Multi - Feature Video Tracking Algorithm Based On Particle Filter

With the development of computer science and related disciplines, computer vision has made vigorous improvement, and moving target tracking has become a central issue in the field of computer vision. With the emergence of other new technologies, such as extended Kalman filtering and particle filtering, which push the video tracking technology to be better. Kalman filtering theory is a successful solution of target tracking under the condition of the linear or Gaussian. With the increasingly complex application environment, the theory of particle filter was proposed to solve more non-Gaussian nonlinear problems, and it turns out to be an excellent one.The main research work of this thesis is as follow:1. Use edge, color and motion features to describe the goal which cannot be described by the one feature well. Besides, the thesis proposes the feature-oriented multi-feature fusion method. The color feature is the color attribution of the object, the thesis uses second order histogram to solve the insufficient information problems. Edge feature is the high-frequency properties of the image. The thesis uses Canny to get the edge features and do expansion operation. The motion feature refers to motion attributes of the objects, when the background of the video changed slowly, the target tracking based on the motion features becomes more prominent. The thesis compares the multi-features and single feature tracking results of different environment and proves that the multi-feature method is effective.2. Concerning the particle degeneracy phenomenon and samples shortage phenomenon this paper applies mean shift algorithm to particle filter. Mean shift is taken to cluster the particles before the resample, after which partial resample is taken. The algorithm is a trade-off between the particles degeneracy and the phenomenon of samples shortage. This algorithm not only relieves the particle degradation phenomenon, but also keeps the diversity of the particles. Also due to the convergence effect of the mean shift, by using the algorithm proposed by this thesis fewer particles can get the same effect as the classic one, which improve the real-time performance of the algorithm.3. By analyzing the improved algorithm and multi-feature fusion, this paper puts forward an improved particle filter algorithm based on the multi-feature fusion. The experiments of tracking some targets in occlusion case prove that the algorithm can get a fast and robustness tracking.