Target Tracking Based On Bayesian Filter In Colored Image Sequence

As a challenging research topic in computer vision, colored traget tracking of dynamic scenes attempts to detect, recognize and track moving objects from image sequences. This thesis focuses on the research on foreground measurement and filtering, which are two key techniques in visual Bayesian tracking.Based on color vision and colorimetry theory, pixel classification and target region segmentation are firstly discussed. In order to meet the practical requirement in real-time and effectiveness, the pixel classifications based on constant threshold method are compared and analysed in three kinds of color spaces respectively.The representation of Bayesian tracking is given based on Markov hypothesis and Bayesian formula, and three implementation methods of Bayesian tracking are introduced and disscussed in succession. In order to satisfy the requirement of RoboCup, a colored ball tracking method is proposed which combine the Kalman filter with search window updating method. Instead of global searching in image, local segmentation greatly decreases the computation burden and improves the accuracy of tracking results.In order to implement tracking task in non-Gaussian measurement environment, a novel approach of hybrid particle filter is presented to process the target's position and shape respectively, whose states updating is on the basis of data fusion between Kalman filter and particle filter. A method of achieving least uncertainty measurement is also used to complete the optimization of particles' measurement matching. The proposed method can not only pave the way for a low-complexity particle filter algorithm in dealing with higher dimensional tracking problem, but also cover the drawback of Gaussian restriction in Kahnan filter.Color distributions provide an efficient feature as they are robust to partial occlusion, rotation and scale variant. This thesis presents the integration of weighted color histogram into particle filter, which takes into account the target's shape as a necessary factor in target model. Bhattacharyya distance is employed to estimate the similarity between the target model and each hypotheses of the particle filter, which makes samples' weight updating more reasonable. A series of experiment results and data analysis demonstrate the method's validity and practicability.