Graphics Reference
In-Depth Information
operations in filtering n B scan bands of width w B each and length N w is equal to
2 n B w B N w k 2 . The next step involves horizontal projections in each band, which is
w B N w n B addition operations. The shift and add operation involves N w multiplica-
tions and comparisons per band resulting in a total of 2 N w n B operations. Therefore,
the total number of operations for lane feature extraction in the proposed method is
given by
k 2
N prop =
2 n B w B N w (
+
1
)
(10.2)
This is a simplified model but it is sufficient to evaluate qualitatively the effect of
scan bands on the overall computation cost efficiency. Figure 10.7 shows a scatter
plot between number of operations N prop and the detection rate for different possible
number of scan bands and scan band widths. The top left corner in the graph, i.e.,
high accuracy but less number of operations, is the ideal place to be in and we can
seeinFig. 10.7 that using eight scan bands of width w B =
5 gives similar detection
rate as eight bands of w B
10 but at 50% lesser number of operations. Also,
when compared to conventional methods wherein the entire image is processed for
filtering alone, the proposed method gives orders of magnitude savings in the number
of operations. Other constraints for embedded realization such as total computation
cycles, latency, energy cost, total memory accesses, etc., are also directly related to
the number of operations by different factors.
=
Fig. 10.7 Number of operations versus detection rate for different scan band widths in Set 1 and
Set 2
Search WWH ::




Custom Search