3. Istanbul 2004
calculated in [5|
l hypotheses as
acter of stopped
es with vehicle
E
Cast Shadow
rrection
segmentation
w-Corrected&
isked Nadir image
tion of Noise |
T segmentation|
Y = =
chicle region
emoval
Y
ons Likely
'ehicle Parts
he Street
Ÿ
ular Polygons |
tted by |
Likely Regions |
p
eight of |
ular Polygons
with |
ig correlation
EIE
ikely Regions
iar fitted polygon
1 height
ration
TON
Our algorithms
| supplementary
topped vehicles
to discriminate
| from stopped
se of omission
; to additionally
tomatically and
chicle detection
] algorithm of
nage processing.
st vehicle-likely
it-and-matching
ed by selecting
ained from pre-
€ independent.
vities of 'those
pair distances,
ry, hierarchical
um.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Vehicle Candidate Generation
- Select Vehicle-Likely Regions agreed with
Thresholding of Height-and-Matching-Correction
po» - Vehicle Candidate Grouping
Hierarchical Cluster
- Fitting Rectangular Polygon
enclosing Vehicle Candidate
i
i
Removal of committed Road Surface |
as Vehicle Candidate |
- Analysis of Matching Correlation Curve!
Vehicle Candidate Validation |
- Agreement with Stopped Vehicle Model
No E “Agreed with x
>< Vehicle Model
i Yes
| Detected Stopped Vehicle
at first order expansion
T
2nd order > . Yes Detected Stopped Truck Candidate
. Expandability of‘ »——— —» Generation
“Detected Vehicle” - Second Expansion
Validation of Stopped Truck
| - Agreement with Stopped Vehicle Model
+
^^ Agreed with
Vehicle Model,
Ra T Yes
[Detected Stopped Vehicle =
Vehicle candidate are generated by detecting natural nearest-
No
Figure 4 Systematic diagram of stopped vehicle detection
Figure 5 exampies of Detected Stopped Vehicles
region groupings in the hierarchical tree. Vehicle candidate
regions are fitted with rectangular polygon by our rectangular
polygon fitting algorithm and height of this rectangular polygon
is calculated by area-based stereo matching algorithm. Vehicle
candidates with their rectangular polygon and height, which are
agreed with stopped vehicle model, are detected stopped
vehicles.
Moving Vehicle Detection is our proposed algorithm of
moving vehicle detection by using multi-TLS image processing.
Firstly, vehicle likely regions of stopped vehicles and their
neighbourhood regions along the road direction are removed by
using neighbourhood relation network with road-direction
constraint. Secondly, regarding non-vehicle likely region such
as road surface regions at pre-processing stage, isolated vehicle-
like regions surrounded by road surface regions without any
neighbourhood vehicle-likely regions, which are not agree with
vehicle width and orientation thresholding, are removed.
Seeding Point Detection
- Removal of Stopped Vehicle Regions and their
neighborhood regions along the road direction
- From preprocessing, Removal of Isolated vehicle-
likely regions surrounding by road surface not agreed |
with vehicle size |
4
| Rest of regions as seeding points
i from Removal stage above
!
Vehicle Candidate Generation
- Expanding along the road lane direction and
merging seeding point with regions found
by ‘expansion proceed
- Fitting rectangular polygon enclosing
Vehicle Candidate
| Vehicle Candidate Verification
- Using moving vehicle model
|
3 No |. 7 Agreed with : ;
Moving Vehicle Model
| Yes
Redundant result of Detected Moving Vehicle
4
Remove redundant results and Polygon Storage
- Vehicle polygon with Maximum Edge pixels
|
| Detected Moving Vehicle
Figure 6 Algorithm of Moving Vehicle Detection
From two processing stages, the rest of vehicle-likely regions
are the seeding points for the ‘Expansion Proceed’ algorithm of
vehicle candidate generation at third stage. For the description
of ‘Expansion Proceed’ algorithm, a selected vehicle-likely
region as seeding point expands along road to detect
neighbourhood vehicle likely regions between both sides of
seeding point along road direction and then merge detected
regions with seeding point to be generate one cluster or moving
