Stereo Vision Algorithms Suited to Constrained FPGA Cameras - Advances in Embedded Computer Vision - page 122

Fig. 5.11 The four paths used to compute permeability terms and aggregated cost for algorithm
[ 5 ]
techniques [ 21 ] potentially suited to the constrained FPGA architecture, thanks to the
reduced (and constant) number of operations required with respect to explicit cost
aggregation approaches inspired by bilateral filtering. Despite these positive facts,
the results provided by these constant time algorithms are comparable to those based
on explicit cost aggregation, and hence these algorithms are potentially suited for
implementation in the outlined target platform.
Concerning FPGA implementations based on adaptive weights strategy, [ 8 ]
reported a hardware friendly implementation of the original AW approach [ 46 ].
In [ 24 ], a further simplification of the AW approach based on simpler binary weights
was devised. A method based on the mini-census transform for cost computation
and the two-pass approach [ 43 ] for cost aggregation is [ 4 ] while [ 49 ] used, with the
same cost function, the two pass orthogonal cost aggregation strategy proposed in
[ 48 ]. An FPGA implementation of a cost aggregation strategy based on segmenta-
tion is reported in [ 38 ]. Finally, an interesting method based on adaptive weight cost
aggregation, identification of reliable points and disparity refinement, by means of
an effective method aimed at enforcing local consistency [ 20 ] of the disparity field,
was proposed in [ 16 ].
5.5.1.3 Algorithms Based on Unconstrained Supports
According to the taxonomy provided in [ 29 ], the algorithms reviewed in the previous
sections clearly belong to the class of local algorithms. However, there are some
local algorithms that significantly diverge from traditional approaches, in particular
for what concerns the support regions used for cost aggregation.
An interesting local approach, referred to as Permeability, was proposed in [ 5 ].
This technique performsmultiple 1Dcost aggregations constrained by an information
permeability term (see [ 5 ] for a detailed explanation) computed along horizontal or
vertical scanlines, as shown in Fig. 5.11 , without enforcing any explicit smoothness
term. The permeability term, computed along the horizontal scanline from left to
right is defined as follows: