Abnormal Event Detection In Videos Using Two-Stream Convolutional Network

The growing video surveillance terminals raise a vast amount of surveillance video,and video abnormal event detection is the most critical application field.Abnormal events refer to an unusual behavior or pattern in a video scene,such as a car on a sidewalk,a pedestrian crossing a fence,etc.At present,deep learning methods have made significant progress in the field of video content understanding.However,due to the inherent inexplicability and weak robustness of deep learning methods,and the uncertainty of the definition of abnormal events,and the limitation of data sets,it is still extremely challenging to design a video anomaly detection framework that can effectively support decision-making.To this end,this thesis proposes a video anomaly detection method based on a two-stream convolutional network to achieve more efficient and reliable video anomaly detection.The research contents of this thesis are as follows:1.Two-stream abnormal event detection networkThe key of video abnormal event detection is to fuse practical information within and between video frames.This thesis designs a two-stream neural network framework,which uses a self-encoder to encode and decode spatial and temporal information in two-stream.The features are learned in the encoding stage,and the initial video frame sequence is restored in the decoding stage to generate reconstructed video frames.Specifically,the spatial flow sub-model divides the video frame from the spatial domain to learn appearance features;the temporal flow sub-model divides the motion information between video frames and learns the regular pattern of the video.On this basis,postfusion is used to improve the efficiency and accuracy of the detection results.2.Multi-granularity video anomaly detectionBased on two-stream convolutional network construction,multi-granularity video anomaly detection is performed,including video frame-level anomaly detection and object-level anomaly detection.Video frame-level anomaly detection uses reconstruction errors to determine abnormal events.Calculating the differences between the initial frame and the reconstructed frame determines whether an abnormal event occurs in the current video frame.If the current video frame is abnormal,the error between the reconstructed frame and the initial frame will be relatively large.Otherwise,no abnormality occurs.Object-level anomaly detection first uses neural network post-interpretation technology to visualize fine-grained features.Next,it displays the pixel-level visual elements learned by the convolutional layer through a heat map.And then identifies the individual objects that cause scene anomalies to achieve object-level video anomaly detection.Through qualitative and quantitative experimental comparisons,the effectiveness of the two-stream convolutional network proposed in this thesis is verified.It achieves more efficient video frame-level and object-level abnormal event detection,making the decision-making basis for abnormal event detection more sufficient.