Architectures for Stereo Vision - Signal Processing Systems - page 482

Fig. 6 Processing steps for disparity estimation using rank transform, semi-global matching, and
optional median filter. Numbers in parenthesis refer to the respective equations and R denotes the
number of paths

minima. The summed path costs S (thick black line) clearly identify the

minimum at a disparity level of 32 resolving all ambiguities. However, the cost
difference for the positions 32 and 33 is minimal indicating that the correct position
is located a sub-pel precision.
Finally, left/right check according to ( 3 ) and post processing can be applied, e.g.
a median filter in its most basic form. An overview of the processing steps is given
in Fig. 6 .

( p , d )

3 Architectures

The variety of architectures and implementations to compute the stereo correspon-
dence easily rivals the variety of the underlying stereo matching algorithms. Today
very efficient implementations for local and global stereo methods are available on
FPGAs, ASICs, GPUs, and DSPs. For real-time image throughput, local methods
have been and continue to be favored by many researches because of their efficient
implementation possibilities. However, with advances in computational power,
many global methods are also implementable in real-time.
Early work includes a complete stereo vision system from 1996 featuring
rectification and stereo matching with an SSD variant on a custom hardware board
consisting of off-the-shelf components and a DSP array [ 49 ] . At 30 fps 200 200

images with 5 bit depth resolution could be computed. Other noteworthy early
implementations have been presented in 1993 using a DSP array [ 21 ] and in 1997
using a single DSP [ 52 ] . In [ 52 ] also an early overview of implementations is
provided. An FPGA array was used in 1997 to implement a census transform
stereo matching method [ 95 ] . All of these implementations directly compute the
disparity information from the matching costs without cost aggregation. Early work
includes a complete stereo vision system from 1996 featuring rectification and
stereo matching with an SSD variant on a custom hardware board consisting of
off-the-shelf components and an DSP array [ 49 ] . At 30 fps 200

×

× 200 images with