| Human 3D motion estimation has been a hot research issue in computer vision field in recent years. There are three chief reasons: 1)the popularity of non-professional digital video equipment;2)widely application, such as virtual reality, animation design, computer game, Advanced Human-Computer Interaction, video surveillance, assistant sports training, robot action design and so on;3) the rapid development of computer vision theory.Nowadays, more and more researcher has been shifted reseach emphasis to the monocular markless video because multi-camera system and monocular video with marker can not deal with human natural motion. But this topic still is a challange problem because of low computional efficiency caused by the high-dimensional state space, and ambiguity caused by occlusion and self-occlusion in the human motion.This dissertation focuses on study the related technologies of human 3D motion estimation based on monocular markless video, with the emphasis on human 3D motion estimation based on particle filter. This work investigates on how to improve the computional efficiency of particle filter, and to slove the problem of ambiguity caused by self-occlusion in the human motion, and has obtained some good resalts. The human motion on the diferent stats adn tries to estimate the human 3D motion on the diefferent states using different strategies was studied on the basis.This dissertation includes the following aspects:1. This dissertation studies decomposition of state space, introduces the graphical model to the human 3D motion estimation and propoes the state space decomposition strategy based on the limbs graphical model. So the high dimensional arm motion state space could be discomposed into low dimension subspaces, and the top-down strategy. Based on those works, the thesis introduces the concept of joint chain, and proposes the articulated graphical model to estimate the arm 3D motion and the particle filter are used for tracking the arm motion. A new particles projection algorithm is proposed to strengthen the human motion constraints. The algorithm has shown the good computional performance, and improved the computational efficiency in some degree.2. There is very strong correlation among limbs in the course of human motion. Based on the strategy of learning motion correlation, the dissertation analyses the speed correlation between right big arm and right forearm, proposes limb motion speed correlation model. So right forearm motion can be linear represented by right big arm motion via limb motion speed correlation model. As a result, human self-occlusion can be sloved at some degree while improvement of computional effciency.3. The dissertation studies pedestrian limbs self-occlusion detection using Skin-color feature. Skin-color feature shows some regularity on the different stats in the cause of human motion. The algorithm combines the prior knowledge of human motion with the analysis of the regularitie of skin-color feature, classfies human walking into four self-occlusion states according limbs self-occlusion, and transforms the detection of pedestrian limbs self-occlusion to the calculation of the self-occlusion state transition probability via hidden markov model.4. How to estimate the human on different stats using hybrid tracking model is studied. Traditional methods train samples to slove the problem of self-occlusion in the course of human motion. Hybrid tracking model trains limb motion correlation model online, uses the result of detecting limbs self-occlusion to estimate the human on different stats via different tracking algorithms, and shows good profermance.The proposed hybrid tracking model does not need a large number of sample data to train limb motion correlation model, can be adaptive under multi states of the human body motion estimation. The contributions made in this thesis realize automatically estimated human walking in door at the beginning step, is the initial exploration of whole automatic estimation of human 3D motion, and obtain good result.
|
PDF Full Text Request