Groups Segmentation And Abnormal Behavior Detection Base On Streakline

With increasing capability of computer processing and analysis of image, using computer tracking and analysis of groups of behavior has become a hot research areas of computer vision. Group segmentation and abnormal crowd behavior anomaly detection are the first step in the analysis of group movement and basic premise. This article introduced the concept of streakline in the fluid mechanics to computer vision, using the properties of fluid mechanics to extract the motion characteristics of the groups, to achieve liquidity groups’ segmentation and crowd abnormal behavior recognition.In computer vision, the first obstacle for calculating the motion characteristics of the groups that must be overcome is finding a good way to identify patterns of fluid without tracking individuals, because the tracking of individuals in a dense crowd is impractical and not necessary. Another obstacle to overcome when the scene content and dynamic population change in a long range, you need to find a new way to understand the changes in behavior. Optical flow is interframe instantaneous rate of change of the two-dimensional image pixel point; it can not capture the movement characteristic in the long range. Under Lagrangian framework for fluid dynamics, particle flow concept is introduced into computer vision. But spatial variations is omitted from the application of the particle stream, and with a significant time delay.In order to solve these problems, this article introduced the streakline. It calculated to obtain the velocity field on the basis of the optical flow field, and to obtain the collection of moving particles of a certain period, this paper combines the concept of extension particle to describe the time correlation of the particle movement, using the extension particles to construct streakline; its gradient is streak flow, according to obtaining streakline and streak flow to calculate the similarity of their temporal and special neighborhood,different regions is divided by the watershed algorithm.The potential function that calculated directly from the optical flow velocity field is sensitive to light and other external interference, while the streakline reflects crowd interframe macroscopic motion trend,potential function calculated by the streakline can be overcome noise interference, in the simple hydrodynamic model,potential function is decomposed Stream function and velocity potential according to Hlmholtz decomposition theorem, the crowd flow abnormal behavior recognition is the singular value extracted for each frame of the crowd flow velocity potential and stream function, and realizing description of movement characterization for each video frame; groups movement in the video is a continuous process, for video segment this article use PCA dimensionality to obtain descriptor of describing the motion characteristics of this segment of the video crowd flow; Finally, using support vector machine classified positive and negative sample library to discriminate whether crowd activity abnormal.DMN and UCF data set are used in this article, test results show that the recognition rate in DMN datasets is better than the current method of identifying test; recognition rate in DMN datasets is slightly lower than in the UCF datasets; this article prove that this method can overcome the interference of the background obscured、the light changes, the crowd dress changes and other external factors, and effectively identify abnormal behavior in group movement.