The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3h. Beijing 2008 
With the increase of the survey image sets, the size of the 
hypothesis set might quickly explodes. Hypothesis management 
step is designed to keep a practical-sized hypothesis set and 
keep the diversity of that hypothesis set as much as possible. 
Several rules are introduced: first, only limited amount of 
hypothesises are maintained; second, the hypothesises that are 
not changed for several image sets will be combined into other 
hypothesises, or deleted; third, only limited hypothesises are 
allowed to be added at each image set. 
The MHA processing goes through all the image sets. At the 
end, H* —the hypothesis with highest likelihood— will be 
viewed as the best connection configuration of all the LLS 
notes. Each links in H* is a lane line. 
7. LANE LINE CLASSIFICATION 
The classification of LLS attributes is affected by many facts: 
occlusion; worn-out painting of lane lines; variation of the side 
pavements or grass; variance of road surface materials; and the 
unreliability of the feature extraction algorithms. So, instead of 
classify the LLS, we calculate the feature of a lane line base on 
all the LLSs included in that lane line; and classify the type of 
the lane line as a whole. 
Lane line usually extends for hundreds of meters or even 
kilometers. According to the traffic design, the type of lane line 
may change during the extension. For example, a dashed white 
lane line may change into a solid near road crossings, to keep 
the vehicles from lane change. However, the above mentioned 
lane line detection and linking procedures are not able to 
separate the lane line type changes. In order to solve this 
problem, type-changed point detection is introduced in our 
system. 
Denote a lane line as LA, LA includes a set of LLS, LA = { /,• | 
i=0,n}. Denote judge function Ej as : 
E = j° ™ hen F(l i _ d ,...,l i ) = F(! i _ d ,...,l i ) (5) 
' jl when 
where d is the buffer size. p(J k ) is the classification 
function that decide the lane line type based on the 
characteristics of LLSs {/^ / j. In practice, we use a KNN 
classifier as f(). If the preceding lane line segments 
l t _ d and the successive lane line segments / ; - 
are not with the same type, then E j is 1, and LLS / is viewed 
as a point where the lane line type changed. Figure 9 illustrate 
the finding of type-changed point. 
h-df“'h //>••• h+d 
A 
• 
a 
• 
w 
Figure 9: detection of type-changed point 
Once a type changed point is detected, the lane line will be 
broken at that point. The separated two parts of the lane line are 
to be classified independently. 
areas. Test results show that ARVEE is robust and ready to 
serve the real world applications. Video of the results can be 
found at http://mms.geomatics.ucalgarv.ca/Team/Current/ 
Collaborators/cheng/AVREE demo/ARVEE demo.htm 
Figure 10 shows a road geometry extraction result of ARVEE. 
The extracted lane lines are superimposed on the original 
images (only two of the four cameras are shown in the figure). 
There are four lane lines within this site, all correctly extracted 
and classified. The two lines in the middle of the road are 
dashed white lane line (marked as dashed white line), the one 
in the left is a yellow solid lane line (marked as solid yellow 
line), and the one on the right side is a solid white line (marked 
as solid white line). Figure 11 is the bird eye view of the 
extracted road geometry. At this point, it should be stressed that, 
although there are other vehicles occlude the sight view to the 
right side lane line, it is still successfully extracted. This shows 
the robustness of the ARVEE against partly occlusion. 
a b 
Figure 10: ARVEE result in partly occlusion 
Figure 11: Bird eye view of result in Figure 10 
Figure 12 shows ARVEE result at a shadowed road. Despite of 
the tree shadows, all visible lane lines are correctly detected, 
linked and classified. This shows the robustness of the ARVEE 
against shadows. 
Figure 12: Road geometry result in shadows 
Figure 13 shows the detected road geometry overlapped on the 
digital map. The extracted road geometry fits the map perfectly, 
but with much more details and much higher accuracy. The 
result can greatly improve the current GIS database. 
8. EXPERIMENTS 
ARVEE has been tested over massive real mobile mapping 
survey data from VISAT™, including data from urban and rural 
519