B3, 2012 
er mainly deals 
easons such as 
e ellipse cover 
in many cases 
ur method is to 
ible ellipse after 
is extracted for 
) detect arcs and 
id the minimum 
non-maximum 
“hysteresis High 
hold as arc. All 
It. 
0 covers for the 
used to fine the 
ipe and size of 
'ameters as long 
vas solved with 
according to the 
are then used to 
Jecause only the 
g axle is likely 
:d to 45 degree. 
tween 0.5m and 
n the interesting 
s. The LIDAR 
se false arcs by 
r is 0.2m higher 
«cluded as false 
ature for cover 
ching method is 
| to the deferent 
ination etc, it is 
h all candidates 
ing for adjacent 
detection results 
; combined with 
enough texture 
ithin the ellipse 
yme covers have 
(see Fig. 3d). In 
€ to the car and 
se covers. The 
2 deducted from 
n eq.1, xy and X 
ge and ground 
rotation matrix 
by GPS/IMU 
ent observations 
LIDAR data, À 
| eq.1 while Z 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
X 2 AR, ^ T, a) 
x, 2 RI(X - T) 
X=Rx +7 (2) 
x, 2 RI (ARx, * T, - T) 
According to eq.2, all the matching images can be resampled to 
the reference image. Then the least square image matching is 
done between the patch of candidate covers in reference image 
and corresponding area in matching images. 
Taking into account that there are some occlusions in covers 
and only a part of cover can be detected, the matching method 
deals with two situations. Firstly, the whole ellipse is matched 
and if the matching successfully, this step ends and if not, the 
part of ellipse according to the arc been detected by EDISON is 
then matched. 
After matching, the floating objects such as wheels and floating 
shadows with round shape will be excluded. 
Then the forward intersection of corresponding image points is 
done to get ground location with EOs supplied by GPS/IMU. 
     
  
C 
     
Figure 3. Variety of manhole covers 
Figure 4. Edge detection 
b c 
Due to the high accuracy of EOs, the matching result will be 
regard as failure if the REMS of intersection is more than 0.2m. 
In most cases, we consider two images for one single cover. The 
front cameras take two images with nearest cover presentation 
of 2m interval in object space. The farer images are not used for 
relatively small and vague texture (see Fig. 5d). If the two 
images are correctly matched, we consider that's a successful 
cover detection and location. 
3. EXPERIMENTS 
3.1 Test data 
The test data include 1100 images and 26 distinct manhole 
covers. In this test the manhole covers exhibit different textures, 
sizes, and different view points, especially the influence by 
shadows and occlusions. The positioning accuracy of GPS data 
is about 0.05m and the rotate angle accuracy of IMU is more 
than 30 . The locating accuracy of LIDAR is more than 0.05m 
within 60m. 
  
d 
Figure 5. Matching and locating 
3.2 Edge detection evaluation 
After the interesting area is extracted the EDISON edge 
extraction is done to find edges which may contain covers, see 
figure 4a, 4b. Then the ellipses are fitted for each independent 
edge. If more than 60% points in edge segment is fitted to an 
ellipse with REMS less than 2 pixels, we consider it a 
successful fitting. In fig. 4c, there are 3 edges can fitted to an 
ellipse among 7 segments. Obviously, there are two fitted 
ellipses are not belong to covers but belong to traffic lines. In 
the next step, the area and direction constraints are used to 
exclude those two ellipses, and in fig. 4d only the right ellipse 
corresponding to the cover remains. 
The red line in fig. 6a shows the missing rate via false positives 
(FP) rate only use edge information. With the FP increases the 
   
missing rate decreases but slowly. In the FP rate of 0.05 the 
detection rate is 79%. It should be noted that the missing covers 
which cannot be extracted by EDISON also been included in 
the statistics. 
3.3 Matching and location evaluation 
The adjacent images with candidates are then matched for 
further confirm and for locating. The part and whole ellipse are 
separately matched according to the arcs been extracted. Firstly 
the whole ellipse is matched only in two adjacent images for the 
third images about 6m away lose most of the texture (see Fig. 
5c). If the matching success, part arcs will not be matched. The 
fig. 5d shows the situation that the whole ellipse can not match 
but the part in the arc (white line) can match. The Fig. 5e shows