Model driven segmentation and registration of articulating humans in Laplacian Eigenspace

We propose a general bottom-up approach using Laplacian eigenmaps to segment 3-D voxel data of human subjects into different body parts. The voxels are first transformed into a high dimensional space which is the eigenspace of the Laplacian of the neighbourhood graph. We exploit the properties of this transformation and fit 1-D splines to the voxels belonging to different body segments in eigenspace. We show that the properties of Laplacian eigenmaps are particularly suitable for the purpose of fitting 1-D splines and segmentation. The boundary of the splines is determined by examination of the error in spline fitting. We then use a probabilistic approach to register the segmented body parts by utilising their connectivity and prior knowledge of the general structure of the subjects. The probabilistic approach combined with the properties of the transformation enables us to deal with complex poses. We present results on real data, containing both simple and complex poses, as well as synthetic data, which helps us evaluate the algorithm. We also present algorithms to automatically estimate the human body models using the output of the segmentation algorithm. While we use human subjects in our experiment, the method is fairly general and can be applied to voxel-based registration of any articulated or non-articulated object which is composed of primarily 1-D parts. Index Terms Markerless motion capture, Human body model estimation, Pose estimation, Segmentation.