Research On Object Tracking Using Sift Feature And Particle Filter

With the development of economy, computer vision technology and multimedia processing technology has been rapid development. One of key area of computer vision researching is moving target detection and tracking technology which is the foundation of higher-level of motion recognition, event monitoring and video analysis etc. At the same time, it is widely used in many fields including intelligent car, intelligent traffic monitoring system and video retrieval.This thesis focuses on the moving target detection and tracking, and lays strong emphasis on the update of the Learning Rate about Gaussian mixture model in the moving target detection and the reduce of the background disturbance in the moving target tracking. The main contributions of this thesis are as follows:1. Our paper do a comprehensive summary of the meaning of moving target detection and tracking technology, domestic and foreign research status, introduce the whole process of this technology, and classification of this technology and do a simple presentation.2. This paper proposed a new method on the update of the Learning Rate about Gaussian mixture model. The learning rate should be updated on-line according to the numbers of each model matched of each pixel. First, the initial background was established using the traditional Gaussian mixture model with a global learning rate. Second, the adaptive learning rate was used for each pixel according to the number of matches when the background was updated. Finally, the moving objects were detected by background subtraction. Experimental results show that comparing to the moving objects detection approach based on the conventional Gaussian mixture model, this approach has a desirable stability and learning ability.3. Following a brief analysis of Kalman filter tracking and Mean Shift tracking. A new method to the target tracking which combines Kalman filter tracking with Mean Shift tracking is introduced. By using Kalman filter to predict locations where the target probably appear and Mean Shift to search in the corresponding areas and match the target. Although the effect of the combined method is better than that using only one method, the size of the window cannot reflect the size of the target, so this paper proposed a method that combined the Kalman filter tracking with the improved Mean Shift tracking. Experimental results show effectiveness of the improved approach.4. Simply using color histogram for object tracking is vulnerable to the background disturbance. This article presents the feature points matching into particle filter which has typically been used in combination with color histogram to solve the background disturbance. This tracking algorithm is based on particle filter in which a new method of computing each sample weight is proposed. Each sample weight can be obtained through measuring the similarities of color histogram and feature points between the object model and each sample. Experimental results show that comparing to the moving objects tracking approach based on the conventional particle filter, this approach can alleviate the background disturbance.