Chen Yuen Teoh
2 RECOIL FRAMEWORK
RECoiL is developed based on Road Recognition from Aerial Images using Inductive Learning, RAIL, by Sowmya and
Singh (14, 15). RAIL uses supervised multi-level learning to derive rules based on arduous example selection, while
RECoiL bypasses this problem by implementing a clustering module, which is useful at levels where comprehensible
recognition rules are not required.
The road detection process for RECoiL is conceptualised into four levels: road segment detection, linking road segments,
junction detection and linking roads to junctions.
At each separate level, different features are targeted. Examples are pixel-based ones at lower levels and, relations between
pairs of edges at higher levels. The recognition rules learned at one level are applied at that level, and the resulting image
becomes the input to the next level, and eventually road-liked objects are recognised from the images. The structured
diagram of RAIL is shown in figure 1. C4.5 is the learning program used for learning rules.
| | | Vista Canny |
| | Original Image {> Gradient Image
| |
uw |
| Operator
Edge Image
| Edges | =
ol D
| Curve | Image with | |
Ed iid
rrp
ip etre re
| x un i
1 —€—7 f | Clustering Level 1 f
|| | GUI Example | Applyleamed rules |-1— — vum
Selection "| Generate Road Segment | | Level | 7 iSi =
—————— be 3] —| Clustering Level 2
— — re L——À-| Clustering Level 3
I — rs oT
| GUI Example | | Apply learned rules |
| Selection [ ^| Generate Linked Road Segment
GUI Example | | Apply learned rules |
Selection | | Generate Junction |
| Fusing output of level 2 and level 3
s
Road recognised and extracted
Figure 1: Proposed System
3 REVIEW OF RECOIL
The operations of RECoiL can generally divided into three steps for each level of RECoiL (17). For supervised learning,
the three steps are examples selection using custom built GUI system, rules learning using C4.5 and applying rules to
input image and generating output image. Using the clustering approach, the three steps are generation of cluster data,
applying K-means algorithm to cluster the generated data and finally display/select/saving the required cluster.
As mentioned before, RECoiL is based on a hierarchical multilevel framework, which ranges from level 1 to level 4. Prior
to level 1 is a preprocessing level. At this level, the primary aim is to extract an edge image from the remote sensing raster
image. There are three separate steps involved, which are edge tracking, edge linking and curve splitting.
Level 0 is an abstract and optional level for RECoiL. It resembles the clustering module at level 1, and the output of level
0 will be used as input for level 1. The targeted features at level O are similar to the features used in level 1 described in
figure 2 (a).
Level 1 aims to extract road pairs which are in the form of pairs of edges. The targeted features at this level are described
in figure 2 (a). Another special attribute is also used for extracting road pairs, which is spatial overlapping that is required
for edges to form a road pair. Meanwhile, level 2 aims to link the pairs of road segments extracted at the previous level to
form linked road segments, without any connectivity with junctions based edges. Again, targeted features at this level are
described in figure 2 (b).
Level 3 of RECoiL is mainly focused on junction based edges extraction and will be described in detail in section 4. As
for level 4 of RECoiL, it is still under development and will focus on fusion of the outputs from level 2 and level 3 of
RECoiL to form the road-liked edge image.
Due to the implementation of hierarchical multilevel framework, RECoiL allows combination of different modules at
different levels. Even though the possible combinations are plentiful, only a few important and useful combinations are
described in Table 1.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 883
