Two dimensional upper human body pose modelling system

• Main Author: Rosalyn R Porle
• Format: Thesis
• Language: English; English
• Published: 2011
• Subjects: TA1501-1820 Applied optics. Photonics
• Online Access: https://eprints.ums.edu.my/id/eprint/39134/1/24%20PAGES.pdf; https://eprints.ums.edu.my/id/eprint/39134/2/FULLTEXT.pdf; https://eprints.ums.edu.my/id/eprint/39134/

Human Body Pose Modelling detects the human body parts and estimates their size, position and orientation in image sequences and then represents them using a specified model. It has wide applications such as user interface system, intelligent visual surveillance system and motion analysis in sports and medical applications. The human body can be divided into three main parts, which are the head, the torso and the limbs. In this thesis, the head, the torso and the arms of the human body are detected from the captured video. In this system, the human subject is assumed to be alone, in standing position, facing the video and wearing a short-sleeve clothe. The body parts are estimated and then represented in an image using rectangle shape. The input image sequences, which are acquired from the video, are processed using Single Level Haar Wavelet Transform decomposition. From the decomposition result, three features; namely the silhouette, the edge and the colour are extracted. Silhouette extraction, wavelet-based edge extraction and histogram analysis methods are employed to extract these features respectively. Rectangular shape is used as a model for each targeted body part. The parameters of the model such as the corner position, angle of rotation, width and length are computed using pose estimation methods. Four methods are proposed to estimate the head and the torso. These methods are Windowing Method I, Windowing Method II, Histogram Analysis I and Histogram Analysis II. All these methods use the extracted silhouette feature. Then, the arms pose is estimated using template-based matching. The arms pose are classified into non-occluded and occluded pose. Depending on this classification, different input features are applied. Silhouette or edge feature can be used for the non-occluded pose estimation. Meanwhile, colour feature is used for occluded pose estimation. The developed system is tested using 30 image sequences of different users. The best method for the silhouette extraction achieved 88.62 percent of correct silhouette detection and 98.68 percent of correct non-silhouette detection. For the colour extraction, the best colour format achieved 55.01 percent of correct skin colour detection and 97.71 of correct non-skin colour detection. For the pose estimation, the best pose estimation method for specified body parts achieved more than 90 percent of correct estimation. The proposed methods deals with: 1) partial occlusion, e.g. when the torso is not detected or partially detected, the size and position of the torso can still be estimated; 2) self occlusion, e.g. when the upper and lower arms are occluded with each other; and 3) hand crossing, e.g. when the right and left arms crossed with each other.