Zouhour Ben Azouz
Authors: Chang Shu and Zouhour Ben Azouz
Institute for Information Technology, Visual Information Technology,
National Research Council of
Body scanners provides tremendous amount of information about the shape of the human body. Processing the surface data collected by the sensors, however, proves to be a challenge. The main difficulty is that the digitized models have different number of points and there is no correspondence between different models. Anatomical landmarks are important features to establish this correspondence. These points are efficiently identified on the 3-D models when markers are placed on the body before scanning. The marking process is time-consuming and most likely will be eliminated in future 3-D anthropometric surveys. In this paper we present an approach for automatic locating of anthropometric landmarks on 3D human scans. Our method is based on learning landmark characteristics and the spatial relationships between them from a set of human scans where the landmarks are identified. The learned information is formulated by a pairwise Markov network. Each node of the network is a random variable corresponding to the position of a landmark. The edges of the network represent correlations between the positions of landmark pairs. Probabilistic inference is then performed over the Markov network to locate the landmarks. We evaluated the algorithm on 30 human models with different shapes. The results showed good accuracy for most of the landmarks.