A Template Matching and Ellipse Modeling Approach to Detecting Lane Markers - Advanced Concepts for Intelligent Vision Systems - page 182

varyinshapeandcolor,thetemplatematching procedure is repeated using the
remaining combination of templates and color channels.

4.3 Morphology

The residual labeled regions acquired after template matching are fed as inputs
to the Lane Region Finder (LRF) as shown in Fig. 5a. The labeled regions of the
Y, Cb, and Cr channels are prefixed with “lbl” e.g. lblY, lblCb, and lblCr. The
default mask passes all labels untouched (all-pass). After each LRF iteration, the
mask is updated to ignore areas that have already been detected. This updated
mask is then fed to the next LRF module. Finally, the results of all three LRFs
are combined using a bitwise OR operation to produce Ensemble Lane Regions
(ELR).

lblY

lblCb Mask lblCr

Inputs

Residual Label Growing Residual Label Growing

Double Lane Marker Normal Lane Marker Wide Lane Marker

lblY
lblCb
lblCr

lblY
lblCb
lblCr

lblY
lblCb
lblCr

Default
Mask

Yellow Lane Region

Dilate & Invert

Lane Region Finder Lane Region Finder Lane Region Finder

Residual Label Growing

Updated
Mask

Updated
Mask

White Lane Region

Lane
Regions

Lane
Regions

Lane
Regions

Dilate & Invert

Outputs

Ensemble Lane Regions

Updated Mask Lane Regions

(a) Flowchart illustrating the creation (b) Flowchart illustrating the inner work-

ings of the Lane Region Finder (LRF).

of ensemble lane regions.

Fig. 5. Steps involved in the Morphology

The ordering of input labels plays an important role in the success of the lane
detector as the mask used in LRF is updated after each operation. When an
off-line test involving the computation of NCC coecients between equal sized
test samples and their corresponding templates was conducted, on average the
double lane templates produced the highest cross correlation score in the center
of the coecients map. This was followed by the scores of narrow and then wide
lane templates. The double lane templates were able to produce a higher cor-
relation score as multiple intensity oscillations in the vertical direction account
for strong discriminating features in comparison to the other templates. As a
result, the LRF sequencing in Fig. 5a starts by detecting double lane markers.
The updated mask is used in the LRF operation in collaboration with residuals
labels corresponding to normal lane markers, and then with wide lane markers.