研究目的
Investigating the capability of reconstructing non-rigid 3D models of a human body by using a single depth camera with skeleton information.
研究成果
SegmentedFusion is able to effectively handle fast motions and topological changes in reconstructing non-rigid human models by using a single depth camera with skeleton data. The system segments the human body into multiple nearly rigid parts and builds the canonical bounding box for each part, using bone motions to estimate a volumetric 6D motion field.
研究不足
1. Cannot reconstruct twist motions of the arms due to insufficient accuracy of skeleton information and lack of salient geometric features for the ICP algorithm. 2. Fails in the case where large occlusions exist such as when crouching due to segmentation failure. 3. Some gaps appearing in the armpit due to the weights that blend the transformation of parent and child bones producing some collapse of the 3D model.
