Characterization of Models

  Most approaches do not use knowledge-based representation formalisms (cf. Table 1). The only exception is [6], which uses a semantic network. The approaches are based mainly on geometry, even though by utilizing profiles a radiometric component is added. The kind of representation is exclusively 2D and generic with the exception of [25,3]. 3D-Objects like bridges are not considered at all by most approaches, because their percentage of the road network is usually small. Overall, there is a trend from simple topology-building to more complex network construction, for example, by active closing of gaps [17]. Since roads are at least locally planar and continuous with respect to their height, many approaches do not use a sensor model for road extraction. Nevertheless, most of the approaches are based on orthophotos, which allows for a direct link to object coordinates. The scene and object models of roads are relatively simple (cf. Table 2). McKeown and Denlinger's approach [19] is based on a local disturbance model of so-called ``road features'' like cars, occlusions, etc. In [24,25], intersections as well as cars on the road, in [3] intersections and markings, and in [6] only markings are modeled. Network construction, which can be seen as a part of the scene model, is only used by relatively new approaches like [25,29]. The approach described in [25] extends the network construction with the modeling of other objects like shadows, cars or trees, while in [3,17] bridging shadows by means of snakes is used. The function of objects is only used in [6,25,29] by means of network construction. The latter is, more or less, implicit in [6], and more explicit in [25,29].

In summary, there is an evolution from general techniques to approaches customized for the object type: It is considered to be very important and feasible to utilize the knowledge, i.e., the models and strategies, of an application. For roads, 3D-information (bridges, tunnels, mountainous roads) is seldom used. Because the network character of the road forces the model to consider larger areas, a global treatment is of interest.

 

Table 1: Characterization of the Models for Road Extraction - I

Approach	Representation Formalisms	Geometry and Radiometry	Kind of Representation	Sensor Model
Roads as Lines in Low-Resolution Imagery: Fischler et al. 1981[8]	implicit	geometric	2D; generic (but no explicit network construction)	no
Road-Grouping and Network Creation: Vasudevan et al. 1988 [36]	implicit	geometric	2D; generic (simple topology)	no
Road Detection and Tracking: Zlotnick and Carnine 1993 [40]	implicit	geometric	2D; generic (parallelism, continuity and smoothness)	no
Road Detection and Tracking: McKeown and Denlinger 1988 [19]	implicit	geo- + radiometric	2D; generic (profile)	no
Road Network Construction by Interpreting the Local Context: Ruskoné 1996 [24,26,25]	implicit	geo- + radiometric	2D and 3D; generic (network construction)	no
Multi-Resolution, Semantic Objects, and Context for Road Extraction: Baumgartner et al. 1997 [3,29,17]	implicit	geo- + radiometric	2D and 3D; generic (network construction)	no
Utilization of Maps and GIS: de Gunst 1996 [6]	semantic network	geo- + radiometric	2D; generic + parametric (profiles, markings)	no

 

Table 2: Characterization of the Models for Road Extraction - II

Approach	Object Model	Scene Model	Function
Roads as Lines in Low-Resolution Imagery: Fischler et al. 1981[8]	no	weak (minimal spanning tree)	no
Road-Grouping and Network Creation: Vasudevan et al. 1988 [36]	no	medium (network construction)	no
Road Detection and Tracking: Zlotnick and Carnine 1993 [40]	no	no	no
Road Detection and Tracking: McKeown and Denlinger 1988 [19]	medium (disturbance model)	no	no
Road Network Construction by Interpreting the Local Context: Ruskoné 1996 [24,26,25]	medium (crossing and cars)	medium (network construction and shadow, tree, field)	little; implicit (network)
Multi-Resolution, Semantic Objects, and Context for Road Extraction: Baumgartner et al. 1997 [3,29,17]	medium - high (crossing, markings)	medium - high (network construction, shadow bridging)	little; implicit (network)
Utilization of Maps and GIS: De Gunst 1996 [6]	medium (road marking)	medium (network construction)	medium, implicit (network, max. climb, curvature)