Research And Application Of Unsupervised Abnormal Event Detection Algorithm For Surveillance Video

The purpose of abnormal event detection for surveillance video is to identify all kinds of abnormal events in real surveillance video and deal with them accordingly in time,so as to better maintain social public security.In recent years,abnormal event detection algorithms based on deep learning for surveillance video have made good progress in improving the detection performance of model,but there are still some problems,the influence of spatio-temporal redundancy information in surveillance videos on the detection accuracy of abnormal events,and unclear demarcation between normal events and abnormal events.Therefore,abnormal event detection for surveillance video is still a challenging task.In view of this,an unsupervised abnormal event detection algorithm based on memory guided attention is proposed to solve the problem that the spatio-temporal redundancy information affects the detection accuracy of abnormal events.And an anomaly event detection algorithm based on unsupervised domain adaptive is proposed to solve the problem of unclear demarcation between normal events and abnormal events.Finally,an abnormal event detection system in elevator cage based on computer vision is designed to realize the engineering application of abnormal event detection algorithm for surveillance video.The main research results of this thesis are as follows:(1)An unsupervised abnormal event detection algorithm based on memory guided attention is proposed.There are a lot of spatio-temporal redundant information in surveillance video.When detecting abnormal events in surveillance video,the spatio-temporal redundancy information will affect the detection accuracy of abnormal events.To solve this problem,the algorithm introduces a memory-guided attention module to make the network model pay more attention to the valuable information in video frames when learning.To be specific,the frame difference method is firstly used to obtain the valuable information of normal events,and its feature information is stored in the memory unit of the memory-guided attention module.The input video sequence is then matched with the feature information stored in the memory unit.Finally,the matching information is fused with the feature information of the input video sequence for video frame prediction.In this way,abnormal events can be identified according to the size of the prediction error.In addition,a redundant information suppression constraint is proposed to reduce the influence of spatio-temporal redundant information on detection results.Experiments show that the algorithm model with memory-guided attention module and redundant information suppression constraint performs well on multiple data sets.(2)An anomaly event detection algorithm based on unsupervised domain adaptive is proposed.Existing unsupervised abnormal event detection algorithms only use normal event data for training and lack prior information of abnormal events,which leads to unclear discriminant boundary between normal events and abnormal events and easy to produce false detection.To solve this problem,the algorithm adopts unsupervised domain adaptive method to introduce the priori knowledge that can well define normal events and abnormal events in the source domain into the target domain.Specifically,supervised pre-training is performed on the source domain dataset to clarify the discriminant boundary between normal events and abnormal events in the source domain.Then,an unsupervised domain adaptive method based on adversarial learning is used in the target domain to introduce the priori knowledge of the source domain into the target domain,and clarify the discriminant boundary between normal events and abnormal events in the target domain,so as to improve the detection performance of abnormal events in the target domain.In addition,adversarial learning can align the data distribution of source domain and target domain,reduce the domain offset,and improve the scenario applicability of the algorithm model.Experiments show that this method can effectively improve the performance of abnormal event detection and make the algorithm model have better applicability to scenarios.(3)Engineering application of abnormal event detection algorithm in surveillance video.In the field of video surveillance,the method of detecting abnormal events by human vision requires huge manpower and is easily affected by visual fatigue.To solve this problem,this paper designs an abnormal event detection system in elevator cage based on computer vision.For the detection of dangerous goods and fire in elevator cage,the corresponding datasets are constructed respectively,and according to their different characteristics,the corresponding abnormal event detection algorithm is used to detect them.By deploying the above algorithm in the abnormal event detection system in the elevator cage,and then using the developed software system to detect the abnormal events in the elevator cage in real time,and display the real-time results of the abnormal event detection on the system visualization interface,so as to realize The engineering application of abnormal event detection algorithm in surveillance video is presented.