Research Of Human Pose Estimation In Still Image

Human pose estimation is an essential issue in computer vision area since it hasmany applications such as human activity analysis, human computer interaction andvisual video surveillance, it main purpose of human pose estimation is that detect theposition、scale and direction of parts of people.Human pose estimation is often approached in avideo setting, within the context of tracking. Recent focus in the area has expanded tosingle-image pose estimation, because such algorithm are likely useful for tracker(re)-initialization. In this dissertation, vision-based human body estimation isinvestigated. Main contributions of this thesis are follows：(1)First the general representation of the object and several image feature extractionalgorithms are introduced, such as texture, edge, color information features of the image,and also several image matching algorithms are introduced.(2)Because of human detection can provide a lot of prior knowledge for human bodyestimation, such as position information, so in this thesis, we introduce a humandetection algorithm named extended histograms of oriented gradients which is proposedbased on differences in human characteristics which extract the HOG descriptor of imagewith different blocks, this method give a better performance in dense population thantraditional HOG method which have a lower detection rate in dense population.(3)Then a framework of human body estimation based on pictorial structure isprovided. This kind of probabilistic framework mainly for modeling appearance and poseof the human body and supports computationally-efficient inference, and introduce apictorial structure model and statistical model of human, the model parameter estimation.(4)At last, a new part appearance for human estimation is proposed. Because of theresult in the iterative human pose estimation based on tree-structure model is susceptibleby the background, In this thesis (consider static images), Part Appearance Model isimproved, a appearance model based on color and texture for iterative human poseestimation is proposed, Experimental results show this method give a better performanceand Accuracy while reduce the search space by people detector and grab cut.