Research Of Human Pose Estimation Based On Pictorial Structure Models

The purpose of human pose estimation is detecting the position, scale anddirection of parts of human body. It has a wide range of potential applications,including intelligent monitoring, advanced man-machine interaction, imageannotation and so on, making it arguably one of the popular problems in machinelearning vision. Due to the diversity of human body and human body pose, humanpose estimation refers to multiple field of view, so to effectively represent it need tobuild many models and to detect it needs much time.Lately, the most popular methods of human pose estimation are based onpictorial structure (PS) models. PS models decompose human body into a collectionof parts and connections between certain pairs of parts, among which, parts aremodeled by the appearance models and connections are modeled by the deformationmodels. Based on PS models, inference can estimate human pose estimation.Based on PS models, researches were taken to achieve better human poseestimation. The following is the main work:(1) Based on graph models, this article introduced the theories of two commonmodels of pose estimation， DPM model and PS model, and decided to choose PSmodel after making a comparison. Finally this article chose the standard protocol toevaluate the performance of the algorithm.(2) To overcome the problem that classic PS models can’t model human bodyvery well and the problem that the hypothesis space of human pose estimation is toolarge, this article augmented classic spring models by modeling the relationshipbetween direction of the children of the neighboring parts and the relative position ofthem, which is called mixtures of PS models. The model captures some local rigidityof human body and it takes advantage of some priori-knowledge of human body.Finally this article evaluated results using the Image Parse dataset and the Buffydataset, and the experimental results express that this model improved the accuracy ofhuman pose estimation.(3) To overcome the problem that some poses can hardly be distinguished in asingle frame, this article introduced local motion of joints into human skeletal pose to estimate human pose based on temporal continuity of sequential frames. This methodcomputed the motion using a histogram of oriented space-time gradients, andcombined it with skeletal pose to form dynamic pose. Then this article used bag ofvisual words to generate some visual words, and represented sequential frames ashistograms of frequency of these visual words, and took them as the input of classifierfor training and classification. Finally this article evaluated results using the KTHdataset, the UCF-Sports dataset and the Weizmann dataset, and the experimentalresults express that this model improved the accuracy of human pose estimation.