Human motion tracking has been received increasing attention from researchers in the fields of image processing and computer vision during the past few years. It has a lot of applications in video edit, virtual reality, advanced user interface and human motion analysis, etc. The procedure of the articulated human motion tracking and analysis in monocular action image sequences involves three main stages: (1) human body segmentation in a complex scene; (2) human motion tracking and body structure re-construction; (3) motion analysis and action recognition. As the base of the human action recognition and understanding, human motion tracking and body structure re-construction is the key of the whole procedure. Action recognition is the highest level of motion analysis, but now it is far from application. The researching work mainly focuses on the first stage and second stage. This paper proposes a novel method of tracking human motion from a single view. The correspondence between the human region of the image and the human body model is established through analyzing the feature extracted from the image sequences and estimating the human body pose, then optimizing the whole state sequences in the form of motion trajectory, which makes the MAP (Maximum A Posterior) solution of the whole state sequence of the pose efficiently obtained in the recursively pose updating process, and ultimately recovering the sequence of human motion. In this thesis, we study human pose tracking from monocular image sequences which involves some basic problems in image processing, pattern recognition and energy function optimization, etc. The main contributions of this thesis include the following issues: No learning process of human model prior from image database is needed, and no limits on the video scene are required. By integrating a post-process called K-Best Viterbi Search into the conventional Viterbi, the global constraints are elegantly added to get reasonable results while the time complexity of the inference keeps quite low. The pose ambiguity is carefully analyzed, and the missing data caused by occlusion is connected by the trajectory smoothing. In a word, in this thesis, we have made fruitful attempts and satisfactory progress on tracking articulated human motion in monocular action sequences.
修改评论