Research On Video Anomaly Detection Algorithm Based On Deep Learning

With the development of information technology and the explosion of big data,the amount of video is getting larger and larger.Video anomaly detection has also become one of the very important research directions in the field of computer vision.This article summarizes domestic and foreign research on surveillance video abnormal event detection,introduces the relevant knowledge of anomaly detection,uses deep learning methods for video anomaly detection,and makes a comparative analysis of the final experimental results to draw conclusions.The main contents of the thesis include the following:(1)Anomaly detection based on reconstruction error.In this paper,the Conv-LSTM unit is used to construct an auto-encoder model to reconstruct error.The model trains only video frames composed of normal scenes,with the goal of minimizing the reconstruction error between learning input data and output data.The anomaly detection system uses adaptive threshold settings and uses different data sets to detect the model.The experimental results show that the method has achieved excellent performance in these several data sets,and the accuracy of the algorithm is better than the ordinary method.Especially after adding the adaptive setting threshold algorithm,the accuracy rate of video anomaly detection is improved in four different data,and the recall rate is also improved.The adaptive threshold setting greatly improves the performance of anomaly detection and has good robustness.(2)Anomaly detection based on future frame prediction.U-Net is used as the basic GAN prediction network.In order to generate more realistic future frames,in addition to adversarial training and appearance constraints,loss constraints are also implemented in time and space to ensure that the optical flow of the predicted frame is consistent with the true value.This can both guarantee the generation of normal events of appearance and motion,and determine these abnormal events by comparing predictions with real values.The improved method compares and analyzes the experiment by adding different constraints.The experimental results show that each time a constraint is added to the future frame prediction method,the AUC value of the model becomes larger.The future frame prediction based on the appearance and motion constraints is larger than the existing one.The video anomaly detection method has better performance;on the other hand,the AUC based on the future frame prediction model in the Kindergarten dataset also reaches 72.8%.The future frame prediction model can effectively detect abnormal events in the video in real scenes.