Video Abnormal Event Detection Based On Perturbation Visual Interpretation

With the rapid growth of video data,it is not only necessary to identify the target and its behavior,but also to detect suspicious behavior in a large amount of ordinary data.The task of video surveillance is to detect abnormal events such as crimes,illegal activities,and traffic accidents.In recent years,since deep neural networks can reduce the labor intensity of feature extraction,they have become more and more popular in video abnormal event detection.However,due to the inherent “black box” properties of deep neural network models,it is impossible to understand the model’s decision-making process and the basis for the model to make decisions,which makes the detection results of abnormal events lack credibility.Existing video abnormal event detection methods cannot satisfy both high efficiency and interpretability.Aiming at the problems of current video abnormal event detection methods,this thesis proposes a video abnormal event detection method based on perturbed visual interpretation.The research content of this thesis is as follows:1.Video abnormal event detectionThis thesis uses the spatiotemporal autoencoder model to learn the spatial characteristics and timing relationships of video frames.The spatio-temporal autoencoder is mainly divided into two parts: the spatial autoencoder and the time encoder-decoder.The spatial autoencoder is used to extract the spatial structure of the video frame,and the temporal encoder-decoder is used to learn the temporal mode of the spatial structure extracted by the spatial autoencoder.The output of the model is the reconstructed video sequence of its input video sequence.By calculating the regularity score between the input video sequence and the reconstructed video sequence,and then setting a threshold for the regularity score,it can be judged whether an abnormal event has occurred.2.Perturbation visual interpretation of abnormal video eventsThis thesis proposes a perturbation visual interpretation method suitable for video abnormal event detection.This method is improved on the basis of the traditional perturbation visual interpretation method.On the basis of abnormal events,delete other objects and background areas that interfere with the abnormal objects,retain the objects that the model focuses on in the decision-making process,and locate the object areas where abnormal behaviors have occurred.This method improves the reliability of the detection results,and the experimental results show the effectiveness of the method.