Research On Crowd Group Behavior Analysis Method Under Dynamic Background

With the development of computer vision technology,intelligent crowd behavior analysis methods based on monitoring systems are increasingly used for security control of various crowd scenarios.People in crowd scenes usually form many small-scale groups due to behavioral interactions between individuals and restrictions of the surrounding environment.Within each group,the pedestrians share similar motion patterns and exhibit collective behaviors.As the primary components of a crowd,groups convey sufficient semantic information.Therefore,crowd behavior can be well identified by analyzing the behavior and changes of groups.This paper proposes a crowd group behavior analysis method under dynamic background according to the application requirements of the crowd scene monitoring system for related algorithms.This method can well identify different crowd behaviors by detecting and analyzing crowd groups.For the crowd scene with a fixed camera,the traditional crowd trajectories extraction method is susceptible to interference from dynamic background such as illumination changes,grass motions,etc.,and cannot well tackle tracking drifting problems caused by crowd occlusion.To solve these problems,a crowd trajectories extraction method based on contour tracking is proposed in this paper.Different from the traditional method based on gray level modeling,this method firstly extracts foreground edge by edge gradient background modeling,which improves the efficiency and accuracy of foreground detection.Then,this method extracts stable and complete foreground contours by extending the vertical edge.After extracting contours,this method tracks the feature points on the contours stably by optical flow forward-backward refinement method to filter out the contours of dynamic background.Next,this method can get the crowd foreground area according to the crowd contours.Lastly,this method can get robust crowd trajectories by tracking crowd foreground area feature points.The test results reveal that this method can capably overcome the interference of dynamic background,suppress the drift phenomenon in tracking,and is robust to crowd scenes.For the crowd scene with a shaking camera,to overcome the influence of global motion,this paper proposes a crowd trajectories extraction algorithm based on homography transformation of background feature points.Firstly,to improve the accuracy of global background motion estimation and foreground motion compensation,this method distinguishes foreground and background trajectories by analyzing the moving distance of feature points.Next,to reduce the misjudgment of short distance trajectories caused by crowd occlusion,the method further classifies background trajectories obtained by the first step according to the homography matrix remapping error.Finally,the camera jitter is estimated according to the background trajectories,and then the camera jitter is used to make foreground motion compensation.The test results reveal that this method can precisely distinguish foreground and background trajectories,eliminate background motion,and obtain stable and smooth crowd trajectories.After obtaining the stable and reliable crowd trajectories,to better analyze crowd behavior,this paper proposes a crowd group behavior analysis method based on the group center.Firstly,the similarity of the movement direction and the spatial position distribution of the feature points is first utilized to find the group centers,which reveal the motion dynamics of groups.Next,according to the topological relevance to the group centers,feature points are classified into local clusters.Finally,by analyzing the similarity of motion vectors of group centers,the distribution of motion vector intersection points,the distance between the groups and the change of group area,three different behaviors of the detected crowd groups are identified,including gathering,evacuation and running in the same direction.Then the abnormal behavior alarms can be given in time.Experimental results on the running monitoring system reveal that this method can achieve real-time crowd behavior detection with low false alarm rate and high recognition rate under complex dynamic background.