Graphics Reference
In-Depth Information
Fig. 5.1
Preliminary experimental results for three algorithms implemented in the target computing
architecture. Disparity maps are concerned with frame
66 of the KITTI dataset [
10
], using a
simple x-Sobel filter as prefiltering step. From
top
to
bottom
rectified reference image, disparity
map computed by the FW implementation, disparity map computed by a modified version of the
[
5
] algorithm using two paths, and disparity maps computed by a modified version of the SGM [
13
]
algorithm using four paths. Additional experimental results are available at this link:
http://www.
consumer/embedded applications due to their high power requirements, cost, and
size. Computing architectures, such as those based on high-end FPGAs, are often
too expensive as well, while solutions based on custom
application specific inte-
grated circuits
(ASICs), despite the limitations regarding their reconfigurability and
time to market
compared to FPGAs, represent a less expensive solution in large
volumes. Finally, we point out that interesting low-power, low-cost, reconfigurable
architectures for real-time dense stereo vision are represented by embedded CPUs
coupled with integrated DSPs, such as the OMAP platform [
11
], extensively used for
stereo vision. A recent and detailed review of stereo vision algorithms for different