Research On Abnormal Behavior Proposal And Detection Based On Spatial-Temporal Attention

The purpose of abnormal behavior detection is to automatically detect and analyze unusual motion behavior from the scene of intelligent video surveillance,which is still of great significance for the intelligent video surveillance system to monitor the motion in the actual application scene,and further data analysis of the content information in the video scene.However,in the face of a large amount of video data generated by the huge intelligent monitoring system,how to effectively and accurately express the behavior or motion information in the intelligent monitoring video scene,and accurately analyze the behavior and motion in the intelligent monitoring video scene is still a major challenge.Many existing abnormal behavior detection models can detect the abnormal behavior in the video scene accurately,but there are still some limitations: the complex and changeable background information and the interference of other irrelevant actions make the video contain a large number of irrelevant redundant information;in addition,most abnormal behavior detection models rely on the underlying features of the action,which is difficult to be accurate to express the movement features of abnormal behavior.In this paper,based on the original motion and appearance features,three different forms of violence flow features are defined to analyze the abnormal behavior features in the video scene,and 3D convolutional neural network is used to deeply extract the spatial-temporal context information of the behavior,at the same time,combined with spatial-temporal attention mechanism,the interference of redundant frames is removed,and the significance of the features is enhanced,thus enhancing the abnormal behavior For the detection performance of the detection model in complex scenes.In view of the above problems,this article mainly focuses on the following work:(1)In view of the low efficiency,poor performance and high computational cost of most of the current motion proposal methods,this paper further proposes a motion proposal method based on spatial attention mechanism,which can suppress the interference of irrelevant background noise information and focus on the key areas of motion by integrating spatial attention mechanism into the 3D convolutional neural network,so as to enhance the performance of the proposed model The significance of the sign.In addition,the training of motion proposal model in this paper does not need to rely on the real bounding box of motion behavior,which greatly reduces the calculation cost of motion proposal model,and further improves the analysis efficiency and accuracy of motion proposal model in this paper.Experiments are carried out on UCF101-24 and JHMDB,which prove the validity of the proposed method based on spatial attention.(2)In order to solve the problems of low probability of abnormal behavior and complicated and changeable appearance in practical applications,this paper proposes three different forms of violent flows features based on motion and appearance features: obstacle force,contact force,and aggression force.To explore the deep motion features of behavior.Before calculating these violent flows features,the temporal attention mechanism is used to preprocess the input video frames to reduce the interference of irrelevant redundant frames and to select the key frames with important motion information.Then,the original video frame,optical flow and the calculated violent flows features are learned through a multi-channel 3D convolutional neural network,and the spatial attention mechanism is integrated to enhance the saliency of the features and remove the interference of redundant information such as background noise.Experimental results on the UCF-Crime dataset show that the method proposed in this paper can effectively detect abnormal behavior.