Research On Video Abnormal Detection Method Based On Image Reconstruction And Prediction

In recent years,the use of surveillance cameras has been gradually prevalent and covered from major area with huge traffic to inconspicuous region.Both the increased amount of data and the development of computer science techniques makes manual inspection,which takes huge labour,no longer be the best option.Meanwhile,Artificial intelligence(AI)is rising more attention from researchers as it can significantly improve the efficiency and accuracy of video abnormal detection models,enabling them to filter out abnormal events from a multitude of videos after learning normal features,in which image reconstruction methods and prediction methods play the important role of video abnormal detection relying on its outstanding and stable performance.Firstly,to solve the problem that the detection information is relatively simple,a video abnormal detection method based on reconstruction(Recon＿VAD)which performs detection based on the non-temporal information in a single frame and a video abnormal detection method based on prediction(Pred＿VAD)which performs detection based on temporal information in multiple consecutive images are proposed.To solve the problem that the mask edge length cannot be selected automatically and the local feature extraction ability is limited in Recon＿VAD,a mask edge length calculation strategy is proposed and a dense U-Net++network structure is applied,aiming to make the model can reconstruct the input image according to the learned normal features and use the reconstruction error of each pixel position as the basis to detect abnormal events,which makes the feature extraction performance and the video abnormal detection accuracy is improved.Secondly,to solve the problem that the feature extraction ability is limited when the Pred＿VAD is oriented towards dense flow of people,coordinate attention mechanism is applied to make the model more sensitive to changes of the location information of targets,aiming to make the model can predict subsequent frames based on accurate normal characteristics and compute the pixel-wise error used as the criterion of predicted and observed frames,which makes the detection accuracy of Pred＿VAD model is improved.Finally,to solve the problem that the boundary between normal event and abnormal event is not clear enough in existing video abnormal detection models,a combined method called Recon＿VAD+Pred＿VAD is proposed which uses the output reconstructed image from Recon＿VAD as the input of Pred＿VAD.By combining non-sequential and sequential information,abnormal events can be more accurately separated and detected.In addition,the Recon＿VAD+Pred＿VAD model is verified to have better video abnormal detection performance by comparing with some classical models in the video abnormal detection field on UCSD Ped1 and Ped2 datasets...