Research And Implementation Of Human Behavior Analysis Technology Based On Deep-learning

With the continuous development of information technology and the improved level of application of urban informatizaion,the surveillance event has gradually become out of control.At this stage,it is necessary to consume a,lot of manpower to distinguish the human behavior in video.The use of automated behavior analysis system can effectively alleviate this problem.This paper completes the framework of human behavior recognition based on deep-learning.The framework mainly includes the steps of object detection,object tracking,ROI extraction and behavior recognition.Based on the exixting object detection method,this paper proposes the pedestrian behavior sequence extraction method.In order to better extract the spatiotemporal information of human behavior in video,a new Two-stream architecture is proposed for spatiotemporal fusion during the process of extracting spatiotempoal features.The concrete work is as follows:1)In this paper,based on the frame-based object detection method(You only Look Once version two,referred to as YOLOv2),the video level object detection and tracking method is proposed to establish the relationship for the same target between the adjacent frames.And finally realize the detection and tracking function in video.2)Two-Stream method could not fully extract the efficent spatiotemporal information because of the separation of extractiong temporal and spatial information.To solve such problem,The 3D convolution and 3D pooling operation are used to extract the spatiotemporal information for a more convincing video-level descriptor.3)In this paper,the above methods are conbined to realize the framework of human behavior analysis based on depth learning,which can be used to analyze the behavior of multi-objective subject in the surveillance video.The experimental results show that the pedestrian behavior sequence extraction method can real-time detect and track human in video.The C3D-based Two-stream behavior recognition method can better extract and integrate temporal and spatial in-formation and enhance the video The average accuracy rate of human behavior recog-nition.By combining the two methods,the external disturbance is eliminated and the accuracy of behavior recognition is improved.