Research On Anomaly Detection Method Based On Unsupervised And Weakly Supervised Learning

Traditional video surveillance system has low timeliness and high invested cost,so it can't fully play the role of surveillance system.The intelligent monitoring system combines anomaly detection algorithms with computer equipment to achieve automatic monitoring and make up for the shortcomings of traditional monitoring systems.Abnormal behavior detection technology studied in this paper is the core technology of intelligent monitoring system.There are three main difficulties in anomaly detection:(1)the video has the characteristics of high dimension,noise and large amount of information.(2)The definition of abnormal behavior and normal behavior is vague.(3)Abnormal behavior is low probability and rarity.The characteristics of video data and abnormal events make it very difficult to acquire abnormal behavior data,and the task of information annotation is huge.This paper mainly studies the anomaly detection method based on unsupervised learning and weakly supervised learning.In this paper,three methods are proposed:(1)Unsupervised anomaly detection based on autoencoder and SEblock.The video data not only has spatial dimension and temporal dimension,but also has a lot of redundant information.Because too much redundant information will lead to poor anomaly detection results,this paper proposes an unsupervised anomaly detection method based on autoencoder and SEblock.Firstly,the method uses the stacked small-sized spatial convolution to learn the hierarchical feature representation of the video data and the ConvLSTM module to fuse the spatial information.Secondly,the redundant information is suppressed and main information is enhanced by SEblock module.Finally,the data distribution of normal behavior is learned by autoencoder.The experimental results show that SEblock module can restrain redundant information to some extent.(2)Unsupervised anomaly detection based on autoencoder and optical flow features.The redundant information in video data is mainly static background information.This method incorporates optical flow features into the input data to guide the model to focus on the moving foreground objects and ignore the static background information.In addition,several micro networks are added to the model to help the model learn more complex non-linear distribution functions.The experimental results show that the model can be effectively guided to focus on moving objects by incorporating optical flow features into the input data.(3)Weakly supervised anomaly detection method based on multi-instance learning and two-stream networks.Because the definition of abnormal behavior and normal behavior is vague,the model may not be able to accurately judge whether the behavior is normal or abnormal without using supervisory information.In order to solve these problems,the multi-instance learning method adds coarse-grained monitoring information to the training data that guide the model to learn the differences between normal and abnormal behaviors.In order to further improve the detection results of the abnormal behavior detection system,this method combines the spatial-temporal features of video data with the motion information of optical flow features to get a better representation of video data.The experimental results show that the two-stream network can effectively improve the detection effect of anomaly detection by fusing spatial-temporal features and optical flow features.The three methods proposed in this paper are tested on public datasets such as UCSD ped1,ped2,Avenue,Subway Exit,UCF,etc.,and compare other similar algorithms.The experimental results show that the three methods can achieve good results.