Research On Video Abnormal Behavior Detection Based On Autoencoder Model

Video abnormal behavior detection is an important branch in the field of computer vision with wide applicability in many fields such as intelligent transportation and intelligent security.Over the years,researchers have proposed many methods in the field of video abnormal behavior detection,among which the most representative methods are automatic encoder and generative adversarial network.However,there are still many problems in the field of video abnormal detection that need to be solved,such as the interference caused by complex backgrounds on object feature learning and the relationship between the definition of abnormal behavior and the environment in which the behavior is located.This paper mainly focuses on how to effectively remove the interference information from complex backgrounds and how to use environmental information to assist the discrimination of abnormal behavior in video abnormal behavior detection.The specific implementation scheme is as follows:(1)To solve the problem that complex backgrounds interfere with object feature learning,this paper proposes a video abnormal behavior detection algorithm based on background separation network.First of all,in order to reduce video jitter,this paper proposes a video stabilization algorithm based on SURF feature extraction to eliminate video jitter.In order to solve the problem that the traditional background separation algorithms are prone to incomplete foreground object extraction,this paper constructs a background separation network based on ViBe background extraction algorithm and Mask R-CNN instance segmentation algorithm.The network obtains the motion pixel distribution map of the current video frame by ViBe algorithm.Then,the image is segmented by Mask R-CNN,and the required object mask map is obtained by setting the detection tags.After that,the motion pixel distribution map is used to correct the foreground target mask map to keep the correct foreground target and finally obtain the foreground image.In the model training stage,in order to obtain better training results and improve the robustness of the model,this paper proposes a random masking module,which mask a certain percentage of the foreground targets in the input foreground image as the background.The processed images are then fed into the autoencoder to train the model,and finally the video abnormal behavior detection results are obtained by anomaly discrimination.In this paper,we validated on UCSD ped2,CUHK Avenue and SHTech Campus,show that the accuracy of the video abnormal behavior detection algorithm based on background separation network is 96.2%,80.7% and 72.3%,respectively,which is better than the video abnormal detection algorithm using the original image for feature extraction.(2)In order to solve the problem that the definition of abnormal behavior is related to the environment in which the behavior is located,this paper proposes a dual-channel autoencoder network based on the background separation network.The dual-channel autoencoder network consists of two encoders and an attention decoder,one encoder is used to extract the original image features and the other encoder is used to extract the foreground image features extracted by the background separation network.Then the foreground features and the original features are later feature fused as a way to introduce the spatial contextual information of the object,and then the features are fed to the decoder.Because background information is incorporated into the features,in order to make the network pay more attention to the features of important objects in the image,this paper presents an attention decoder module based on the traditional decoders.When the feature is input to the attentional decoder module,the attention decoder will pay more attention to the reduction of the foreground objects in the process of feature reduction,reduce the propagation of noise in the process of feature reduction,and at the same time can make better use of the correlation information between channels.In the model testing stage,the final video abnormal behavior detection results are obtained by calculating the reconstruction errors of the reconstructed image and the input image.In this paper,we validated on UCSD ped2,CUHK Avenue and SHTech Campus,show that the accuracy of video abnormal behavior detection algorithm based on background separation network and dual-channel autoencoder network is 97.1%,82.6% and 73.8%,respectively,which is superior to the corresponding algorithm.