CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
Figure 6: Ground truth used to evaluate the classification results; 
white pixels are buildings, blacks one are remaining objects 
the combination of Ah, Ap and NDVI has the best performance 
in different indexes. A detailed description of indexes is 1 : 
DR - Detection Rate: DR = TP /{TP + FN + UP) 
FPR - False Positive Rate: F PR = FP/ (TN + FP + UN) 
FNR - False Negative Rate: FNR = FN/(TP + FN + UP) 
UPR - Unclassified Positive Rate: UPR — UP/(TP + FN + 
UP) 
OA - Overall accuracy: OA = (TP + TN)/(TP + TN + 
FP + FN) 
R - Reliability: R = TP/(TP + FP) 
TUR - Total Unclassified Rate: TUR = (UP + UN)/(TP + 
TN + FP + FN + UP + UN) 
Classifier 
DR 
FPR 
FNR 
UPR 
nDSM 
94,49 
10,69 
5,51 
0,00 
AdaBoost 3F 
87,44 
1,33 
7,31 
5,25 
AdaBoost 5F 
91,17 
3,95 
7,08 
1,75 
AdaBoost 7F 
88,84 
1,57 
4,76 
6,40 
Classifier 
OA 
R 
TUR 
nDSM 
91,24 
83,95 
0,00 
AdaBoost 3F 
96,13 
97,50 
8,16 
AdaBoost 5F 
94,66 
93,18 
4,30 
AdaBoost 7F 
96,97 
97,10 
8,96 
Table 1: Results of pixel-based classification using different sets 
of features and metrics 
AdaBoost 3F, 5F and 7F differ for the set of features; 3F classifier 
uses Ah, Ap and NDVI, 5F adds Green and Blue; AdaBoost 7F 
classifies data using all features (excluding GNDVI). The Ada 
Boost 3F guarantees the best performance if compared with Ada 
Boost 5F/7F; adding more features other than Ah, Ap and NDVI, 
the classifier misclassifies data due lack of spectral separability 
(confirmed by ReliefF). All the classified data are also used for 
the road extraction; in particular the binary image obtained by 
considering land (bit set to one) and remaining classes (bit set 
zero) represents the input for roundabout and road extraction; the 
approach and results are presented in the following section. 
1 TP/FP = true/false positive TN/FN = true/false negative UP/UN = 
unclassified positive/negative 
4 ROAD EXTRACTION 
In this section we present preliminary results on road/roundabout 
extraction starting from classified data; the proposed approach 
works fine when the area is urban; modem cities often grows 
around main ancient perpendicular roads (cardus-decumanus). The 
key idea behind the algorithm is the “line growing”; more details 
about algorithm are discussed in next sub-sections. 
4.1 Filtering 
Filtering is a preliminary process before road extraction; this ac 
tivity is necessary for two main reasons: the first one is the pres 
ence of noisy classified data, because pixel-based classification 
suffers of noise; other approaches based on regions (object-based 
classification) can reduce it. The second problem that influences 
the quality of road extraction is the presence of trees/canopies; 
the chosen approach is a non-linear filter; if pixels that appertain 
to tree class have neighbours classified as “land”, then they are 
assigned to land class. The advantage of using this filter, is the 
reduction of effect produced by occlusions. In Fig. 7 the result of 
the filtering process is shown. 
Figure 7: Filtering. In the top image white pixels are classified as 
land; classification is noisy due to the presence of small objects 
as vehicles; in the bottom, the non-linear filter allows to reduce 
significantly the effect of noise and occlusions 
Non-linear filter consists of two steps: the first one is the reduc 
tion of noise using morphological operators. We applied three 
algorithms: opening to remove small objects, morphological re 
construction to retrieve boundaries and closing to fill small holes; 
the structuring element used was disk of size two. Second step is 
16