Profile Recovery System Using a Differential Identity of Binocular Stereo Images

A Parallel Algorithm and Its Experimental Evaluation

Abstract

This paper describes a new gradient-based depth recovery algorithm for binocular vision systems with emphasis to the usage in real time measurement environment. Two image sensors are assumed to be matched mechanically. Their common object planes define a reference plane of height on the 3-D object. Extracting a relative height distribution with respect to this plane supplies us extensive information about the object shape near the plane. Approximating local variation of object intensity and describing it using left and right images, we obtained a differential identity with good symmetry which has a meaning that a ratio of a difference image of left and right images against a derivative of a sum image is equal to the disparity. Applying local least squares estimation, we obtained stabilized identity among statistical correlation between the sum and difference images. Advantage of parallel structure of this algorithm is emphasized for applications to dynamical binocular sensor systems. This algorithm was tested with experimental image data. Depth recovery results as well as evaluation of accuracy, resolution, and rangeability are shown.