Robust motion from space curves and 3D reconstruction from multiviews using perpendicular double stereo rigs

摘要

This paper proposes a robust curve based method to reconstruct 3D model of an object from image sequences captured by two perpendicular stereo rigs. First, corresponding points and the geometry of points are computed in stereo images by extracting unique space curves. A new algorithm is proposed to extract unique space curves from plane curves in stereo images based on curvature and torsion consistency. The proposed method provides accurate geometry of the curve points with extremely reduced number of outliers. Contrarily to the standard sparse approaches that need sub-pixel accuracy to compute structure and motion, the proposed curve matching method deals with pixel accuracy information. More importantly, it finds the correspondence based on curve shape and does not use any photometric information. This property makes the matching process very robust against the color and intensity maladjustment of stereo rigs. Second, the recovered space curves are employed to estimate robust motion by minimizing the curve distance in the next sequence of stereo images. An efficient structure of stereo rigs - perpendicular double stereo - is proposed to improve accuracy of motion estimation. We discuss and prove its properties mathematically. Third, a set of calibrated virtual cameras are constructed from estimated motion information to take advantage of the shape-from-silhouette using multiple views and extract the object's visual hull as fine as possible. As a whole, a complete automatic and practical system of three-dimensional modeling from raw images captured by calibrated perpendicular double stereo rigs to surface representation is proposed. Fine motion estimation is the main advantage of the proposed method using perpendicular stereo rigs, which makes use of the space curves in a large base line camera setup. While the previous methods of motion estimation suffer from the statistical bias due to quantization noise, measurement error, and outliers in the input data set, the proposed method overcomes the bias problem even in pixel-level information. Experimental results demonstrate the privileged performance of the proposed method for a variety of object shapes and textures.