Architectures for Stereo Vision - Signal Processing Systems - page 474

Considering all aspects (algorithmic performance, implementation performance,
architectures) a huge design space is unfolded. For embedded systems the choice is
invariably on low-power solutions, e.g. based on application specific architectures
implemented on FPGAs or ASICs. However, with the recent rise of GPUs for
high performance computing, GPUs offer a cost-efficient alternative for stationary
systems where power consumption is not an issue.
This chapter addresses high performance disparity estimation considering both,
algorithmic and implementation performance. The chapter is structured into an
algorithmic and an architectural section; these being Sects. 2 and 3 . An introduction
to the fundamental principles of the stereo image matching ( epipolar geometry )
and a minimal practical stereo vision system is given in Sect. 2.1 . The algorithmic
and architecture sections both give a comprehensive overview of recent works.
It is followed by a detailed discussion of the semi-global matching algorithm (SGM)
[ 37 ] (Sect. 2.3 ) and two exemplary implementations on FPGA (Sect. 3.7 ) andGPU
(Sect. 3.6 ) , respectively.

2 Algorithms

A minimal system for disparity estimation from a real camera setup consists of
two processing steps: The first step is camera lens undistortion and rectification of
non-ideal stereo camera setup (Sect. 2.1 ) while the second step is the actual stereo
matching (Sect. 2.2 ) . All other image preprocessing steps (e.g. noise reduction,
equalization) and disparity map post-processing steps (e.g. whole filling, interpo-
lation of pixel with missing stereo information) are optional.

2.1 Epipolar Geometry and Rectification

The objective is to find corresponding pixels in the two images of a stereo camera
setup. Due to the underlying epipolar geometry [ 34 , 83 ] of a stereo camera setup,
the search space for corresponding pixels is one-dimensional. As shown in Fig. 2 a ,
for a given pixel in the base image all potential correspondences project onto the
epipolar line ( e bm ) in the match image and vice versa. Strictly speaking the possible
projections are bound by the epipole and the viewing rays for a real-world point at
infinity.
For efficient correspondence search implementations a preprocessing step, the
rectification , is employed. Both images are warped such that epipolar lines in both
images are parallel to the scanlines and are row-aligned, i.e. corresponding pixels
are in the same horizontal line [ 34 , 103 ] . Thus, efficient memory access patterns and
parallelism over independent scanlines can be obtained. After rectification, the focal
axis are parallel to each other and perpendicular to the line joining the two camera
centers ( baseline ) and the disparity for points at infinity is 0.