Research And Analysis On Non-rigid Video Track Algorithm

Target Tracking is an important issue in computer vision technology. There are many important applications in many areas for the image sequences to track of moving targets, such as medical, digital video surveillance, missile guidance and intelligent transportation systems.Video tracking method can be divided into two categories:probabilistic tracking methods and deterministic tracking methods. Deterministic tracking methods have efficient tracking performance and stability, high accuracy. Mean Shift algorithm is a typical representative of these methods, the method start with framing image of a target for kernel function weighted histograms in the spatial as feature extraction. Due to employee the weighted kernel function the searching problem can be optimized to a gradient searching in a continuous space, and thus the positioning of the target object equivalent transformation in the spatial. The method uses probability density function of the Bhattacharyya coefficient between the evaluation function to represent the target object model and the matching degree between the search object models. Finally, manipulate Mean Shift Algorithm to solve the optimal solution in spatial searching, find the target object model most similar to the search object model, thus completing the target location. Kalman filter and particle filter is a typical representative of the probability tracking methods. Kalman filter have strictly limitation on system modeling and the posterior distribution, so it can only deal with linear, Gaussian, single mode problems, whereas in the image tracking applications, the posterior probability distribution is often non-linear, non-Gaussian, multi-modal, and therefore the application of Kalman filtering subject to certain restrictions. In contrast, there are no special requirements for the system model in Particle filter, and it is able to maintain the state's multi-modal distribution, less susceptible to the impact of clutter, the tracking algorithm has been greatly development in this field. However, the existence of conventional particle filter tracking algorithm calculations and sampling are in low efficiency. The complexity of the actual track scene presents a great challenge to the conventional tracking algorithms.In this thesis, a number of images tracking algorithm which play an important role in tracking was studied and proposed, such as the object representation model, tracking the object template update methods and the adaptive update tracking window width. A joint color and shape characterization methods, kernel density estimation based on template matching adaptive update method and the optical flow based method for window adaptive updating method are proposed in this paper. The above improved algorithms are simulated for analysis and research based on Mean Shift algorithm and particle filter algorithm. Finally, the algorithm parameters influence to tracking accuracy is discussed. Simulation results show that the novel algorithms can track the target more accurate and more robustness than the traditional tracking algorithms.