This thesis focuses on robust contour and region tracking under mobile camera. Specifically, we mainly discuss the following four sub-topics: (1) contour tracking with abrupt motion; (2) improving the accuracy of contour tracking; (3) the construction of observation model for region tracking. The main components of our thesis are listed as follows: 1. We firstly address the problem of abrupt motion and accuracy in the contour tracking area. In order to track the contour of object with abrupt motion, we propose a particle swarm optimization based contour tracking framework. In this framework, the problem imposed by the abrupt motion is efficiently solved. Meanwhile, in order to improve the accuracy of contour tracking, we propose a contour tracking algorithm that based on the discriminative model. The Adaboosting algorithm based discriminative model is adopted in constructing the energy function for the contour evolution. Our algorithm improves the accuracy of contour tracking and its ability in handling noisy environment. 2. In the second part of our thesis, we mainly concentrate on constructing a robust appearance model that can provide more robust region tracking results. Firstly, we propose a probabilistic index histogram for the appearance of the target. Our model can efficiently overcome some limitations of the histogram representation. We also introduce spatial distance and bin-ratio dissimilarity for more robust histogram comparison. Secondly, we explore the problem of fusing different types of appearance model. We propose a Volterra embedding based discriminative model for the appearance of the target. Furthermore, we compensate the discriminative model with the generative object model through a unified graph embedding framework. So our model has the complemental advantage of both models and can provide more robust tracking results. At last, we explore the problem of multi-cue fusion in constructing the appearance model. We propose a multi-task joint sparse representation framework for robust object tracking. The problem of multiple features fusion in the sparse representation is efficiently handled through the joint optimization of multi-task. 3. In the last part of our thesis, we focus on design a robust tracking algorithm for the hand. This is because hand is one of the most important tools in the human-computer interaction, it is meaningful to design a particular algorithm for it. We propose a RCD criterion based multi-cue fusion ...
修改评论