A novel computational model of stereo depth estimation for robotic vision systems

Abstract. This paper presents a novel computational model of stereovision for improving the accuracy of three-dimensional data extracted from a stereo-pair image with no eect of changes in focal length. For decades, most previous studies on stereovision have focused on the establishment of stereo matching, and have made conclusions on the premise of a xed focus. In general, error in the depth estimate becomes bigger when the focus and aperture are unknown or not xed. For that reason, a three-stage framework is proposed in this paper to modify the conventional stereovision model for improving the accuracy of depth estimation. The rst stage is to modify the computational model of conventional stereovision for varifocal cameras. Then, the spacing of depth intervals in the non-uniform spacing of discrete depth levels can be altered, in particular, to be unaected by changes in focal length. Finally, by considering the ane transformation, we add the deformation coecient into the modied stereovision model for correcting three-dimensional ane deformations. Experimental results demonstrated that the depth estimation from stereo images using the proposed scheme was more accurate than conventional methods. The percentage error of most estimates fell between 0.06%-0.82%, and the error value increased from 0.02 cm to 2.21 cm within 6 m.