is playing an increasingly important role in public safety.Abnormal event detection is one of the main tasks of video surveillance.Through monitoring video,it can detect whether abnormal events have occurred in the current scene,and give early warning of abnormal events.Therefore,monitoring-oriented abnormal event detection has become an important research direction and has attracted extensive attention from researchers.In a video,abnormal events refer to abnormal events in the current scene,such as traffic jams,fights,crowd gatherings,etc.In the detection model of video abnormal events,the characteristics of normal events are generally learned based on the video before the current video frame.A large difference is generated,and it is determined whether an abnormal event has occurred according to the degree of difference from the normal scene event.However,most existing methods do not combine video context information when judging abnormal events,which leads to misjudgment and under-reporting of upcoming abnormal events.To solve this problem,this paper uses deep learning as the core tool and proposes a series of video anomaly detection methods,as follows:(1)For the problem of not combining video context information,this thesis believes that the earlier the abnormal event in the video is detected,the faster the abnormal event can be responded to,so an abnormal event detection method based on generating adversarial network connection video context is proposed.The video context information at the current moment judges whether an abnormal event has occurred,so as to achieve the effect of early detection of abnormal events in the video.This method uses the generative confrontation network as the basic framework.The current context in the video will be used by the generator to predict the video frame at the current moment,and the context at the current moment will be used by the generator to look back at the video frame at the current moment.Through this context learning The method can effectively detect abnormal events that are difficult to determine,and improves the performance of video anomaly detection.Experiments on multiple datasets demonstrate that the method is effective for improving the performance of video anomaly detection.(2)On the basis of the above work,a video anomaly event detection method based on a multi-branch generative adversarial network is proposed to solve the problem that video anomaly event samples are scarce and unbalanced with normal event samples.In most methods using generative adversarial networks,the discriminator only judges the generated results of the generator in the training phase,and only relies on the image quality of the generator to complete anomaly detection in the anomaly detection phase,so that the discriminator cannot play the role of detecting anomalies.At the same time,because the abnormal events in the video are rare,and the sample ratio of normal events will cause the problem of sample imbalance,which further reduces the accuracy of abnormal event detection.To solve the above problems,this thesis proposes a multi-branch generative adversarial network method based on context learning to detect abnormal events in videos.This method uses the pseudo anomaly module to randomly generate abnormal events from real normal events in the video.These generated abnormal events do not actually occur,so they are called pseudo anomalies.In addition,the performance of the generator training is at a low level in the early stage,when the quality of the video frames generated by the generator is poor,and these generated low-quality video frames can also be used as pseudo anomalies.The pseudo-anomalies are used to train the discriminator in the GAN to improve its performance in judging whether anomalous events have occurred in video frames.Moreover,the method uses the trained generator and discriminator at the same time in the anomaly detection stage,and the discriminator can be used to assist the generator to better judge whether an abnormal event has occurred in the current video frame,thereby improving the performance of the model.(3)Combining the above two work of video anomaly event detection,this thesis designs a video anomaly detection system based on video semantic information.The system is built by Django-Python framework,equipped with database,user interface,background data processing,software development of encapsulation algorithm Kit(SDK)and actual functions such as file transfer.The system integrates the above two video anomaly detection methods.After the video to be detected is uploaded to the server database,it can detect whether there are abnormal events in the video.First,the preprocessing operation of extracting video frames is performed on the video,so as to cooperate with the algorithm SDK to take video frames as input.The algorithm encapsulated in the system will calculate the normal score frame by frame.When the normal score of the video frame to be detected exceeds the threshold set by the system,it indicates that an abnormal event has occurred in the video frame,and the abnormality detection system will highlight the current video frame.After all the video frames of the current video are detected,the normal event video frame and the highlighted video frame will be merged into a result video as the output of the algorithm.Among them,the abnormal events that can be detected include sprinting,abnormal individuals,throwing objects,and people falling.The system has convenient features such as user-friendly operation,intuitive algorithm results,and historical record viewing,which can meet individual needs and facilitate function expansion.