Studies On Three-Dimensional Model Based Posture Estimation And Tracking Of Articulated Objects

In this thesis, the problem of estimating three-dimensional posture and motion of articulated object in video sequences is studied. It plays an important role in lots of applications in areas like motion capture for animation and virtual reality, video indexing, natural human-computer interaction and robotic control. This research aims to estimate full-degree-of-freedom posture and motion of articulated objects in high-dimensional and multi-model state space from monocular video sequences. The complexities of articulated structure, shape, and the large ambiguous in image observations caused by the loss nature of camera projection and high-dimensionality required for three-dimensional representation, render the solution non-trivial.To address these problems, a model is proposed that incorporates realistic three-dimensional shape model, full-degree-of-freedom kinematics, linear dynamic model, and robust global observation feature. Three novel search techniques are derived respectively, which aim for posture initialization, motion tracking and simultaneous estimation of motion parameters and mode parameters.In posture initialization aspect, estimation is regarded as an inverse process of model-based observation generation. Bayesian estimation is introduced as the whole framework and a regression integrating with optimization approach is proposed. With the initial learning of state distribution in regression step, evolutionary computation achieves much quicker and more efficient global search in high-dimensional state space. Motion tracking, as a sequential estimation problem is always solved with a filter. Again, the high-dimensionality is the main obstacle that current filter methods have encountered. Evolutionary particle filter is proposed to deal with the tracking of complex high-dimensional articulated motion. With the global observation model and global search concept, the filter makes a significant improvement of tracking performance compared with conventional tracking methods. Automatic update of model parameters is a key problem in a model-based visual analysis system. General state space model is introduced for simultaneous estimation of dynamic posture and fixed model parameters in video sequences. An EM framework with particle filter and estimation in fixed block window is proposed to overcome the conflict between online posture estimation and stable model parameter estimation.The methods give general approaches to the estimation in the high-dimensional multi-model and non-convex general state space so often encountered in computational vision or other areas.