The Research Of Tracking-learning-detection Based On Non-rigid Target

Target tracking is a cutting-edge researching area in computer vision. Traditional Tracking-Learning-Detection algorithm, though efficiently solves the problem of target reappearance and occlusion, is limited only to deal with rigid targets. Considering the above problems, we proposed a brand new Tracking-Learning-Detection algorithm for non-rigid target. By modeling the non-rigid target with a probabilistic model, an efficient tracking algorithm for non-rigid targets with unknown geometrical shape has been realized in this thesis.Firstly, as a whole, we proposed a new Tracking-Learning-Detection framework based on a probabilistic model. Our framework integrates a tracking module with a cross-skeleton model and a detecting module with a probabilistic-template-matching method. Meanwhile, by performing an online learning mechanism for updating the model in the tracking module, we finally realized an efficient long-term tracking framework of Tracking-Learning-Detection for non-rigid target tracking.Secondly, for the issues of drifting and losing of tracking points which frequently happens in the tracking module, the selection strategy of optical-flow feature points has been raised. This method adopted a singular value decomposition algorithm to reduce the dimension of optical-flow points which finally solved the selection and renewal problems of foreground point sets with a proposed cross-skeleton model. Then, we also adopted an on-line learning strategy for training the proposed cross-skeleton model which finally solves the problem of constantly changing of target’s geometrical shape.Finally, for the template matching problem occurred in non-rigid scene, we also proposed a novel template matching method called probabilistic template matching which effectively solved the problem of foreground matching for non-rigid target in the detecting module. The methods working by combining the proposed cross-skeleton model in the tracking module with an effective image segmentation algorithm. Based on this, our thesis proposed an efficient probabilistic model for the expression of foreground target with which we finally realized a probabilistic template-matching algorithm for non-rigid targets. In the general public non-rigid data sets, our algorithm obtained very good results, the average rate of succeed tracked frames was 65%, compared with the traditional TLD, our algorithm got an increased rate of 124%, and to the other algorithm that had compared in the experiment, our algorithm also received a better effect.