On-line Generation of Customized Human Models based on Camera Measurements
Koen Buys, Dorien Van Deun, Tinne De Laet, Herman Bruyninckx · 2011
This paper presents two algorithms, one monocular-camera and one multiple-view camera based, to estimate several anthropometric parameters of human skeletons underlying the human avatar consisting of a human skeleton and a mesh model. This results in an on-line implementation, which can use cheap sensors to give adequate results for biomechanical tracking with only four to seven parameters in the estimation process. The first, single-view camera based (monocular-camera), algorithm uses the MakeHuman (Bastioni and Flerackers, 2007) mesh model to estimate several anthropometric parameters that are directly or indirectly observable, such as length, hip and waist circumference, age …. The second, multiple-view camera based, algorithm uses a voxel-based method to estimate both the length parameters and the weight of the person. This is achieved by a multiple-view visual hull reconstruction that overfits the person in order to iteratively increase the number of voxels to better match the actual human's shape. This voxel-technique is furthermore used for person segmentation (head, torso, arms, and legs). The proposed methods are currently being extended to combine visual images with a depth image (Microsoft Kinect) or a laser range image (Hokuyo), since these sensors directly provide a point cloud of the person, hereby replacing the otherwise computational expensive voxel technique.