Research On Crowd Target In Video Scene

Computer vision has become a very active research area, which involves signal acquisition, image processing, machine learning, pattern recognition, behavior control and even cognitive science and many other subjects. It researches on the target of video image sequence for detection, tracking, behavior analysis and so on. Target detection and tracking in video is an aspect of computer vision. Current research is only concentrated on single or multiple targets and human monitoring and tracking. The number of targets is limited to 10 or less for multi-target detection and tracking.With the increase of population and the acceleration of urbanization process, public places is becoming more and more crowded and large gatherings are also increasing. As a result, the requirements for crowd target monitoring have become urgent. But few researches have involved in crowd target detection and analysis. Concerning this problem, we have done the following work:First, an overview of motion field and optical flow field is given in this thesis. Motion field are vectors that describe the target motion. Without considering the effect of illumination, the optical flow field can be used to express motion field. Optical flow field is referred to the apparent motion of image brightness patterns. It can be obtained by adding constraints to solve the optical flow constraint equation. Lucas-Kanade optical flow method is suitable for crowd target detection and classification. We use the method to calculate optical flow field in this thesis.In Chaotic dynamics, the Lagrangian algorithm is an approach of dealing with fluid, which attempts to track the motion trajectory of each pixel. Due to the high density, crowd target can be recognized as fluid. Finite time Lyapunov exponent(FTLE) shows the mixing and separation of particles. It reflects the degree of separation between particles. According to flow diagram, FTLE images can be obtained by calculating the Runge-Kutta-Fehlberg equations. The thesis compares cubic interpolation and inverse distance weighted interpolation algorithm in the impact on the results. The simple inverse distance weighted interpolation algorithm is more suitable for real-time crowd motion target detection and analysis.Motion area is obtained based on FTLE image. This thesis presents an improved Bernsen adaptive binary algorithm to get FTLE binary image. Then the motion area is obtained by morphology. There are holes in the motion area. In order to solve this problem, we can use Freeman contour extraction.The motion area needs to be analyzed. The analysis is in two ways:the direction and density. For the direction, optical flow field is used. An improved K-means clustering approach is presented to obtain optical flow direction image. Duo to the quality of video, the results is mottled. The small-contour fusion algorithm is proposed to improve the results, which effectively removes all impurities. For density analysis, this thesis adopts the approach of texture. Gray level co-occurrence matrix (GLCM) analysis is an approach of texture analysis. Its characteristic parameters describe the image texture from different aspects. Contrast reflects the clarity of texture. The density of motion area can be got. According to the different density of all directions and the Bayesian classification, different targets are classified:sparse, middle and dense.All the proposed algorithms are implemented by C and Intel OpenCV library. We treat crowd target as fluid. Firstly, FTLE field is got. For solving Runge-Kutta-Fehlberg equations, inverse distance weighted interpolation algorithm is used. Non-chaotic points are excluded, and the rest represent the flow region. Secondly, motion region can be obtained after morphological processing. Then, K-means clustering is applied for motion direction analysis. Impurities are absorbed by small-contour fusion algorithm to optimize the results. Finally, GLCM is obtained for density analysis. Density information is given by the contrast of GLCM. The results show that the algorithms are effective. It brings a way to crowd target research.