Efficient 3D voxel reconstruction of human shape within robotic work cells

In this paper an efficient voxel reconstruction method within a robotic work cell by means of shape from silhouette is presented. A procedure is proposed which includes occlusions from the robot's kinematics into the reconstruction step, i.e. the time-dependent occupied space. Axis values together with surface models from the robot are used for modeling occlusions coming from the moving device. An image-based method for simplified in-place frame calibration is shown and compared to a regular marker-based method. A learning method for background models is introduced, which includes overlapping image areas from the robot and offers improved learning rates. For efficiency reasons the implementation was done on the GPU. The achieved performance using five cameras in an industrial setting is shown and discussed with respect to quality and frame-rates.