Research On Anomaly Detection Technology In Surveillance Video Based On Unsupervised Learning

With the rapid development of the economy,people's requirements for urban life safety are constantly increasing,it is becoming more and more common to install surveillance cameras in public places.At present,most surveillance video systems rely on manual monitoring.However,after a long time of work,people will be distracted and missed due to fatigue.In order to overcome various problems of manual monitoring,surveillance video anomaly detection technology emerges as the times require.It can intelligently analyze video data through computer programs,and can automatically warn of abnormal events occurring in video scenes,which greatly improving detection accuracy and detection efficiency.In surveillance video anomaly detection,it is mainly divided into three types of anomaly detection methods based on supervised learning,semi-supervised learning and unsupervised learning.The anomaly detection method based on supervised learning can only detect specific abnormal behaviors in the video,and has great limitations in use.The anomaly detection method based on semi-supervised learning is too dependent on the constructed model.When the model is biased,it is likely to lead to false detection and missed detection.The anomaly detection method based on unsupervised learning directly detects the test video,which has high flexibility,and has no model dependency problem,and has broad research prospects.This dissertation studies the surveillance video anomaly detection technology based on the unsupervised learning method.We divide the anomalous events into local and global anomaly events,and perform anomaly detection from two different perspectives.In the local anomaly detection,we firstly extract and extend the motion blocks in the video.The extended motion blocks is used as the basic unit of local anomaly detection.Then we perform anomaly detection in three dimensions: time,space,and time-space.In the global anomaly detection,we firstly extract the moving targets in the video.The moving targets is used as the basic unit of global anomaly detection.Then we use the sliding window on the moving targets sequence and perform anomaly detection in each window.We finally optimize the test results based on temporal and spatial consistency.Based on the above anomaly detection algorithm,this dissertation designs and implements a surveillance video anomaly detection system.The main work of this dissertation is as follows:Firstly,we provide a local anomaly detection algorithm in the surveillance video based on unsupervised learning.We firstly extract the motion blocks in the video and extend them.Then,we calculate the time anomaly,spatial anomaly and space-time anomaly of each motion block separately.Finally,we fuse the three anomalies and setthe local anomaly threshold for abnormal judgment.The algorithm does not need to use the training sample to build the model,which solves the model dependence problem in the surveillance video anomaly detection algorithm based on semi-supervised learning.And only the motion area where abnormality may occur is detected,which improves the algorithm detection efficiency.By diversified detection dimensions,the algorithm detection accuracy is improved.Secondly,we provide a global anomaly detection algorithm in the surveillance video based on unsupervised learning.We firstly cluster the motion blocks in the video to extract the moving targets.Then,we use a sliding window on the moving target sequence to make abnormal judgment in each window according to the set global abnormal threshold.Finally,we optimize the test results based on temporal and spatial consistency.The algorithm also does not need to use training samples to directly test the test video,which solves the model dependence problem.And the algorithm uses each moving target in the video as the basic detection unit,which ensures the unity of the detection unit and improves the detection accuracy.The algorithm adds the Unmasking operation to the detection,which makes the result of the anomaly detection more accurate.What's more,the detection results are optimized based on temporal and spatial consistency,which reduces the false detection caused by improper selection of thresholds,and further improves the detection accuracy.Finally,we design and implement a surveillance video anomaly detection system based on unsupervised learning.We firstly introduce the overall design of the system.Then,we explain the implementation of system related functions.Finally,we analyze and evaluate the effectiveness of the system.