Flexible body scanning without template models

The apparition of low-cost depth cameras has lead to the development of several reconstruction methods that work well with rigid objects, but tend to fail when used to manually scan a standing person. Specific methods for body scanning have been proposed, but they have some ad-hoc requirements that make them unsuitable in a wide range of applications: they either require rotation platforms, multiple sensors and a priori template model. Scanning a person with a hand-held low-cost depth camera is still a challenging unsolved problem.

This work proposes a novel solution to easily scan standing persons by combining depth information with fiducial markers without using a template model. In our approach, a set of markers placed in the ground are used to improve camera tracking by a novel algorithm that fuses depth information with the known location of the markers. The proposed method analyzes the video sequence and automatically divides it into fragments that are employed to build partial overlapping scans of the subject. Then, a registration step (both rigid and non-rigid) is applied to create a final mesh of the scanned subject. The proposed method has been compared with the state-of-the-art KinectFusion [1], ElasticFusion [2], ORB-SLAM [3, 4], and BundleFusion [5] methods, exhibiting superior performance.