英文摘要 | Object tracking plays an important role in visual analysis and understanding of human motion, it is the intermediate-level vision part. Object tracking is to detect, locate and track moving objects in video sequences captured by cameras. Through object tracking, we obtain the motion parameters and the trajectories of the object, which are the foundation of the high-level activity understanding and recognition. Recent years have witness great advance in tracking literature, it still remains a great challenging task due to clutter background, dynamical motion, appearance variations, and especially the occlusion problems. Pose estimation, a high-level vision task, is the task of determining the location, orientation and scale of each object. It is important for many vision understanding applications, e.g. visual interactive gaming, immersive virtual reality, visual surveillance, content-based image retrieval, etc. This thesis focuses on robust object tracking and pose estimation under complex scenes. Specifically, we mainly discuss the following three sub-topics: (1) graph embedding learning based appearance modeling; (2) sample impoverishment in tracking process; (3) high dimensional state space parsing in pose estimation. The main contributions of our work are summarized as follows: 1.We propose a graph embedding framework to simultaneously learn the subspace of the target and its local discriminative structure against the background. Then a robust appearance model is constructed based on the learnt subspaces. Experimental results demonstrate that, compared with two state-of-art methods, the proposed tracking algorithm is more effective, especially in dynamically changing and clutter scenes. 2.We propose two tracking algorithms, which combines of deterministic and stochastic tracking frameworks: (1) kernel Bayesian tracking framework; (2) singular value decomposition (SVD) based Kalman particle filter framework. (1) First, the kernel method is applied to the current frame to obtain the offset of the object state, and then these information are incorporated in Bayesian filtering framework as a heuristic priori. This algorithm can effectively combine the advantages of deterministic and stochastic tracking frameworks. And this algorithm is general to any form of appearance models. Experimental results demonstrate that, compared with Bayesian and kernel based tracking frameworks, the proposed algorithm is more efficient and effective. (2) Firstly, a s... |
修改评论