Abstract: At present, existing single-image-based human body modeling methods still cannot
effectively deal with the complex occlusions of body parts either due to the arms or clothes or to the
changes of viewpoints. To solve this problem, using the distribution characteristics of skeletal joints
in the SMPL model, we designed a human body reconstruction method by improving the traditional
segmented hinge transformation model. The method uses the accurate annotation of the skeletal joints
to identify the node of the model transformation and combines the image contour boundary constraint
map to propose the non-rigid registration method of forward piecewise regression and probability
expectation minimization (FPR-PEM). The iterative model was used to linearly interpolate the thin
plate splines at the deformed joint to ensure the independence of the point cloud shape on the model
surface, which effectively registered non-rigid deformation models under various postures and better
solved the reconstruction challenges brought by complex occlusion. Then regression adjustments
were performed with regard to the model posture so as made to achieve accurate human body
modeling. Experimental results show that the proposed method works effectively to build a fine and
smooth model of human body reconstruction.

Key words: human body reconstruction, piecewise hinge, non-rigid registration, probability model;
SMPL