Video Anomaly Detection Model Based On Reinforcement Learning

In recent years,the development of multimedia and video applications has become more and more rapid,which has also more and more closely related to the daily life of human beings with the rapid increase of population,the security issue has attracted more and people's concern.Violent conflicts,terrorist attacks,explosions and other emergencies cause a lot of damage to human resources and material resources.How to deal with emergencies in time or even effectively prevent emergencies has become a key to solving security problems.Although there are surveillance cameras everywhere in our life,the existing anomaly event processing technology has not kept pace with the times,and failed to accurately detect abnormal events,the benchmark datasets used for research has also been serious shortcomings.Compared with normal activities,abnormal events are generally rare.Traditional video anomaly detection methods require a large amount of manual screening of large amounts of video information,and analysis and processing of useful information,which will lead to a great increase in workload.At the same time,it is also easy to detect wrong,such as error determination of event exceptions and error positioning of the location of the event.In the process of research abnormal event detection,there are also many problems in the existing abnormal datasets,such as the anomaly type is too simple,there are a large number of artificially forged videos and the inaccurate abnormal video sequence annotation,which makes the recent research in this field has been slow,so it is necessary to propose a novel abnormal video datasets.The research on video abnormal event detection integrates advanced technologies such as image processing and artificial intelligence,which can intelligently capture the occurrence of abnormal events,thus improving the efficiency of detection and saving manpower and material resources.This thesis constructs an anomaly video dataset for research,at the same time,an anomaly detection model is proposed to solve the existing problems with reinforcement learning technology.The main contents of this paper are as follows.(1)In this study,a new large-scale video benchmark dataset for video anomaly detection is constructed.The dataset can also be used in the field of video classification.The dataset consists of 2,000 video sequences,including 14 anomalous categories,such as Fighting,Trampled,Thiefing,and so on.Each video sequence is labeled on the video-level(abnormal/normal,video anomaly types)and frame-level(abnormal/normal video frames).The dataset is also divided into training set and testing set,which can be more convenient for research.The dataset proposed in this paper with the largest number of video sequences and the largest abnormal types in the world.(2)This study proposes an efficient anomaly detection model based on reinforcement learning.In this model,firstly,the trained residual network is used to extract high-level visual features of video frames.Then,a group of consecutive video frames is regarded as a combination,the features of the video frames in the combination are concatenated together,and put them into network to obtain the spatiotemporal features.Finally,based on learned features,a reinforcement deep network architecture with end-to-end training is constructed.Because anomaly detection is a problem related to time series,we transform this problem into a Markov Decision Process(MDP)and use a novel reward function to optimize the whole model.This is the first model that combines reinforcement learning and anomaly detection,which can better analyze and capture abnormal events.Experimental results show that the model is superior to existing advanced anomaly detection models.