McGlone - 9 
could be in practice, due to the limited accuracy of recovering the parameters and 
the limited dimensionality of the description space. 
4.4. MAPS lab site model format 
The maps lab site model format is designed for the representation of cartographic 
objects, in a topologically-structured boundary representation. The site modeling 
package is used within a number of cartographic feature extraction systems within 
the laboratory [McKeown et a/., 1994]. Some support is included for volumetric 
primitives, for objects such as spheres and cylinders. The fundamental object is the 
point, which is represented in local coordinates with attached covariance informa 
tion. Edges are defined as chains of points, surfaces are defined as chains of edges, 
and models are defined as a set of linked surfaces. 
A problem with boundary representations is that they are almost too general, i.e., a 
cube has the same boundary representation as any rectangular prism, but sometimes 
we specifically want to represent a cube. We address this by allowing the specifi 
cation of constraints between points, edges, surfaces, and objects, thereby fixing 
the geometry of the object as precisely as we require. The constraints between ob 
jects allow us to emulate CSG by combining primitive shapes to form more complex 
objects. 
In order to deal with objects at a higher level we define a set of models, including 
flat-roof building, peaked-roof building, and rectilinear building, by predefining the 
topology and constraints in a standard way and furnishing a corresponding set of 
access and manipulation functions. The construction and utilization of common 
building models is therefore simplified. These predefined models can be tied to 
each other using constraints, for instance, to model an 1-shaped building as the 
combination of two simple rectangular buildings. 
Our site model format has proven to be effective for our applications, including 
generation using edge geometry based techniques and rendering for simulation ap 
plications. The model has been extended as necessary; we have added a triangulated 
irregular network (TIN) terrain representation using point and edge objects. The 
inclusion of constraints and covariance information makes information explicit for 
translation into parametric models or manipulation as higher level objects. As with 
all boundary or surface representations, the computation of volumetric properties is 
complicated. 
4.5. TraX—representation space 
Common statements in papers on object modeling (including this one) are that 
the object model must be task-specific, and that the optimum type of model is 
determined by the types of objects to be represented, this presents a dilemma for 
general-purpose systems, Since many types of objects may need to be represented, 
at a variety of resolutions. 
The TraX system, developed at SRI [Bobick and Bolles, 1992] for use in autonomous 
vehicle navigation, attempts to address this problem by defining a “representation