This paper presents a novel 3D image-based system for capturing full human body models. The system uses three Microsoft Kinects, a turntable for rotating the human object, and multi-view registration algorithms to fit the Kinect data to the human anatomy.
Kinect is used to capture depth information. To ensure better quality of the depth data, three Kinects are used to capture the body profile in parts. As the human object turns with the table, the three Kinects capture multiple synchronized views (frames) for the respective body parts. These frames are then stitched together using feature points obtained through the optical flow analysis of pairs of consecutive frames. This is then suitably post-processed using global deformation registration to resolve the errors accumulated over complete rotations.
Using this 3D imaging system, the authors have generated multiple different 3D full-body models. In general, the quality of the models is good, and notably, many of them capture delicate geometric details such as faces, dress, and hairstyles. However, the authors wish to use more sophisticated denoising and super-resolution approaches to further improve the quality of the models. They also plan to improve the registration algorithm to completely eliminate the misalignment problem.
Overall, the paper presents a low-cost, good quality 3D full-body modeling system for humans, with a lot of potential for domestic applications.