Video Anomaly Detection Using Spatiotemporal Information

The public safety has always been the focus of attention.However,abnormal events affecting public safety still occur occasionally.In order to find and handle abnormal events timely,video surveillance has appeared on various occasions in life.The number of cameras is also rising sharply,and the coverage is getting wider and wider.However,it is time-consuming and laborious to manually detect abnormal events in videos.How to automatically detect abnormal events in videos has become a research hotspot.As a kind of high-dimensional time series data,video contains rich spatiotemporal information.Extracting the spatiotemporal features to represent videos is the key to distinguish between normal events and abnormal events in videos.We use video cubes as the basic unit,and combine high-order VAE coding features with low-order HOF features describing motion information intuitively into a local spatio-temporal feature set to represent the spatio-temporal information in the video.Based on the K-means clustering algorithm,an algorithm for calculating the feature deviation score is proposed.The deviation score and the VAE reconstruction error are weighted to calculate the last anomaly score.In addition,a video anomaly detection algorithm based on anomaly tracking is proposed.The cubes that abnormal objects may appear next time are tracked with finer granularity around cubes with high anomaly scores,and tracking of abnormal objects is continuous.Due to the complexity of the video scene,the anomaly scores of the same abnormal object in different video positions may vary greatly.In order to reduce the rate of abnormal omissions,the tracking cube with a high probability of abnormality is weighted with the corresponding abnormal tracking starting cube to improve anomaly score of the tracking cube.Experiments are performed on three public data sets such as UCSDped2,Avenue and Shanghai Tech.Compared with the state of the art video anomaly detection algorithms,we obtrain the competitive experimental results,which verifies the effectiveness of the algorithm.