Abnormal Activity Detection Network Based On Optical Flow Estimation And Its Application In Surveillance Video

In today's society,video monitor has become one of the main means of maintaining public security,among which the surveillance personnel pay the most attention to abnormal events.In order to realize the automatic detection of abnormal action,namely to quickly locate the abnormal sequence interval in the video stream and identify its abnormal action for the reference and treatment of surveillance personnel,this paper studies the detection algorithm of abnormal temporal activity based on deep learning,and the algorithm process is divided into two steps: one is the location of abnormal sequence;the other is abnormal action recognition.For the lack of abnormal sample data caused by the nature of rarity about the unusual activity,abnormal temporal sequence location adopts a semi-supervised learning method to eliminate algorithm's dependency to abnormal data.Because abnormal activity is unpredictable,a network model which only learn the prediction methods about normal video data,can have a better performance in normal video frames prediction,however its predictions about abnormal frames are not good,therefore it can distinguish and locate the abnormal sequence in video streams.Prediction model uses retrospective prediction scheme,by adopting the constraints in terms of optical flow ? intensity ?gradient and adversarial training on forward and backward prediction to improve the prediction accuracy.Considering the problems such as over-fitting and learning on motion feature of video sequence caused by a small-scale of abnormal samples,an action recognition model based on optical flow estimation is used to identify abnormal sequences,and the optical flow map sequence about the input was calculated rapidly through the optical flow estimation network,which was used as the input of the subsequent 3D-Conv Net action recognition model.Finally,the anomaly sequence localization model reached the frame-level AUC of95.725% on the UCSD Ped2 dataset,and the abnormal action recognition model achieved a classification accuracy of 94.286% on the anomalous dataset.These experiments proved that the temporal localization of a semi-supervised learning model can accurately locate the abnormal sequence in the video,which can avoid the over-fitting problem on supervised learning under the situation of lack of abnormal samples.At the same time,the 3D-Conv Net,combined with the abundant motion and appearance information contained in the optical flow of the moving object,can greatly improve the accuracy of abnormal action recognition.The proposed framework can achieve better results on detection and recognition of abnormal activity.