**Layer of the topologic knowledge**

While exploring the railway domain, lots of standard topological rules are imposed; such rules are used to help the driver and to ensure the passengers' security. From our point of view, the created rules are helpful also to verify and to guide the annotation process. In fact, topological knowledge represents adjacency relationships between scene elements. For instance, and in case of the Deutsche Bahn scene, the distance between the distant signal and the main one corresponds to the stopping distance that the trains require. The stopping distance shall be set on specific route and is in the main lines often 1000 m or in a rare case 700 m. Add to that, three to five Vorsignalbake are distant from 75m while then the last one is distant 100m from the distant signal, Fig.11.

At semantic view, topological properties describe adjacency relations between classes. For example, the property Topo:isParallelTo allows characterizing two geometric concepts

From Unstructured 3D Point Clouds to Structured Knowledge - A Semantics Approach 235

gemetry

Geometry

Geometry containing Perpendicular elements

> Geometry containing Parallel elements

The specialized classes of the Alg:Algorithm axiom are representing all the algorithms developed in the 3D processing layer. They are related to several properties which they are able to detect. These properties (Geometric and semantic) are shared with the DC:DomainConcept and the Geom:Geometry classes: By this way, a sequence of

SubPointCloud Line\_3D 3D Lines Segmentationin2D)

None

LinesDetectionin3DbyRANSAC

LinesDetectionin3DbyRANSAC

gemetry VerticalObjectsDetection

gemetry Segmentationin2D

height Segmentationin2D

front face Segmentationin2D

**Algorithm name has Input hasOutput isDesignedfor hasSuccessor** 

PointCloud SubPointCloud Vertical

**Detection** PointCloud Point\_2D Vertical

**BoundingBox** SubPointCloud Point\_3D Vertical

**FrontFaceDetection** SubPointCloud Boolean Geometry with

Point\_2D

**ApproximateHeight** SubPointCloud number

**CheckPerpendicular** Line\_3D Boolean

**CheckParallel** Line\_3D Boolean

Table 3. 3D processing algorithms and experts observations

algorithms can detect all the characteristics of an element.

Fig. 12. Hierarchical structure of the Algorithm class.

**Vertical Objects** 

**Segmentationin2D**

**RANSAC Line Detection** 

Fig. 10. Topological rules.

by the feature of parallelism. Similarly relations like Topo:isPerpendicularTo and Topo:isConnectedTo will help to characterize and exploit certain spatial relations and make them accessible to reasoning steps.

#### **5.3.2 Layer of 3D processing knowledge**

The 3D processing algorithmic layer contains all relevant aspects related to the 3D processing algorithms. It contains algorithm definitions, properties, and geometries related to each defined algorithms. An importance achievement is the detection and the identification of objects, which has a linear structure such as signal, indicator column, and electric pole, etc., through utilizing their geometric properties. Since the information in point cloud data sometimes is unclear and insufficient, the various methods to RANSAC (Tarsha-Kurdi, et al., 2007) are combined and upgraded. This combination is able to robustly detect the best fitting lines in 3D point clouds for example. Fig11 presents the Mast object constructed by linear elements, ambiguously represented in point cloud as blue points. Green lines are results of possible fitting lines and clearly show the shape of the object that is defined in the ontology. The object generated from this part is a bounding box that includes all inside geometries of the object and a concept label.

Fig. 11. Mast detection.

Next to the 3D expert recommendation, knowledge within the Table 3 is created linking a set of 3D processing algorithms to the target detected geometry; the input and output.

234 Semantics – Advances in Theories and Mathematical Models

by the feature of parallelism. Similarly relations like Topo:isPerpendicularTo and Topo:isConnectedTo will help to characterize and exploit certain spatial relations and

The 3D processing algorithmic layer contains all relevant aspects related to the 3D processing algorithms. It contains algorithm definitions, properties, and geometries related to each defined algorithms. An importance achievement is the detection and the identification of objects, which has a linear structure such as signal, indicator column, and electric pole, etc., through utilizing their geometric properties. Since the information in point cloud data sometimes is unclear and insufficient, the various methods to RANSAC (Tarsha-Kurdi, et al., 2007) are combined and upgraded. This combination is able to robustly detect the best fitting lines in 3D point clouds for example. Fig11 presents the Mast object constructed by linear elements, ambiguously represented in point cloud as blue points. Green lines are results of possible fitting lines and clearly show the shape of the object that is defined in the ontology. The object generated from this part is a bounding box that includes

Next to the 3D expert recommendation, knowledge within the Table 3 is created linking a set of 3D processing algorithms to the target detected geometry; the input and output.

Fig. 10. Topological rules.

Fig. 11. Mast detection.

make them accessible to reasoning steps.

**5.3.2 Layer of 3D processing knowledge** 

all inside geometries of the object and a concept label.


Table 3. 3D processing algorithms and experts observations

The specialized classes of the Alg:Algorithm axiom are representing all the algorithms developed in the 3D processing layer. They are related to several properties which they are able to detect. These properties (Geometric and semantic) are shared with the DC:DomainConcept and the Geom:Geometry classes: By this way, a sequence of algorithms can detect all the characteristics of an element.

Fig. 12. Hierarchical structure of the Algorithm class.

From Unstructured 3D Point Clouds to Structured Knowledge - A Semantics Approach 237

the detection process will result bounding boxes, representing a rough position and orientation of the detected object. Table 4 shows the mapping between the 3D processing built-ins, which is computer and translated to predicate, and the corresponding class.

The layer of the topological knowledge represents topological relationships between scene elements since the object properties are also used to link an object to others by a topological relation. For instance, a topological relation between a distant signal and a main one can be defined, as both have to be distant from one kilometer. The qualification of topological relations in to the semantic framework is done by new topological Built-Ins. This step aims at verifying certain topology properties between detected geometries. Thus, 3D\_Topologic built-ins have been added in order to extend the SWRL language. Topological rules are used to define constrains between different elements. After parsing the topological built-ins and its execution, the result is used to enrich the ontology with relationships between individuals that verify the rules. Similarly to the 3D processing built-ins, our engine

Upper 3D\_swrlb\_Topology:Upper(?x, ?y) Upper(?x,?y) Transitive

Intersect 3D\_swrlb\_Topology:Intersect(?x, ?y) Intersect(?x,?y) Symmetric

Actually, the created knowledge base aims to satisfy two basic purposes, which are, guiding the processing algorithm sequence creation based on the target object characteristics, and facilitate the semantic annotation of the different detected objects inside the target scene. Let's remember that one of the main ideas behind our suggestions is to direct, adapt and select the most suitable algorithms based on the object's characteristics. In fact, one algorithm could not detect and recognize different existent objects in the 3D point clouds, since they are distinguished by different shapes, size and capture condition. The role of knowledge is to provide not only the object's characteristics (shape, size, color...) but also

(?x, ?y,?d) Distance(?x, ?y, ?d) Symmetric

translates the rules with topological built-ins to standard rules, Table 5.

**Function** Correspondent topologic Built-Ins Correspondent object

*3D\_swrlb\_Processing: VerticalElementDetection (?Vert,?Dir)* 

*3D\_swrlb\_Processing: HorizentalElementDetection (?Vert,?Dir)* 

Table 4. 3D processing Built-Ins mapping.

**6.2 Integration of topological operations** 

Distance 3D\_swrlb\_Topology: Distance

Table 5. Example of topological built-ins.

**6.3 Guiding 3D processing algorithms** 

**3D Processing Built-Ins** Correspondent Simple class

*Geom:Vertical\_BoundingBox(?x)* 

*Geom:Horizental\_BoundingBox(?y)* 

property Characteristics

The following section presents in details the semantic integration process undertaken in the WiDOP solution to detect and annotate semantically the eventual semantic elements.
