Ref. e = A à DS | Scheel um ale es BS 
Road 10.59 | 0.06 | 0.47 | 0.19 | 0.39 | 0.08 | 0.75 | 2.49 | - [001] — [0121021| — [68.73 
Isl. V 010 loo | - Shr ai. | 00L ied we] ei wn] we |.— Le- [45 
Sidew. 1.28 - [0335| 0.13 | 0.32 | 0.05 | 021|069| —- | - | - [009|004|] — [1100 
Build. 0.38 — | 0.04 | 13.02 | 0.48 | 0.06 | 0.18 | 130| — | — | — | 0.16 | 0.18 | 023 [81.17 
Grass 0.30 — | 0.14 | 0.57 |i448| 1.03 | 0.16 | 0.94 | 0.05 | 0.40 | 0.06 | 4.61 | 0.07 | — [63.51 
Agr. 0.04 — | 0.17 | 0.25 | 0.74 | 7,03 0.85 | 0.25 | 0.04 | 0.01 | 3.18 | 0.05 | — |s5.70 
Water 0.08 = - oo] 09] - F5] oo) - | - | - | - 1002] —. [4947 
Sealed 3.67 0.52 | 0.81 | 1.08 | 0.28 | 0.40 | 3.19] — | - | 001 | 033 | 0.23 | 0.01 [3022 
Isl. A 0.01 = - mm VOL] wnt | wo ct BA PL] = EE 
Beach 0.01 - E TT -|-Tz 
Railw. 0.05 - Joe ['- ped lees] Lo] - | — 133: /004/ - | — | 264 
Tree 0.21 — | oo: | 012 | 2,57 | 0.14 | 001 | 010] — | - | - H9] oo | — [8180 
Car 0.05 =. | 002 | 002 | 003 [001 | - loi] - ] - | - [oo E933] — [49,6 
Bridge 0.08 > - dee] >|5—-| — 1006| - | — | - | — [ou at + 
Corr. 62.84 | 6.74 | 19.10 | 85.45 | 71.54 | 80.51 | 11.07 /31.98| — | — |10.60|62.59|28.14| — 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Table 1: Confusion matrix for the experiment using 14 classes and the car confidence feature. All values are given in [%]. Overall 
accuracy: 63.50%. Abbreviations: Ref.: Reference; Isl. V: traffic island with vegetation; Sidew.: sidewalk; Build.: building; 
Agr.: agricultural; Isl. A: traffic island with asphalt; Railw.: railway; Comp. / Corr.: Completeness / Correctness. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Class =: » i 5 3 s 5 3 a = z ; v à 
Ref. Z£izlelsls|sis|izis|s|sz|c '$l81l35 
Road 10.25 | 0.08 | 0.50 | 0.18 | 0.35 | 0.12 | 0.78 | 2.56 - 0.01 - 0.12 0.43 | 0.02 | 66.56 
Isl. V 0.10 | 0.01 -- - - = 710.01 | - = E = = = — 857 
Sidew. 1.22 - 0.39 10.13 | 0.30 | 0.08 | 0.21 | 0.65 - - - 0.10 0.10 == 112.22 
Build. 0.23 - 0.02 [13.18] 0.43 | 0.09 | 0.17 | 1.20 - - - 0.17 0.35 | 0.20 | 82.18 
Grass 0.14 -- 0.12 | 0.56 |14.06| 1.41 | 0.15 | 0.90 | 0.04 0.40 0.04 4.84 0.13 - 161.68 
Agr. 0.03 -- 0.15 | 0.25 | 0.68 | 7.21 0.84 | 0.16 - 0.01 3.19 0.09 - | 57.11 
Water 0.07 -- - 0.04 | 0.02 - 0.21 | 0.05 - = - - 0.03 == | 49.39 
Sealed 3.47 0.01 | 0.55 | 0.80 | 0.97 | 0.43 | 0.40 | 3.10 - - 0.01 0.34 0.46 | 0.01 | 29.36 
Isl. A 0.01 - -- = 0.01 - - - 0.00 - - - - -- - 
Beach = - - - - - - 1010 | - 0.00 | 0.01 - - - - 
Railw. 0.04 - 0.09 - 0.09 | 0.07 - 0.04 - - 0.01 0.04 - - 2.80 
Tree 0.01 - 0.02 | 0.12 | 2.38 | 0.26 | 0.01 | 0.09 - - - 14.57 | 0.02 - 183.43 
Car 0.03 -- 0.02 | 0.02 | 0.03 | 0.02 - 0.19 - - - 0.01 0.33 - 150.02 
Bridge 0.08 -- - 0.09 - - - 0.03 - — - - 0.03 | 0.00 - 
Corr. 65.37 | 6.92 | 20.85 | 85.78 | 72.75 | 74.45 | 11.02 | 31.80 | — - 13.33 | 62.29 |16.75| -- 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Table 2: Confusion matrix for the experiment using 14 classes without the car confidence feature. All values are given in [%]. 
Overall accuracy: 63.32%. Abbreviations: see caption of Table 1. 
5. CONCLUSION future we want to improve our method by integrating more 
expressive features, e.g. HOG features or features related to car 
In this paper, a method for the classification of crossroads using 
MRF was proposed. It considered 3D information in the form of 
a DSM generated from multiple overlapping aerial images, as 
well as a car confidence feature to avoid problems with 
occlusions of the road surface by cars. Distinguishing 14 classes 
relevant in the context of crossroads, an overall accuracy of 
about 63.5% could be achieved. The main error sources were 
the confusion of object classes that are only distinguished by 
their relative alignment, but not by their appearance in the data 
(road, sidewalk, sealed) and errors in the DSM generation 
process. After merging the classes that are most similar in 
appearance, the overall accuracy was increased to 74.8%. In the 
trajectories, and we will also integrate multi-scale features. 
Furthermore, we want to build a more sophisticated model of 
context based on Conditional Random Fields (Kumar & Hebert, 
2006), also using improved models for the association 
potentials linking the class labels to the data. Finally, we will 
investigate whether the results can be improved by using image 
segments as the nodes of the graphical model. 
ACKNOWLEDGEMENT 
This research was funded by the German Science Foundation 
    
     
    
   
   
   
    
   
   
   
   
   
    
     
     
  
  
  
  
  
  
   
   
   
   
   
    
   
    
   
   
   
    
    
     
   
   
   
    
   
   
   
   
   
   
  
  
   
    
  
  
  
  
  
  
  
  
  
  
  
  
Build 
Gras: 
Agr. 
Beact 
Tree 
car 
| Bridg 
  
  
Corr. 
Table 
  
  
Figure 
Barsi, 
detecti 
XXXI 
| Birchf 
| that is 
401-4(