Robust sequential surface based object segmentation in noisy range data

A range image consists of noisy, discrete samples of object surfaces. The perception of surfaces plays a key role in range image understanding and 3D object recognition. A statistically robust and computationally efficient function approximation scheme for surface segmentation is developed. In the existing surface characterization techniques, surface completion is done after segmentation, using heuristics and rules to piece together parts of the complete surface. Our novel method generates the complete surface hypotheses in the parameter space in one step. Object segmentation is automatically achieved in the reduced parameter space instead of the image space. The algorithms are tested and supported by extensive experiments on real and synthetic depth maps that exhibit surface coherence property.<>