A novel solution for registration of stereo digital torso images of scoliosis patients.

This paper presents a procedure for registration of a pair of stereo digital images giving an improvement in accuracy and speed over existing methods. It does so by a novel approach combining color based image segmentation and differential geometry. It involves three stages: image segmentation, adaptive local pixel matching, and deferential geometry in a tree weighted belief propagation procedure. The registration was compared to 2 existing registration procedures, segment-based adaptive belief propagation (adaptive BP) and color-weighted hierarchical belief propagation (hierarchical BP). A 3D scan of a mannequin was obtained and errors in reconstruction were measured for each of the 360 cross sections of the mannequin. The proposed procedure outperforms existing methods, particularly for high curvature regions and significantly large cross sections. Its accuracy of reconstruction ranged from 85-100% compared to 75-100% for other existing methods. It was 35% to 40% faster. This work provides a solution to the registration problem and is an important step in developing a cost effective technique for measuring torso shape and symmetry of scoliosis patients using stereo digital cameras.