Research On Abnormal Event Detection Method Under Video Surveillance

With the rapid development of our country's society and economy,the frequent occurrence of public safety accidents has become more and more serious,and emergencies such as trampling,theft,and fighting frequently occur.Therefore,it is very important to carry out comprehensive monitoring in public places.Because traditional video surveillance methods have been unable to meet public needs due to many problems such as high cost and low efficiency,intelligent security surveillance technology has emerged.As one of the important branches of intelligent video surveillance technology,abnormal event detection technology plays an important role in maintaining public safety.Therefore,video-based abnormal event detection technology has important research significance and application value.The main work of this thesis is as follows:(1)Aiming at the problems that the existing abnormal event representation methods do not fully consider the redundancy of video background information and the spatio-temporal correlation between frames,and it is difficult to apply to target occlusion,complex background and other motion scenes,a spatiotemporal feature based on target foreground detection is proposed.method.In this thesis,the background difference method and the frame difference method are combined to complete the foreground target extraction of the video image in the surveillance video,and solve the problem of background information redundancy.On this basis,the spatial-temporal gradient features of the video are extracted by calculating the threedimensional gradient of the pixels in the video spatial-temporal container,and then the principal component analysis algorithm(PCA)is used to perform feature dimensionality reduction,and the features after dimensionality reduction can be better represent the motion scene in the surveillance video.(2)Aiming at the problem of low detection accuracy and speed caused by the fact that sparse combination learning algorithm can't self-updating and the input of existing anomaly detection methods is usually video frame or optical flow image,a self-updating sparse combination learning detection algorithm based on spatiotemporal gradient feature is proposed.First,the foreground processing method is combined with the spatiotemporal gradient model to obtain the spatiotemporal characteristics of the video motion foreground block;then the selfupdating sparse combination learning algorithm is used to iteratively train the normal sample features to obtain the sparse combination dictionary;finally,the dictionary is used to reconstruct the test samples And make abnormal judgments,and select test samples with higher confidence to update the combined dictionary.The experimental results on the Avenue standard dataset and UCSD dataset show that compared with the existing detection algorithms,the algorithm in this thesis not only has a higher accuracy rate,but can also meet the needs of real-time detection.(3)In view of the current anomalous event detection algorithms in complex scenes overly relying on frame-level markers,as well as the long time-consuming and memory usage of the dual-stream extended three-dimensional(3D)convolutional network(Inflated 3D Conv Nets,I3D)model,the research designs a combination of I3 D and Motion Net networks.The dynamic dual-stream extended 3D Conv Nets(Motion Net-Inflated 3D Conv Nets,M-I3D)model is used as the feature extractor,and then an anomaly detection method based on dual-stream network and multi-instance learning is proposed.The method in this thesis uses normal and abnormal videos as packages,and video clips as examples in multi-instance learning.The M-I3 D network is used to extract the features of each video clip,and the extracted feature vectors are input to the three-layer fully connected layer.Learn a deep anomaly ranking model that can predict the score of anomalous video clips.In addition,in order to better locate anomalies during the training process,this thesis introduces a sparse function and a constrained function in the ranking loss function.Compared with several other methods on the UCF?Crime data set,the algorithm proposed in this thesis has higher accuracy and stronger real-time performance.