Research Article  Open Access
AhLian Kor, Brandon Bennett, "A Hybrid Reasoning Model for “Whole and Part” Cardinal Direction Relations", Advances in Artificial Intelligence, vol. 2013, Article ID 205261, 20 pages, 2013. https://doi.org/10.1155/2013/205261
A Hybrid Reasoning Model for “Whole and Part” Cardinal Direction Relations
Abstract
We have shown how the nine tiles in the projectionbased model for cardinal directions can be partitioned into sets based on horizontal and vertical constraints (called Horizontal and Vertical Constraints Model) in our previous papers (Kor and Bennett, 2003 and 2010). In order to come up with an expressive hybrid model for direction relations between twodimensional singlepiece regions (without holes), we integrate the wellknown RCC8 model with the abovementioned model. From this expressive hybrid model, we derive 8 basic binary relations and 13 feasible as well as jointly exhaustive relations for the x and ydirections, respectively. Based on these basic binary relations, we derive two separate composition tables for both the expressive and weak direction relations. We introduce a formula that can be used for the computation of the composition of expressive and weak direction relations between “whole or part” regions. Lastly, we also show how the expressive hybrid model can be used to make several existential inferences that are not possible for existing models.
1. Introduction
Relative positions of regions in largescale spaces, and particularly in the geographic domain, are often described by relations referring to cardinal directions. These relations specify the direction from one region to another in terms of the familiar compass bearings: north, south, east, and west. The intermediate directions northwest, northeast, southwest, and southeast are also often used. Some models for reasoning with cardinal directions are the coneshaped [1, 2], projectionbased models (ibid), and direction matrix [3–5].
Papadias and Theodoridis [6] describe topological and direction relations between regions using their minimum bounding rectangles (MBRs). However, the language used is not expressive enough to describe direction relations. Additionally, the MBR technique yields erroneous outcome when involving regions that are not rectangular in shape [4] Some work has been done on hybrid direction models. Escrig and Toledo [7] and Clementini et al. [8] integrated qualitative orientation and distance to obtain positional information. Isli [9] combined Frank’s [1, 2] cardinal direction relations model and Freksa’s [10] orientation model to facilitate a more expressive reasoning mechanism. Sharma and Flewelling [11] infer spatial relations from integrated topological and cardinal direction relations. Liu and colleagues [12] have developed reasoning algorithms which combine RCC8 [13] for topological relations (discussed in Section 4) and the cardinal direction calculus (CDC [3–5], discussed in Section 2) for direction relations. Li and colleagues’ work [14, 15] focuses on the development and evaluation of an efficient reasoning mechanism for RCC8 and RA (Rectangle Algebra, and further explanation can be found in [16, 17]) which is employed to solve the satisfiability problem of these two joint constraint networks.
Typically, composition tables are used to infer spatial relations between objects. They have been employed to make different inferences about cardinal directions relations [3, 19–24]. One of the advantages of composition tables is that they can lead to tractable computation of inferences [25]. In this paper, we have developed an expressive hybrid model for direction relations. We will describe the binary relations in the model and define “whole and part” relations. Based on this model, we derive two composition tables for expressive and weak direction relations. This is followed by introducing a formula which could be used to compute both expressive and weak direction relations for “whole and part” regions. Finally, we will demonstrate how the model could be used to make several types of existential inferences.
2. Cardinal Direction Models
Frank [1, 2] defines cardinal directions as cones which are related to the angular direction between an observer’s position (in the form of a point) and a destination point. The coneshaped cardinal direction model could have 4, 8, or more partitions (look at Figure 1).
Frank defines the four major cardinal directions (north, south, east, and west) as pairwise opposites and half planes. When the two sets of half planes are combined, it yields four intermediate cardinal directions (northeast, northwest, southeast, and southwest) which are depicted in Figure 2. Ligozat [21] applies the model to points in a twodimensional space. Thus, the referent object, Point B, will be given the four major directions. However, the relations between two objects will be denoted by one of the following basic relations: N, S, E, W, NE, NW, SE, SW, or EQ.
Frank [1, 2] extends the halfplanes to tiles for regions (as shown in Figure 3). In this projectionbased model, the plane of an arbitrary singlepiece region a is partitioned into nine tiles, NorthWest, NW(a); North, N(a); NorthEast, NE(a); SouthWest, SW(a); South, S(a); SouthEast, SE(a); West, W(a); Neutral Zone, O(a); East, E(a). According to Frank, the O tile is considered a neutral zone, because in this tile, the relative cardinal direction between two regions cannot be determined due to their proximity.
Frank compares and contrasts reasoning with the coneshaped and the projectionbased models for cardinal directions. The reasoning capability for both the systems is limited and weak though they do not differ substantially in their reasoning outcomes. In order to create a more expressive reasoning model, Isli [26] integrates the Frank’s coneshaped and projectionbased models to facilitate reasoning about relative position of points of the 2dimensional space. This hybrid model is well suited for applications of largescale highlevel vision, such as, for example, satellitelike surveillance of a geographic area.
The cardinal direction calculus (CDC) [3–5] is a very expressive qualitative calculus for directional information of extended objects. A direction relation matrix (DRM) in (1) is used to represent direction relations between connected plane regions. Liu and colleagues [27, 28] have shown that consistency checking of complete networks of basic CDC constraints is tractable, while reasoning with the CDC in general is NP hard. However, if some constraints are unspecified, then consistency checking of incomplete networks of basic CDC constraints is intractable.
The cardinal direction of a target object (region b) to a referent object (region a) as shown in Figure 4 is described by recording those tiles covered by the target object. According to Goyal and Egenhofer [4], a matrix is employed to register the intersections between the target object and the tiles of the referent object (see (1)). The elements in the directionrelation matrix correspond to the tiles of the referent object, region a (in Figure 4).
In (1), the symbol represents empty tile while represents nonempty tile. These are used to describe cardinal directions at a coarse granularity level. In Figure 4, region occupies the N, NW, and E tiles of region . Thus, these three tiles are considered nonempty while the rest are considered empty (as shown in (1)).
Goyal and Egenhofer [5] extend the direction relation matrix, so that it will be more expressive. Instead of using the empty and nonempty notations, it registers how much (in terms of proportion) the target region occupies each tile (see (2)). The expressive direction relation matrix in (2) has 6 elements of zero and three nonzero elements which sum up to 1.0. If the matrix has only one nonzero element then it is known as a single element direction relation matrix while a matrix with more than one nonzero element is called a multielement direction relation matrix (ibid).
Coarse direction relation matrix [4]:
Expressive direction relation matrix [5]:
3. Horizontal and Vertical Constraints Model
Every region has a minimal bounding box with specific minimum and maximum (and ) values. The boundaries of the minimal bounding box of a region are depicted in Figure 5. The set of boundaries of the minimal bounding box for region could be represented as , , , , and these values will be employed to define each tile.
The definition of the nine tiles in terms of the boundaries of the minimal bounding box is listed as below. Note, in this paper, all the tiles are regarded as mutually exclusive. Thus neighboring tiles cannot share common boundaries:(i),(ii),(iii),(iv),(v),(vi),(vii),(viii),(ix).
In our previous papers [29, 30], we have shown how to partition the nine tiles (in Figure 5) into sets based on horizontal and vertical constraints called the Horizontal and Vertical Constraints Model. However, in this paper, we shall rename the sets for easy comprehension purposes. The following are the definitions of the partitioned regions.(i)WeakNorth(a) is the region that covers the tiles NW(a), N(a), and NE(a). WeakNorth(a) NW(a) N(a) NE(a).(ii)Horizontal(a) is the region that covers the tiles W(a), O(a), and E(a). Horizontal(a) W(a) O(a) E(a).(iii)WeakSouth(a) is the region that covers the tiles SW(a), S(a), and SE(a). WeakSouth(a) SW(a) S(a) SE(a).(iv)WeakWest(a) is the region that covers the tiles SW(a), W(a), and NW(a). WeakWest(a) SW(a) W(a) NW(a).(v)Vertical(a) is the region that covers the tiles S(a), O(a), and N(a). Vertical(a) S(a) O(a) N(a).(vi)WeakEast(a) is the region that covers the tiles SE(a), E(a), and NE(a). WeakEast(a) SE(a) E(a) NE(a).
4. RCC Model
RCC stands for region connection calculus [13, 18, 31]. It is a firstorder theory employed for qualitative spatial representation as well as reasoning and is based on Clarke’s logic of connection [32, 33]. The connection predicate, , which means “region a is connected with region b”, is the only primitive predicate for RCC. This dyadic relation is both reflexive and symmetric and holds whenever regions and are “connected.” The two main axioms expressing reflexivity and symmetry [18] are as follows: Based on this primitive, a basic set of dyadic relations are defined as shown in Table 1.

The relations P, PP, TPP, and NTPP are nonsymmetrical and will have their respective inverses (Pi, PPi, TPPi, and NTPPi). Of all the listed relations, only 8 relations in the following set {DC, EC, PO, EQ, TPP, NTPP, TPPi, NTPPi} are provably jointly exhaustive and pairwise disjoint (JEPD—which means any two regions are related by exactly one of these eight relations [34, 35]). Randell and colleagues [13] refer this set of relations as RCC8, and they are depicted in Figure 6.
5. Expressive Hybrid Model
In our expressive hybrid model, we have combined our Horizontal and Vertical Constraints Model [29, 30] and RCC8 [13].
5.1. Definitions
If there is a referent region and another arbitrary region , the possible basic binary relations between them can be defined as below.
In terms of weak relations,(i)WeakNorth(b,a): b WeakNorth(a),(ii)Horizontal(b,a): b Horizontal(a),(iii)WeakSouth(b,a): b WeakSouth(a),(iv)WeakEast(b,a): b WeakEast(a),(v)Vertical(b,a): b Vertical(a),(vi)WeakWest(b,a): b WeakWest(a).
In terms of RCC8 relations,(i)DCy(a,b): ydimension of is disconnected from ydimension of b, (ii)EQy(a,b): ydimension of is identical with ydimension of b, (iii)POy(a,b): ydimension of partially overlaps ydimension of b, (iv)ECy(a,b): ydimension of is externally connected to ydimension of b, (v)TPPy(a,b): ydimension of is a tangential proper part of ydimension of b, (vi)NTPPy(a,b): ydimension of is a nontangential proper part of ydimension of b, (vii)TPPiy(a,b): ydimension of is a tangential proper part of ydimension of a, (viii)NTPPiy(a,b): ydimension of is a nontangential proper part of ydimension of a, (ix)DCx(a,b): xdimension of is disconnected from xdimension of b, (x)EQx(a,b): xdimension of is identical with xdimension of b, (xi)POx(a,b): xdimension of partially overlaps xdimension of b, (xii)ECx(a,b): xdimension of is externally connected to xdimension of b, (xiii)TPPx(a,b): xdimension of is a tangential proper part of xdimension of b, (xiv)NTPPx(a,b): xdimension of is a nontangential proper part of xdimension of b, (xv)TPPix(a,b): xdimension of is a tangential proper part of xdimension of a, (xvi)NTPPix(a,b): xdimension of is a nontangential proper part of xdimension of a.
5.2. Basic Binary Relations of the Hybrid Model
In this section, we shall demonstrate how we come up with all possible binary direction relations for the hybrid model. All the possible basic binary relations for each horizontal set are shown in Figure 7. The notations that will be used in this section are as follows.(i)RELy(b,Z) is any basic binary relation between and the horizontally partitioned region, . (ii)RELx(b,Z) is any basic binary relation between and the vertically partitioned region, .
Based on Figure 7, the total number of possible binary relations for the hybrid model in the ydirection is which equals 44 cases. However, due to the singlepiece condition, the following rules apply.
Rule 1. (b WeakNorth(a) ∧ WeakSouth(a)).
Rule 2. Assume to be {WeakNorth(a), Horizontal(a), WeakSouth(a)}. If NTPP where then (NTPP) ∧ REL, where , or (NTPP∧ REL) ∧ REL, where .
Rule 3. Assume to be {WeakNorth(a), WeakSouth(a)}.
If (TPP(,Horizontal()) EC, where ,
then (TPP,Horizontal) EC REL,
where .
Based on the rules above, the total number of feasible binary relations for singlepiece regions in the ydirection is (444234) which equals 13 cases. The thirteen feasible and jointly exhaustive binary relations for the hybrid model are depicted in Figure 8. This means that, in the hybrid model, the number of jointly exhaustive binary relations (in both the  and directions) that hold between two singlepiece regions will be . This concurs with the basic relations in the Rectangle Algebra Model [16, 17].
6. Combined Mereological, Topological, and Cardinal Direction Relations
Mereology (from the Greek μερος, “part”) is the theory of parthood relations: of the relations of part to whole and the relations, of part to part within a whole [36]. In this section, we shall make two distinctions: “whole and part” cardinal directions, as well as “weak and expressive” relations. We shall rewrite the notations used in our previous paper [29]. means that only part of the destination extended region, , is in tile R(). The direction relation means that whole destination extended region, , is in the tile R(). As an example, when is completely in the SouthEast tile of , this direction relation can be represented as shown below:
The “whole and weak” direction relations are defined in terms of horizontal and vertical sets:(i) WeakNorth() Vertical(),(ii) WeakNorth() WeakEast(),(iii) WeakNorth() WeakWest(),(iv) WeakSouth() Vertical(),(v) WeakSouth() WeakEast(),(vi) WeakSouth() WeakWest(),(vii) Horizontal() WeakEast(),(viii) Horizontal() WeakWest(),(ix) Horizontal() Vertical().
The “whole and expressive” direction relations are defined in terms of expressive horizontal and vertical sets. A general form of such direction relation can be represented as follows: where H(a) and V(a) are horizontally and vertically partitioned regions for , respectively, where and R(a) (H(a) V(a)).
7. Composition Table
Composition is a common inference mechanism for a wide range of relations and has been exploited for automated reasoning. It has been employed for reasoning about temporal descriptions of events based on intervals [37], topological relations [5, 38–42], direction relations [1, 24, 29, 30], and combined topological relations with cardinal direction relations [19]. To reiterate, one of the main advantages of using composition tables is that they can lead to tractable computation of significant classes of inference [25].
Given the relation between and and the relation between and , a composition table allows for concluding the relation between and . Bennett [41] defines the concept of the composition of two binary relations as follows. Given a theory Θ which is used to define a set β of mutually exhaustive and pairwise disjoint dyadic relations (i.e., a basis set), the composition, Comp(, ), of two relations and which are taken from is defined to be the disjunction of all relations in , such that, for arbitrary constants , , and , the formula is consistent with Θ.
7.1. Composition of Regions with Parts
In our previous paper [29], the method for computing the composition of cardinal direction relations for part regions is not robust enough, because it does not hold for all cases. In order to address this problem, we introduce a formula (obtained through case analyses) for computing the composition of cardinal direction relations. The basis of the formula is to consider the direction relation between and each individual part of followed by the direction relation between each individual part of and .
Assume that the region covers one or more tiles of region while region covers one or more tiles of .The direction relation between and is while the direction relation between and is . The composition of direction relations could be written as follows: Firstly, establish the direction relation between and each individual part of b: where .
Consider the direction relation of each individual part of and . Equation (7) becomes where .
7.2. Composition of Weak Direction Relations
Firstly, we shall demonstrate how to apply the formula for the composition of weak direction relations followed by more expressive direction relations.
Type 1. (().
Find the composition of .
Use (8) with and :
The outcome of the composition is
This means that the region c O(a) W(a) S(a) SW(a).
Type 2. (().
Find the composition of .
Use (8) with , and :
The outcome of the composition is
Viewing the fact that and , the above outcome can be written as
This means that the region E(a) O(a) W(a) NE(a) N(a) NW(a).
Type 3. () ().
Find the composition of (()). Establish the relationship between and each individual part of . In this case, (), () and () holds (this is not necessarily true for all cases).
Use (8) with and .
Therefore, the above composition can be rewritten as
The outcome of the composition is
This means that the of the minimal bounding box for region is greater than of the minimal bounding box for region and .
Type 4. (().
Find the composition of
Figure 9 has been drawn for this example. Establish the direction relation between each individual part of and .
Use (8) with ; the value of for is , while the value for is :
In part of the above composition, . To simplify the composition, we consider the combined horizontal and vertical sets of all the parts of .Thus, we have the following:
In part of the above composition, , , , , , , . The simplified version of the composition is as follows:
The final outcome of the composition is part part is equivalent to
This means that the region which is the union of all the 9 tiles of region . However, based on Figure 9, region .
7.3. Composition of Expressive Direction Relations
We shall use the following notations to represent the 13 binary ydirection relations:(i)REL1NTPP(,WeakNorth()), (ii)REL2()TPP(,WeakNorth()) EC(,Horizontal()),(iii)REL3()TPP(,Horizontal()) EC(,WeakNorth()),(iv)REL4()TPP(,Horizontal()) EC(,WeakSouth()),(v)REL5()NTPP(,Horizontal()),(vi)REL6()EQy(,Horizontal()),(vii)REL7()NTPP(,WeakSouth()),(viii)REL8()TPP(,WeakSouth()) EC(,Horizontal()),(ix)REL9()PO(,WeakNorth()) PO(,Horizontal()) DC(,WeakSouth()),(x)REL10()PO(,WeakNorth()) PO(,Horizontal()) EC(,WeakSouth()),(xi)REL11()PO(,WeakNorth()) PO(,WeakSouth()) NTPP(,Horizontal()),(xii)REL12()PO(,WeakSouth()) PO(,Horizontal()) DC(,WeakNorth()),(xiii)REL13()PO(,WeakSouth()) PO(,Horizontal()) EC(,WeakNorth()).
Similar notations will be used to represent the 13 binary direction relations (WeakNorth is replaced by WeakEast, Horizontal with Vertical, and WeakSouth by WeakWest).
Example 1. Find the composition of the following:
Establish the direction relation between and each individual part of . Use (8), with and , and :
Use (5), and the above composition can be rewritten in the following expressive form:
Use Tables 2 and 3, and and . Thus, the outcome of the composition can be written as follows:
The outcome of the composition is:
(a)  
 
Note: the data in bold are results of the composition of relations.  
(b)  
 
Note: the data in bold are the results of the composition of relations.  
(c)  
 
Note: the data in bold are the results of the composition of relations.  
(d)  
 
Note: the data in bold are the results of the composition of relations. 
(a)  
 
Note: the data in bold are the results of the composition of relations.  
(b)  
 
Note: the data in bold are the results of the composition of relations.  
(c)  
 
Note: the data in bold are the results of the composition of relations.  
(d)  
 
Note: the data in bold are the results of the composition of relations. 
Example 2. This example is similar to the fourth example in the previous section of this paper.
Find the composition of
Establish the direction relation between and each individual part of .Use (8), with and , and .
The composition in expressive form will be as follows.
For part ,
The regions , , , ; the above composition can be written as follows:
For part , The final outcome of the composition is the composition of part (29a) part (29b).
Apply Rule 3 from the earlier part of the paper, and we will get the following: