Architectures for Computer Vision: From Algorithm to Chip with Verilog

6 Stereo Vision

Binocular stereo vision is one of the major vision modules by which one can induce the depth of the surface shape and the volume information of the objects. Created by a pair of cameras, a conjugate pair of images contains the depth information by means of disparity. Binocular vision naturally expands to other vision, such as trinocular or multi-view vision (Faugeras 1993; Faugeras and Luong 2004; Hartley and Zisserman 2004).

Stereo vision deals with the three major problems: correspondence geometry, camera geometry, and scene geometry. Of these, stereo matching deals with the correspondence geometry and remains the major research area. It can be classified into various categories according to the features, measures, inference methods, and learning methods. A comprehensive survey of recent progress on stereo matching is presented by (Scharstein and Szeliski 2002) for algorithms, and by (Tippetts et al. 2013) for realizations. Also, they have provided a benchmark for quantitative evaluation of existing stereo matching algorithms.

This chapter introduces some fundamental concepts of stereo vision problems and stereo matching. Instead of reviewing all the extensive stereo matching algorithms and realizations, we focus instead on the fundamental constructs of the energy function, classified as appearance model and geometric constraints. The models and constraints are examined in five categories: space, time, frequency, discrete space, and other vision module. Stereo ...