A Distributed Adaptive Architecture for Analog Stereo Depth Estimation

In this paper we present an architectural description of a visual sensor capable of extracting a dense depth map of the scene under construction. The goals of the stereo depth perception is twofold: find an estimation of distance for each point in the image and a confidence measure of this operation. Stereoscopic depth estimation is based on disparity measurements. At functional level, in the flow of processing, we distinguish: (i) adaptive convolution with Gabor filters for measuring phase, (ii) regularization process based on local context information, (iii) a comparison between left and right phase maps to determine disparity. The computation in step (i) occurs in two stages: the first, based on a linear space-invariant array of processing units, accomplishes a convolution filtering with kernels of even symmetry at the dominant spatial frequency of the image; the second one performs a local convolution between the image and a Gabor-like filter of given phase. Such decomposition allows to map efficiently the space-variant filter over a lattice of computational nodes by combining cooperatively the outputs of the computational substrate implemented in the first stage. For step (ii), one can extract the local phase of the image at the frequency specified by the Gabor filter resorting to an adaptive phase-matching technique in space, and then regularize these results. The architecture has been validated respect to both artificial and real-world images