Research On Abnormal Event Detection In Videos Based On Deep Learning

Surveillance Video Abnormal Events Detection refers to the use of computer technology to automatically analyze abnormal events contained in a video sequence.When a suspicious(abnormal)condition occurs,the system can detect and issue an alarm in a timely manner.This technology can help the monitors to effectively deal with dangerous events and minimize the false detection rate of missed detection.Surveillance video abnormal events detection has a good application in safety precautions,disaster prediction and social security.In recent years,the police department has established a "skynet" system,so we can see there are surveillance cameras everywhere in cities.If it can effectively analyze and understand the contents of the surveillance video,timely handling of the anomalous events in the video will have very important significance in maintaining public order.However,the video data taken by the camera every day is huge,and the manual monitoring method takes time and effort,resulting in poor results.Based on the above-mentioned situation,this article uses computer vision technology to carry out research on the detection of abnormal events in surveillance videos.The main contents of this article are as follows:1.Aiming at the problem of video anomaly detection,this paper proposes an improved framework based on Deep Autoencoder.The framework includes two main parts,one for event space feature modeling and the other for event time feature modeling.Specifically,the convolutional spatio-temporal autoencoder model is used to process video frames of normal events in an unsupervised manner so as to obtain appearance features and motion features of normal events.These features are combined to form an event video representation.Then,the samples of the test dataset are input into the trained model.The reconstruction error of the normal events is small,and the reconstruction error of the abnormal events is large.After a suitable threshold is established,the network can judge whether the frame is classified as abnormal according to the reconstruction error of each frame.Finally,the assessment test is performed on the UCSD dataset.Experiments show that the proposed algorithm is better than the state-of-the-art method and the detection rate is increased by 1%?3%.2.Aiming at the blurred generated prediction frame problem,a method of detecting abnormal events by using Generative Adversarial Networks(GAN)is studied.The method discriminates abnormal events by comparing the difference between the predicted frame and the ground truth.Specifically,given a video,we predict the future framework based on its historical observations.We first train a predictor that can predict the future framework of normal training data well.During the testing phase,if a framework is consistent with its predictions,it may correspond to a normal event.Otherwise,it may correspond to an abnormal event.In order to predict the future frames of normal events more accurate,other than the commonly used appearance(spatial)constraints on gradient and motion(temporal)constraints such as optical flow,we also introduce a generative adversarial loss to make the predicted frame more realistic.To assess the effectiveness and superiority of this method,experiments were performed on two representative abnormal event detection datasets:Avenue dataset and Shanghai Tech dataset.Our methods are compared with other abnormal event detection algorithms.The evaluation results show that the algorithm in this paper achieves the desired result.