COMPUTER VISION, GRAPHICS, AND IMAGE PROCESSING 49, 1-35 (1990)

A Solution of Polygon Containment, Spatial Planning, and

Other Related Problems Using
Minkowski Operations
PIJUSH K. GHOSH
National Centre for Software Technology, Gulmohar Cross Road No. 9,
Juhy Bombay 400 049, India

Received May 23,1988; revised February 15,1989

This paper gives a complete solution to the polygon containment problem under translation
and other related problems for all kinds of two dimensional regions. The solution is achieved
in three steps. First, it is shown that the containment and the related problems can be directly
mapped to Minkowski decomposition and addition problems. Minkowski decomposition,
which is intrinsically a geometric problem, is then reformulated in terms of set operations and
set theoretic tools are used to reduce the computational complexity of the problem. Finally, a
new technique, termed as decomposition boundary tracing technique, is devised and employed
for the solution of the decomposition problem. This new technique brings out a unfied
algorithmic approach to solve all kinds of decomposition problems in a two dimensional space,
and thereby the containment problem in its entirety. 0 1990 Academic Press. Inc.

1. INTRODUCTION

1.1. The Problem and the Motivation

The polygon containment problem is to decide whether a polygon B can be
translated to fit inside another polygon S, and if it fits then to compute the set of
safe positions in which B fits inside S. This set of safe positions is also known as
feasible region.
The concept of polygon containment has a number of salient applications in the
real world. An immediate application is in stockcutting, i.e., in cutting successively a
polygonal object B out of a sheet of material S. The tolerance inspection problem
often arises in a manufacturing process, in a vision system, or in template matching.
The problem is as follows. In 2-dimension, the task is to generate a given polygonal
object B, but one is getting a polygon B’, not necessarily equal to B because of
noise, fabrication, or other generation errors. Tolerances are given in terms of two
polygons S, and S,. One, say S,, lies outside B and the other, i.e., S,, lies inside B.
If the generated polygon B’ is completely contained in S,, and S, is completely
contained in B’, then B’ is considered to conform to the desired tolerances [l].
Another important application of polygon containment is seen in spatialplanning.
The spatial planning problem involves placing an object among other objects or
moving it without colliding with nearby objects [9]. Let S be an object that contains
n other objects His. Let B be a new object. In a real situation, S may be the entire
floor space, His may be the obstacles in the floor, and B may be a robot vehicle or a
machine tool. The spatial planning is the problem of finding safe positions for B
inside S, such that there is no intersection between B and any of the obstacles His.
That means, Hi n B = 0. The problem of finding a path for B to move it from
one position to another position is an equivalent problem, since a path is nothing
1
0734-189X/90 $3.00
Copyright Q 1990 by Academic Press, Inc.
All rights of reproduction in any form reserved.
2 PIJUSH K. GHOSH

but a connected sequence of safe positions. In 2dimension, the spatial planning

problem, therefore, is equivalent of computing the feasible region for a polygonal
object B inside a polygonal object S that has n number of holes His in it. The
problems like finding out the intersection of two polygons or determining the distance
between two polygons are also special cases of the polygon containment problem.
Certain practical situations demand that the polygon containment problem must
deal with both translation as well as rotation of B. For example, changing the
orientation of the robot vehicle at some position might make it possible to move the
vehicle through the obstacles without collision, which was otherwise not possible by
translation alone. But in this paper, we tackle the problem by keeping the orienta-
tion of B fixed. The exact problem we handle here can be stated as follows.
PROBLEM 1 (polygon containment under translation). Two polygons S and B are
given; the polygons might have holes in them. Compute the feasible region in which B
can be translated to$t inside S.
We solve the problem by using Minkowski addition (vector addition) and
Minkowski decomposition (inverse of addition) operations. Our approach is very
different from the existing ones. The primary advantage of our approach is that. it
allows us to formulate a unijed containment algorithm for all types of S and B
polygons. The unified method can easily be extended to 3-dimensional as well. In
addition, we also develop a theoretical framework for Minkowski decomposition
(which is the primary tool for solving the containment problem) and show how set
theoretic results can be used to reduce the complexity of a geometric problem. Our
treatment of the problem is briefly sketched in Section 1.3.

1.2. Earlier Works

The problem had been studied by a number of researchers in various contexts.
Vaishnavi and Wood [15] and later Lee and Preparata [8] investigated the rectangle
containment problem that involves only rectangular polygons. Chazelle [l] reported
a linear time algorithm in the case where both S and B polygons are convex. If the
convex polygons S and B have Ns and Ns number of edges, respectively, and are
given in slope-sorted order, Chazelle’s algorithm computes the feasible region in
time O( N, + NB) or O(N), where N = Total number of edges involved = Ns + NB.
Recently Baker et al. [l] also obtained the same complexity, but by using somewhat
different techniques. They reported an 0(( N,N, + Ni)log N,N,) algorithm when B
is convex but S is a simple (nonconvex) polygon.
Although the reported algorithms vary considerably in geometric techniques and
data structures, the fundamental idea is same everywhere. Place B inside S such
that B touches one of the edges of S; translate B continuously along that edge and
test whether B crosses the boundary of S at any point; if B hits the boundary, then
choose the next edge of S and translate B along that edge in the similar manner,
and so on. The translational path of B gives the boundary of the feasible region
(Fig. 1). All these algorithms are also similar in another regard. The basic module
inside the hearts of these algorithms is a containment algorithm for two convex
polygons. To solve the general problem, the two given polygons are first ‘broken up
into convex parts and at the final stage intermediary results are combined to obtain
the solution. These kinds of algorithms can be termed as the move and test
algorithms.
 A SOLUTION OF POLYGON CONTAINMENT 3

FIG. 1. The basic idea behind existing polygon containment algorithms.

In Fig. 1 we demonstrate how the B polygon is translated along the edges of S

and show by dark dotted lines the boundary of the feasible region. Note the fact
that a placement of B is uniquely determined by a placement of a specific point of
B, since only translational motion of B is allowed. We may call this specific point
the reference point r. The boundary of the feasible region is, in fact, the translational
path or the locus of r. If we choose some other reference point, say r’, the feasible
region is only shifted/translated, but its shape remains unchanged. The boundary of
the feasible region with respect to r’ is shown by light dotted lines in Fig. 1. In this
paper we are primarily concerned with the shape of the feasible region, without
worrying much about the location of the reference point. Once the shape of the
feasible region is obtained, it can be translated appropriately.

1.3. Our Approach and Objective

Our approach is completely different from the existing move and test approach.
The problems concerning interactions between objects (for example, containment of
one object within another, movement of an object among obstacle objects, distance
between two objects, etc.) become relatively simpler if the moving object B is a
point instead of an object with finite size. Consider, for example, our polygon
containment problem. If the given B is a point, then the entire S polygon itself
becomes the feasible region. We, therefore, try to derive a transformation operation
so that B in the real space is transformed to a point, say to the reference point r, in
the transformed space. That transformed space is commonly krrown as configuration
space, or simply Cspuce. In the Cspace, since B shrinks to a point, the other objects
must also shrink accordingly. On the other hand, if there are holes in the other
objects, they mu% grow in accordance (Fig. 2).
In Fig. 2 the object S is a polygon with a hole and we choose r as the reference
point of B. After transformation the polygon B is transformed to the point r, while
S is transformed to a polygon T. The nature of the transformation must be such
that when r is inside the T polygon in the Cspace, B must be completely contained
in S in the real space; on the other hand B must not be completely contained in S
when r turns out to be outside of T. The polygon T, therefore, represents the
required feasible region.
However, the fundamental question is: what is that transformation operation?
Because once we have that transformation operation, our task is merely to trans-
form S accordingly, and that transformed S, i.e., T will be our feasible region.
We show that Minkowski decomposition operation takes care of the shrinking of
an object exactly in the way we want to shrink them in the Cspace. The growing of
holes, in accordance with the shrinking of the object, is the inverse process, and is
done through inverse of Minkowski decomposition. This inverse operation is com-
monly known as Minowski addition or vector addition.
4 PIJUSH K. GHOSH

Transformation

Reol space Configuration space

FIG. 2. Transformation from real space to configuration space.

Our solution of the containment problem, therefore, reduces to two steps:

1. to establish first the formal relationship between Minkowski operations and

the containment and other related problems;
2. to devise then the appropriate algorithms to carry out Minkowski operations
on polygonal objects.
To achieve this aim, we organize the paper in the form:
1. In Section 2 we briefly state the primary notations and definitions that are
followed in this paper. The formal definitions of Minkowski addition and decompo-
sition are also included in this section.
2. The connections between Minkowski operations and the polygon contain-
ment, the spatial planning, the polygon intersection problems etc. are established in
Section 3. It is shown that Minkowski decomposition of a polygon S by a polygon
B is exactly the same as the containment of the polygon B inside the polygon S. In
other words, the containment problem can be directly mapped to the Minkowski
decomposition problem. In this paper we decide to concentrate only on the
decomposition problem and not on the Minkowski addition problem, since the
latter operation is needed only if the polygons have holes. Moreover, the problem of
Minkowski addition has been tackled in great detail in other places [3].
3. In Section 4 we start devising algorithms for Minkowski decomposition on
various kinds of polygons. It may appear at first that the algorithms for decompos-
ing a convex polygon by a convex polygon, a simple polygon by a convex polygon, a
simple by a simple, etc., will all be different. Interestingly a new notion, termed as
decomposition boundary tracing, emerges from the way we formulate these algo-
rithms. This notion allows us to devise a algorithmic technique to carry out
all kinds of decomposition within the same framework. This section is divided into
three parts.
(a) In the first part we attack Minkowski decomposition problem from a set
theoretic point of view. The decomposition problem is generalized as the
decomposition of a set by another set in the d-dimensional real euchdean
 A SOLUTION OF POLYGON CONTAINMENT 5

space, instead of a polygon by a polygon. Then the problem is reformu-

lated in terms of set operations. A number of set theoretic results
concerning Minkowski decomposition of two sets are stated and proved.
We show how these set theoretic results could be effectively used to
reduce the computational complexity of our problem that is very much
geometric in nature. The results we obtain in this part are applicable for
general sets and thereby allow us to go beyond 2-dimensional. space and
polygonal regions.
(b) In the second part, however, we decide to restrict ourselves to 2dimen-
sion. We start by presenting an algorithm to decompose a convex
polygon by another convex polygon (wiIl be referred to as the
convex-conuex case for convenience). The important contribution at this
part is to introduce the notion of decomposition boundary tracing in
formulating the algorithm. We also show here that the complexity of our
algorithm does not increase because of our formulation. It is also linear,
as was obtained by the previous researchers.
(c) In the third part we intend to examine whether the notion of a decompo-
sition boundary tracing technique could be extended beyond convexity.
Therefore, we undertake the task of decomposing a simple polygon by a
convex polygon (will be referred to as the simple-convex case). The
tracing technique does not need any modification for the convex parts of
the simple polygon. The nature of output, i.e., the tracing also remains the
same as in the case of the convex-convex case. We show that for the
nonconvex parts of the simple polygon also, the tracing technique re-
mains exactly the same but only the nature of the tracing changes. By
following our approach we are also able to get an algorithm of the same
complexity as obtained by Baker et al. Moreover, it appears that the
average case complexity of our algorithm could be reduced considerably
by devising more efficient data structures. However, we decide not to try
out those finer details in this paper.
4. Once it is established that the same decomposition boundary tracing tech-
nique can be used for both convex-convex and simple-convex cases, the possibility
of forming a unitied approach to all decomposition algorithms in 2dimension
emerges. In Section 5 we analyze the basic concept behind the tracing technique and
propose how this concept can be universally used for all kinds of 2D decomposi-
tions. The concept behind the tracing technique is expressed in the form of two
propositions. The applicability of the tracing concept beyond convexity and also
beyond polygonally defined regions is more strongly established by devising algo-
rithms for the simple-simple case and for regions that are bounded by smooth
boundary curves.
5. Section 6 is the concluding section. We briefly indicate how our approach
could be extended to 3dirnensional and also suggest a few important problems for
further investigation.
We must also mention the following facts regarding our paper.
1. The term con&ration space is not coined by us. Lozano-Perez [9] previ-
ously used the term and also the concept of growing obstacle objects in connection
6 PIJUSH K. GHOSH

with the spatial planning problems. Our definition of configuration space, however,
is a more general concept that accommodates both growing as well as shrinking of
objects, since they are just the inverses of each other. Thus the concept of our
configuration space covers a wide range of problems concerning interactions among
objects and their movements in space.
2. In recent years a general method for image processing, known as mathemati-
cal morphology has become exceedingly popular and is growing rapidly as a separate
discipline. Interestingly, Minkowski operations form the kernel of all the mathemat-
ical morphological operations. In fact, two of the morphological operations, known
as dilation and erosion are nothing but Minkowski addition and Minkowski
decomposition, respectively. Although mathematical morphology is primarily con-
cerned with the processing of digital images, in morphological literatures many
results on Minkowski operations were stated for general sets of points in d-dimen-
sional real euclidean space. Therefore, the statements of some of the theorems and
equations presented in Sections 3 and 4.1 in this paper are not completely new.
However, in certain cases, results were stated in morphological literature without
any formal proof. Therefore, we decided to provide separate proofs for all of them
which are also more algebraic in nature. The interested readers may refer to
Matheron [lo], Serra 1121, or Harlick et al. [6] for alternative proofs of some of the
theorems.
3. The proof of the proposed unified theory of decomposition is presented in a
more informal way compared to the proofs of the theorems in the first part of the
paper. It has served a number of useful purposes. First, quite a few interesting
geometric properties of Minkowski decomposition have been unravelled during the
discussion which could not have been done otherwise. Second, we are able to show
how a general approach to polygon decomposition has gradually evolved as we
progressively reduce the constraints on the input polygons. Moreover, the theory of
Minkowski addition was developed rigorously by a number of researchers [4, 10, 5,
31. Since Minkowski decomposition is the inverse of Minkowski addition, it is
possible to formulate more formal proofs of the unified theory of decomposition
along the same line of addition. But we feel it is more appropriate to introduce a
different line of thought and, therefore, do not make use of any of the theorems of
Minkowski addition. The decomposition theory also becomes more self-contained in
that process.
2. BASIC NOTATIONS AND DEFINITIONS

2.1. Notations
The objects we consider in this paper are the sets of points in E = R”, i.e., in
d-dimensional real euclidean space. These sets are generally denoted by capital
letters A, B, S, T, etc. We consider only the compact sets, i.e., which are both
bounded and closed. Any point p in such a set is a vector with d-tuple (xi,. . . , xd).
Most of the results we prove will be applicable for such general sets of points. A
point is generally denoted by a small letter such as a or p or u.
To distinguish between the boundary of a positive region and a hole, the notion
of oriented boundary is introduced in Section 4.
The notations U, n, 1, and ’ are used to denote set union, intersection,
difference, and complement, respectively. The notation “ + ” is used to de-
 A SOLUTION OF POLYGON CONTAINMENT 7

note Minkowski/vector addition with a point, whereas “ @ ” is used for denoting

Minkowski addition of two sets. Similarly, “ - ” and “ 8 ” notations are used for
Minkowski decompositions.
The letter N is kept reserved for denoting the number of vertices or edges of a
polygon. The other definitions or notations that we use are stated at the places
where they first occur in our presentation.

2.2. Minkowski Operations

2.2.1. Minkowski Addition

Minkowski addition can be regarded as a kind of growing or dilation. A precise
definition of Minkowski addition can be given as follows.

DEFINITION 1 (Minkowski addition). Let B and T be two arbitrary sets in Rd

space. The resultant set S is obtained by positioning B at every point of T, i.e.,,
vectorially adding all the points of B with those of T. We denote this by

where “@ ” stands for Minkowski addition.

A simple example in 2D may clarify the idea. Let B be an elliptical region and T
be an open coplanar cubic curve. If B is considered as a brush and T as a trajectory,
then the resulting figure S is obtained by moving the brush along the trajectory (Fig.
3). This process can also be viewed as growing of the curve T to S, by means of B.
We consider that in Cspace B becomes the reference point r. Note that B @ T in
real space is equal to S @ {r} in Cspace. The same concept is true in higher
dimensions as well.
Minkowski addition of two sets can also be expressed in terms of set union
operation. If A, denotes translate of a set A by the vector p, i.e., A, = A @ ( p},

S=B@T

FIG. 3. An example of Minkowski addition in 2D.

8 PIJUSH K. GHOSH

then it obviously follows that,

S=TeB=BeT= uB,= UT,. 0)

tST beB

2.2.2. Minkowski Decomposition

Minkowski decomposition is the inverse of Minkowski addition operation; it can
be regarded as shrinking or erasing. It is denoted by the symbol “ 8 .” Its precise
definition can be given a~ follows:
DEFINITION 2 (Minkowski decomposition). If S and B are two sets in Rd space,
then

SUB= ns-b= n s-b, (2)

668 -be&

where “ 8 ” stands for Minkowski decomposition operation. We call S 8 B as the

decomposition of S by B. (According to our previous notation, c‘S-b” means the
translate of S by the vector -b, i.e., S 8 ( -b}. For the convenience of notation,
S_ b is also written as “S - b.“)
The set 5 = { - b: b E B } is generally known as the symmetrical set of B with
respect to the origin. Since our final aim is to shrink B to the reference point r, we
choose reference point as the origin point (Fig. 4).
Let us clarify the idea of Minkowski decomposition by an example shown in
Fig. 5. The polygon S, we consider, is a trapezoidal regicn while the polygon B is a
rectangular area. Since B is centre-symmetric, B and B ,are $e same in this case.
For decomposing S by B, S is placed at every point of B and their common
intersection region is determined. The resultant intersection is a triangular region as
shown separately in the figure.
It might be difficult to conceive that the above definition of Minkowski decompo-
sition is indeed the inverse of Minkowski addition. Our claim that the above
definition is the inverse of addition can be established by showing the fact that

(Se B) CBB = S,providedthereexistsaset Xsuchthat S = X0 B.

------

FIG.~. Aset Banditssymmetricalset fiin2D.

 A SOLUTION OF POLYGON CONTAINMENT 9

FIG. 5. An example of Minkowski decomposition in 2D.

The above relation can be easily proved by using Eq. (1) and Eq. (2). One derivation
of it is given in Ghosh [3]. Readers may also consult Matheron [lo]. It is important
to note here the necessity of the condition S = X $ B for proving the relation. The
necessity arises since we have no concept of independent existence of a negative
region. Therefore, there does not always exist a set X for any B and S such that
S = X 0 B. For example, no Xexists if B is a circle and S is a triangle. Now S 8 B
is supposed to give us the X. If no such X exists, then (S 8 B) $ B cannot be equal
to s.
3. RELATIONSHIP BETWEEN MINKOWSKI OPERATIONS AND CONTAINMENT
AND OTHER RELATED PROBLEMS

3.1. Polygon Containment Problem

Let us first define the.polygon containment problem more formally. Let T denote
the set of safe positions, i.e., the feasible region in which a given B fits inside a
given S. That essentially means that, if B is translated to any point t E T, the
translate polygon B, must be within S, i.e.,

T= {t: B,c S}.

In the polygon containment problem, our task is to compute the set T.

We now prove the following relation to establish the connection between the
containment problem and Minkowski operations.
S0B=Feasibleregion= {t: B,cS}. (3)

Proof: Let
XES~B
0 x E VbS-,, where b E B
-Vb(x+b)ES
-B,cS

Thus,
SeB= {t:B&S}.
10 PLJUSH K. GHOSH

i. .
---.y-y---- --
.1
.I
I . ’ . I

vB
;. ’ 7
I ’ . I
I. * I
I. . i
I. . -1 r

FIG. 6. Demonstration of spatial planning problem as polygon containment problem.

Therefore, the problem of computing the feasible region T is the same as

computing the set (S 8 B). In other words,the containment problem can be directly
mapped to Minkowski decomposition problem. Problem 1 can, therefore, be re-
stated in the following form.
PROBLEM la (Minkowski decomposition). Given two sets S and B in Rd, compute
the set S 8 B, where 8 denotes Minkowski decomposition.
Note that the statement of Problem la is much stronger than that of Problem 1. S
and B are two planar polygons in the former, whereas they can be general sets in
Rd space in the latter. However, in this paper, we emphasize computing S 8 B
when both S and B are polygons in R2.
3.2. Spatial Planning Problem
It is already pointed out in Section 1 that the spatial planning problem is
equivalent to the following problem. Let S be a set (the entire floor space, for
example) having a hole H (equivalent of an obstacle) in it. The task is to find for a
set B (say, a moving object), the set of safe positions (i.e., the feasible region) in
which B fits inside the positive region in S. In Fig. 6 we show a typical example.
The feasible region for B is shown dotted. Note the forbidden region around the
obstacle H, where B cannot be placed. It may be viewed as if the obstacle H has
grown to the forbidden region when B has shrunk to the reference point r. The
forbidden region can be considered as a hole inside the feasible region.
Now the positive region, say S’, in S is equal to the entire S minus the hole H,
i.e.,
S’ = S \ H = S n H’.
Therefore, according to our previous notation,

Feasible region = S’ 0 B = ( S n HC) 8 B. (4)

Equation (4) can be simplified by using two equations

S 8 B = (SC @ 8)” (5)

(S, n S,) 0 B = (S, 9 B) I-I (S, 8 B). (6)
 A SOLUTION OF POLYGON CONTAINMENT 11

Proof of equation (5). We first claim that,

(SC)-, = (S-J’.

x E (Q-6
-x+bES’
ox+b4S
@ x 4 S-b
* x E (S-J.

Now.

S’fBB= u (SC)-,=u (se,)’ (according to the above result).

-bcB -be:B

Therefore,

(s’c~ij)== f-) S-,=SeB.

-be.i)

The relation is proved.

Proof of Equation (6). The relation can be proved directly from the definition of
Minkowski decomposition. One may also prove it by using Eq. (5).
Now applying Eq. (5) and Eq. (6) in Eq. (4) we obtain

Feasible region = (S fl H’) 0 B

=(SeB)n(H@@=
or

Feasible region = P \ Q, (44

where

P=(SeB) and Q= (Had).

The physical meaning of Eq. (4a) is: if the moving object B becomes a point, the
entire floor space shrinks from S to P and the obstacle grows from H to Q. Figure 6
demonstrates the same fact.
In solving spatial planning problem, researchers, in general, are more interested in
compu$ng the forbidden region around the obstacle H, i.e., in computing Q =
(H 8 B), rather than computing the entire feasible region (P \ Q). Therefore, the
spatial planning problem can be reduced to the following problem.
~OBLEM 2 (Minkowski addition). Given two sets T and B in Rd, compute the sum
set T @ B, where @ denotes Minkowski addition.
12 PIJUSH K. GHOSH

Problem 2 was elaborately discussed and tackled by Ghosh [3] as well as by

Guibas et al. [5] for 2D regions. Therefore, we shall not handle this problem in this
paper.

3.3. Polygon Intersection Problem/Distance between Two Polygons Problem

The polygon intersection problem is to find out whether two given polygons T
and B intersect or not. We have seen in the previous subsection that if the reference
point r is outside the region T $ $, the polygon B does not intersect T. On the
other hand, if r is inside, B intersects T. Note that in computing ii, we have chosen
the reference point r as the origin. Therefore,

Two polygons T and B intersect if and only if the origin is inside the
sum polygon T CBg.

(See Guibas et al. [5] for more detail.) Note that the intersection problem now
reduces to the point inclusion problem.
Similarly the mirrimum distance between two polygons T and B is the distance of
the polygon (T @ B) from the origin. Therefore, these two problems are essentially
Minkowski addition problem as stated in Problem 2.

4. SOLUTION OF PROBLEM la: COMPUTATION OF S 8 B

4.1. Building Up the Set Theoretic Framework

The definition of S 8 B immediately suggests a computational procedure. Place
the set S on every point of B and find out the common intersection of all the
instances (i.e., translates) of S.
The computational cost cou!d beVtremendously reduced if it suffices to place S
only on the boundary points 8B of B, instead of placing it on all the interior as well
as on the boundary points. By limiting ourselves on a restricted class of sets S, we
can make is possible. Note the following theorem.
THEOREM 1 (boundary reduction theorem). If S is a simply connected set (i.e.. it
has no hole in it), then

where 8B denotes the boundary of B.

Proof: In the tirst place note the fact that,

(8s z S) * (B c S), provided S contains no hole.

Refer to Fig. 7. The set S, shown hatched in the figure, has no hole in it. The
boundary cYBof B is drawn in bold. The topological argument says that if d3 can
be placed within a simply connected set S, it is then possible to continuously shrink
JB up to a point by keeping it always within S. During this topological-shrmking
process the different instances of 6JB covers aU the interior points of B. THUS the
above implication is correct.
 A SOLUTION OF POLYGON CONTAINMENT 13

FIG. 7. ~Topological shrinkage of LJB witbin a simply connected set S.

Now,
XES~JB
- (W, c S
* B, c S, provided S is simply connected
ox~S0B.

Thus the theorem is proved.

In order to improve the computational efficiency in decomposition, we have
proposed to limit the set S to be a simply connected set. But that does not prevent
us altogether from handling a multiply connected S. In that case, we have already
shown through Eq. (4a), that a hole in S can be treated separately and the
intermediary results tlnally can be combined together by the set dilference opera-
tion. But a more important question that crops up at this point is: “Is it possible to
increase the computational efficiency further by limiting to a more restricted class of
sets S?” The answer is, “Yes, indeed.”
Let us now consider S to be restricted to a convex set. We can then show a
considerable improvement in computational efficiency. The results of this investiga-
tion are presented through Theorems 2 to 5.
THEOREM 2 (theorem on convex decomposition by line segment). If S is Q
convex set and L is a line segment whose end points are b, and b,, then

S 8 L = S 8 {b,, b,}.

Proof. Let L = {b,, b,, . . . , bi, . . . , b,}, where bj denotes any intermediate point
on the line segment. Any b, can be expressed as

bi = r. b, + (1 - r) . b,, where 0 -C r < 1 (4

From the definition of the decomposition,

S 8 L = [(S-b,) n (S - b,)] n [(S-b,) n ..a n(S - bi) es.].

Theorem 2 will be proved if we can show that any t E [(S - b,) n (S - b,)] is also
an element of (S - bi).
Let us express t as follows:

t=sj-b,=s,-b,, where sj, sk E S.

14 PIJUSH K. GHOSH

(The representation “sj - 4” means “s, -I- ( - bs),’ etc.) Therefore,

be - b, = sk - s,. 0-J)

Let us assume that, t = x - b,, where x is any arbitrary vector. If we can now show
that x E S, then (x - b,), i.e., t becomes an element of (S - bi), and that will
complete the proof of the theorem.
Since,

t=x-bi=s,-b e,

therefore,

x = r. sj + (1 - r) * Sk (using Eq. (a) and Eq. (b)).

Since S is a convex set and sj, sk E S, therefore, r . sj + (1 - r) . sk must be an

element of S, i.e., x E S. Thus the theorem is proved.
The importance of the theorem may be felt better if we restate it as follows.
THEOREM 2a. If S is a convex set, then the decomposition of S by two points is
equal to the decomposition of S by the entire line segment joining these two points.
Theorem 2 serves as a basis for obtaining some more interesting results. Consider
the following theorem.
THEOREM 3 (theorem on reduction to convex hull). If S is a conuex set, then

S 8 B = S 8 conv(B),

where conv( B) denotes the convex hull of B.

Proof. Let B be a set of points in Rd space denoted by

B = {bi: i = 1,2 ,..., PZ}.

The convex hull conv( B) is the set of all points p which can be written in the form
(known as convex combination form):

p = cw,b, + Cy,b, + . . . +a,,b,,

where,

ai E 19, ai 2 0, a1 + a.2 + . * * +a, = 1.

Any point p of the convex huh can also be rewritten in the form

P = “lb, + (1 - dp’,
 A SOLUTION OF POLYGON CONTAINMENT 15

where,

According to the above representation of p, it is an intermediate point on the line

segment joining b, and p’. And Theorem 2 says, every such intermediate point p
will be automatically included within the decomposition S 8 {b,, p’}.
The point p’ can likewise be expressed in terms of b, and p”, etc., and by
proceeding in this manner we fhmlly get that any point p E conv(B) becomes
automatically included in the decomposition S 8 { b,, b,, . . . , b,, }, i.e., in S 8 B.
Thus the theorem is proved.
The implication of this theorem is significant. The theorem states that if S is
convex, then independent of Ihe type of B, the decomposition of S by B can be
carried out by an algorithm for the convex-convex decomposition only (provided
an algorithm to compute the convex hull of a set is available).

In fact, it is possible to simplify our problem further. Note the following theorem.
THEOREM 4 (theorem on reduction to vertices). If S and C are both conuex sets,
then
S 0 C = S 0 vert(C),

where vert(C) &notes the set of vertices of C.

Proof: The proof turns out to be easy enough if we use the following definition
of vert(C): the set vert(C) is the smallest subset of a convex set C having the
property that conv(vert(C)) = C.
Therefore, if
vert(C) = {q: i = 1,2 ,..., m},
then any point c E C can be expressed as
c = up1 + azuz + -** +a,u, (q E w, ffi 2 0, a1 + . * * +a, = 1).

The rest of the proof then follows from the proof of Theorem 3.
For the purpose of concise representation, we may combine Theorems 3 and 4 by
introducing the notion of the set of extreme points of a set B, usually denoted by
ext( B). Mathematically, ext( B) is the set of vertices of conv( B). See Fig. 8 in which
ext(B) is shown in boldface..
The proposed combined theorem may be stated as follows.
THEOREM 5 (theorem on reduction to extreme points). If S is a conuex set, then
S 0 B = S 0 ext(B),

where ext(B) denotes the set of extreme points of B.

An idea of the reduction in size from the set B to the set ext(B) may be obtained
from the typical example depicted in Fig. 8. The improvement in the computational
cost, in the average case, can also be easily guessed.
16 PIJUSH K. GHOSH

------- conv (Bl

FIG. 8. The cmv( B) (shown dotted) and ext( B) (shown bold) of a set B.

4.2. Decomposition Algorithms in 20

All the results we have proved so far are valid for any decomposition in Rd space,
d being equal to or greater than 2. However, we shall now concentrate in devising
decomposition algorithms in 2-dimension. That is because our primary aim is to
solve the polygon containment problem and not the decomposition problem in the
general d-dimensional space.
The techniques and complexities of the algorithms depend directly on the types of
S and B polygons. For the convenience of the readers, we write down the ditrerent
combinations of S and B that are considered in this paper.
1. Simply connected polygons.

Al. Conuex-Convex. Here both S and B are convex polygons. Our ap-
proach is different from the conventional ones and our technique is
discussed in detail.
A2. Convex-Simple. S is a convex polygon and B is a simple polygon. In
that case Theorem 3 suggests that first we have to determine the convex
hull of B and then algorithm Al can be used.
A3. Simple-Convex. S is a simple polygon and B is a convex polygon. In
that case an algorithm can be devised by making use of the formula,

S = conv( S ) \ (holes)
= S, \ H.

The idea is to represent a simple polygon as a convex polygon minus

some holes. We can simplify the above expression by using the result of
Es. @a),
S 6 B = (S, 8 B) n (H @ -ii)‘= (S, 0 B) \ (H @ &.

Since both S, and B are convex, the (S, 6 B) polygon can be obtained
by using the algorithm Al. The polygon (H @ 5) can be obtained using
any of the standard algorithms given in Ghosh, Guibas et al., Schwartz,
etc. [3, 5, 141.
In this paper, however, we take another approach in the line of the
algorithm Al. This algorithm together with algorithm Al finally pave
the way for a unified algorithmic approach to solve all decornp&tkm
problems, i.e., the polygon containment problems in 2-dimenkon.
 A SOLUTION OF POLYGON CONTAINMENT 17

A4. Simple-Simple. Here both S and B are simple polygons. The algorithm
we present is derived from the unified approach.
A5. Smooth-Smooth. Instead of polygonal boundaries, we consider both S
and B having smooth boundary curves. This algorithm is briefly pre-
sented, primarily to demonstrate the applicability of the unified ap-
proach beyond polygonal boundaries.
2. Multiply connected polygons. No decomposition algorithm for multiply con-
nected polygons (i.e., polygons that have holes in them) is presented in this paper.
However, we must mention that some of the theorems and results we proved so far
can be effectively used for that purpose. For example, Theorem 3 states that the hole
in B polygon can be ignored during decomposition if S is convex. Similarly, Eq. (4)
can be used if S polygon has a hole in it. (Unfortunately, relationships similar to
Eq. (4) do not hold true if B has holes, since

S 8 (B, f-I H’) 1 (S 8 B,) U (S 8 H’).

Note that the equality does not hold true always).

4.2.1. Convex-Convex Case: Algorithm Al

Let S be a convex N,-gon and B be a convex N,-gon. Our task is to determine
S 8 B in this case. The definition of decomposition in Eq. (2) and Theorem 4
together suggest that our task is equivalent to solving the following problem.
PROBLEMAl (Minkowski decomposition for convex polygons). Given two convex
polygons S and B, translate and place S polygon at every vertex of I? polygon (ii is the
symmetrical set of B) and determine the common intersection among all the instances
of s.
The algorithms for intersection of two convex polygons are well known [ll]. Our
task is to find out the common intersection of NB number of convex polygons; each
are identical (only translated by different amounts) having N, number of edges.
Preparata et al. [ll] stated that the common intersection of K number of convex
N-gons can be found in O(NK log K) time. In our case, it will, therefore take
0( N,N, log NB) time.
But could we not expect a better result than that, since in our case all the N,
polygons are exactly identical? Problem Al can indeed be solved in linear time, i.e.
i.e., in time O(N, + NB). There are several ways of achieving this bound by
following conventional techniques [l]. However, we present here a somewhat
different method. This new technique finally enables us to arrive at a unified method
of decomposition in 2-dimension.
4.2.1.1. Two notations. The algorithm we are going to propose needs the follow-
ing two notions.

(i) Oriented boundary curve. If the boundary curve encloses a region, we

consider this region as a positive area and assign counterclockwise orientation to the
boundary curve. On the other hand, if the boundary curve encloses a hole, we
consider this as a void (or negative) area and give clockwise orientation to the
boundary curve. The basic idea is, if we move along the boundary curve in the
18 PIJUSH K. GHQSH

(a) (b)

FIG. 9. Oriented boundary curves.

direction of its orientation, the region on its “left” is a positive region and on its
“right” is void. In our presentation, hereafter, we assume that all the edges and line
segments are directed without explicitly mentioning this (refer to Fig. 9). A region
and its oriented boundary representation are shown in Fig. 9a. On the other hand,
in Fig. 9b we take an oriented boundary curve and show the region it represents.
Note that the triangular portion at the vertex pr is discarded at the time of region
representation, since its orientation specifies it as a void region. In reality a void
region cannot exist in isolation; it must be always surrounded by a positive region
to make it a hole.
(ii) Anti-supporting line of a conuex j&e. A convex figure F can have two
supporting lines in any given direction. (A supporting line 1 of a convex figure F
passes through a point(s) on the boundary of F such that the interior of F lies
entirely on one side of 1.) If we assign an orientation to the supporting lines, then
one of the two, say I,, will follow the orientation of the boundary curve of F while
the other one, say I,, will be exactly opposite to it (Fig. 10). We call the second one
an anti-supporting line while the first one can be called simply a supporting he.
With respect to the supporting line, the figure F lies on its left, while F lies on the
right of its anti-supporting line. The intersections F n I, and F n I, can be called
supporting point and anti-supporting point, respectively.

4.2.1.2. Basic idea of the algorIt+u. We can now come back to our algorithm.
The situation is as follows. The I? polygon has iV, number of vertices. The S
polygon is placed on every one of those NB vertices. Therefore, corresponding to
every edge, say e, of S, there will be NB number of line segments-all of which are
equal, parallel, and similarly oriented to the edge e, only translated di%rently. We

p, (supporting point)
*--- ----- ----- -- I, (supporting line)

n F

+L.l!-- ----------
pL (anti-supporting point)
IsCanti-supporting line)

FIG. 10. Supporting and anti-supporting lines of convex figure.

 A SOLUTION OF POLYGON CONTAINMENT 19

(a) (b)

Le, - seg

(cl (d)

FIG. 11. Decomposition of a convex polygon by a convex polygon.

may consider all these line segments as one bunch corresponding to the edge e of S.
We call it the e-bunch. For example, refer to the typical example depicted in Fig. lla.
The corresponding situation when S is placed on the vertices of B is shown in Fig.
llb. The e,-bunch corresponding to the edge e, of S consists of iVa line segments
which are shown by bold lines. Clearly, if S has N, edges, there will be N, such
e-bunches corresponding to every of its edges.
A cruci;al obseruation. The line segments in an e-bunch could be obtained by
placing B to the two end points ef the edge e and then joining the corresponding
vertices of the two instances of B. This is demonstrated separately in Fig. llc for
the e,-bunch. This observation will be used later in our algorithm.
Therefore, in order to determine the common intersection of all the instances of
S, our task boils down to:
1. Select from each of the e-bunch, the left-most line segment. We call it the
Le-seg. In general a Le-seg is an oriented line segment of length e. But since S is
convex in this case, we may safely assume that each Le-seg here is an i&rite line
and thereby describes a half plane. The final outcome of this selection step will be
N, number of L.e-segs corresponding to all the edges of S. In Fig. lld they are
shown by bold lines.
2. Find out the intersections among those L.-e-segs.In this particular case, it
will be equivalent to determining the common intersection of the N, half planes
they describe.
20 PIJUSH K. GHOSH

4.2.1.3. First step of the algorithm: Me&ion step. Selecting the Le-segyfrom an
e-bunch is not difficult. The Le-seg must be the anti-supporting line of the B polygon
(or a translated instance of B polygon as shown in Fig. 11~).
The proof of the above statement follows from the argeents given below. If a
l@e segment from the e-bunch is an anti-supporting line of B, then alI the points of
B will be on its right by definition. Therefore all other line segments of the e-bunch
will be also on its right (refer to our observation in a previous paragraph). That
means, that line segment must be the left-most line segment of the lot, i.e., the
Le-seg. On the other hand, if thz L-e-seg is not the anti-supporting line, then there
must be at least one vertex of B that will be on the left of the Le-seg. Therefore,
there will be at least one line segment in the e-bunch that will be on the left of the
L.e-seg. This is clearly a contradiction. Therefore, the Le-seg must be anti-supporting
line.

Step 1 of the algorithm can, therefore, be accomplished in two substeps:

la. IdentiJy, for each edge e of S, the corresponding anti-supporting

point at B.
lb. Translate i to the two vertices of S corresponding to the edge e and join
the two instances of the anti-supporting point at B by a line segment.
According to our above arguments, this line segment is the desired Le-seg.
For more clarification, refer to Fig. lld.
Repeat the process for every edge of S.
We can now find out the time bound of an algorithm for executing step 1. If the
edges of S and B are given in slope-sorted order, then the time bound for step la is
O(N, + Ns), since it is equivalent of merging two ordered set of elements. Step lb
needs O(N,) time. Therefore, step 1 takes O(N, + NB) time.

4.2.1.4. Second step of the algorithm Meme&ion step. Once we obtain these Ns
L-e-segs, the task of step 2, in this particular case, is to find out the common
intersections of the N, half planes described by these lines. This can be done in
O(N, log Ns) time [ll]. However, we can obtain a faster algorithm since these line
segments are parallel to the edges of S and are known, therefore, in slope-sorted
order. This order can be exploited to get an O(N,) algorithm.
Baker et al. [l] indicated an algorithm that indeed takes O(N,) time. We present
here a different algorithm of the same complexity with our ultimate intention of
unifying all decomposition algorithms.
The crux of our algorithm lies in describing these Le-segs through a different
representation scheme which can be termed as a decomposition boundaty tracing
scheme. (The decomposition boundary tracing of a decomposition will be denoted
by the letter D.) The scheme is, essentially, traversing or traciug the Le-segs in a
continuous manner.
The new representation of the L.e-segs through tracing is done as follows. Start
with any of the L.e-segs, say Lq-seg corresponding to the edge e, of S (see Fig. 12
as well as Fig. lld), and move along it following its orientation. According to our
previous argument we can assume that the end points of L.e,-seg lie on the
corresponding anti-supporting point at B (or at its translate), say ul. Try now to
trace the next Le-seg, say Lqseg, and let us assume that its corresponding
 A SOLUTION OF POLYGON CONTAINMENT

L!.I!l
23

e2
"1
84 '4
'1
"4
"2 Q '2 ,3
"3
el E,
S

(a) (b)

T= SOB

(exact rule) (easy rule)

(cl Tracing D (d)

e5
EXAMPLE I

‘x_
84
e6
23 -B

0
r2- -h
e7
e2 Le,- seg
6 !I
Le, - seg
el T=S@B
ee
S Tracing D
(a) (b) (cl

EXAMPLE II

FIG. 12. Two typical decomposition boundary tracings. (A line segment B must be considered as a
2D rectangle of zero width.)

a$i-supporting point at i is u2. If u1 and u2 do not happen to be the same point of

B, then first move along the edge(s) of B to reach from u1 to u2, and then move
along Le,-seg. Continue this process until alI the Le-segs are traced in this way.
Obviously we are bound to reach Lq-seg where we started. Two typical decomposi-
tion boundary tracings are shoy in Fig 12. In the figure, the edges in the tracings
corresponding to the edges of B are marked as t’. These t’-edges are keeping the
continuity in tracing the Le-segs.
22 PIJUSH K. GHOSH

Surely we have to resolvf a point before proceeding further. We hnd that, in any
tracing D, some edges of B, i.e., the r’-edges are also i$.tded. That happens when
we move from one of tht anti-supporting points at B to another anti-supporting
point along the edges of t. Now in which way should we have to move? Should we
follow the orientation of B or move in the opposite direction? For example, consider
the first example of Fig. 12. In mo+g from ui to u2, we may either move along u~c’~
by following the orientation of B, or we may go along u1u4u3u2, in the opposite
direction.
In the convex-convex case, it is a matter of choice, since both will result into the
same S 8 B as shown in Fig. 12~. Without explaining at this point we state that the
easy rule, as long as B is convex, is to move in the opposite direction.
The easy rule, however, may introduce inefficiency: there may be too many
self-intersections in the decomposition boundary tracing at times; also the total
number of edges in the tracing might become unnecessarily large as demonstrated in
Fig. 12.
The exact rule is as follows. Check whether the direction of turn of one Le-seg to
the next Le-seg in the tracing is clockwise or counterclockwise. Choose that path
along the edges of i where you can follow the same orientation. An explanation
behind the formation of these rules is clarified in Section 5.1. Until we reach that
section, we shall follow the easy rule.
A few immediate observations about decomposition boundary tracing could be
noted now:
(a) A decomposition boundary tracing D is a closed, oriented, nonsimple
(self-crossing) polygon.
(b) The D polygon is essentially a travel along the Le-segs in a continuous
manner. In that respect D may be regarded as a new representation of the Le-segs
as one single polygon.
(c) The positive region, say T, enclosed by the polygony D is the required
common intersection of different instances of S placed on B, i.e., the intended
decomposition S 8 B.
This particular observation is crucial for our future course and therefore may be
clarified a little further. We knew that the positive region T (i.e., the common region
left of all the Le-segs) enclosed by the Le-segs is the required S 0 B. In the <
polygon, in addition to the Le-segs, we have also added some edges of the B
polygon which are denoted as r’edges. Since all the Le-segs are the anti-supporting
lines of B, the polygon 5 (or its translates) will be always on the right of the
L,e-segs. Therefore, no edge of i can go to the left of the Le-segs, and thereby the
positive regioc enclosed by the Le-segs remains unaltered even after including a few
edges of the B polygon.
(d) For convex-convex decomposition, only the Le-segs of D corresponding to
the edges of S contribute to the boundary aT of the positive region T. The total
number of such L.e-segs is N,. In computing T, the f-edges of D appear to be
redundant.
(e) It can be argued that, in the case of convex-convex decomposition, the
polygon D consists of (N, + NB) number of edges corresponding to every of s
and B, provided we follow the exact rule for traversing the edges of i).
 A SOLUTION OF POLYGON CONTAINMENT 23

(a) Ii is redundant (b) Ii is not redundant

FIG. 13. Checking the redundancy of an edge.

According to our formulation, now the task of step 2 is to find out the
self-intersections of the edges of D, i.e., the intersections of the Le-segs. That will
define the boundary aT of T. We observe that two kinds of situations may happen.
Case I. When all the Le-segs of D are present in the boundary aT (Fig. 12c,
Example I), then step 2 of our algorithm can be accomplished in time O(N,).
Because the task is merely to determine the intersection points between the adjacent
L.e-segs which are given in slope-sorted order.
Ca.se II. When only K out of the N, L.e-segs (K < N,) are present in the
boundary aT (Fig. 12b, Example II), then first we have to identify the (N, - K)
Le-segs that are absent or redundant. To check the redundancy, we may proceed as
follows.
Let Ii-19 Ii, and li+i denote three consecutive Le-segs given in slope-sorted order.
Let the intersection points of the lines I,-, and Ii be a, and that of Zi and li+i be b.
The line ii will be redundant if the direction from a to b does not follow (i.e., is
opposite to) the direction of the line Ii. See Fig. 13.
The logic is easy to see. If the lines ab and Ii are oppositely directed, then the
intersection point c of Ii-i and li+ I must be on the left of Ii. Because if c is on the
right of Zi, then the slope of I,+, (which passes through the points b and c) must be
less than the slope of Ii, and that is a contradiction. Now if the lines liei, li, and
li+i describe respectively the half planes HPi-i, HPi, and HP+i, then the fact that
the point c is on the left of Ii means (HPi-, 17 HPi+,) c HP,. Therefore, HP, (i.e., in
turn li) becomes redundant since the common intersection is nothing but
“ . . - nHPi-, n HP, n HPi+i ... .” Hence by a constant time testing a redundant
Le-seg can be identified and removed. Therefore, the removal of all the redundant
L-e-segs can be done in O(N,) time.
Once we identify only the K number of Le-segs that all are present in the
boundary JT, their self-intersections can be determined in time O(K) by Ending
out the intersection points of the adjacent Le-segs.
Therefore, step 2 can be accomplished in O(N,) time. So the whole algorithm
takes O(N, + NB) time, which is linear.
The technique of this algorithm can be summarized in the following form:
1. Form the decomposition boundary tracing D.
2. Determine the positive region enclosed by D.
We can term this technique as the decomposition bounduly tracing technique, or
simply the D-tracing technique.
24 PIJUSH K. GHOSH

4.2.2. Simple-Convex Case: Algorithm A3

Let S be a simple Nsgon. B remains a convex Na-gon as before. Our task is to
determine S 8 B.

4.2.2.1. Relative supporting line. The first step involves characterization of convex
and simple polygons in terms of a supporting line. Generally a convex polygon is
distinguished from a simple polygon in the following way. For a convex polygon all
its vertices are convex, whereas for a simple polygon only some of its vertices are
convex, the others are reJIex or concave. A vertex is a concave vertex if the internal
angle at the vertex is more than 180’; otherwise it is convex.
We observe that every vertex of a convex polygon can have at least one
supporting (or anti-supporting) line, whereas a vertex, particularly the concave
vertex, of a simple polygon may not have any. Mathematicians, to remedy this
situation, extended the notion of a supporting line, known as a relative supporting
line. A line passing through a point p on the boundary of the figure F can be
considered as a relative supporting line if it is a supporting line not compared with
the whole figure F, but with F in some neighborhood of p. We shall use this general
notion of relative supporting line to tackle the problems with nonconvex figures.
(For convenience, we shall hereafter write simply supporting (or anti-supporting)
line/point to mean relative.)

4.2.2.2. A subproblem: Decompition by a line segment. As a first step toward

solving the main problem, we investigate a situation where S is a simple polygon
and B is a line segment such that it is a relative supporting line at some concave
vertex v of S. Such a situation is depicted in Fig. 14a. Gur task is to determine
S 8 B.
Since S is not a convex polygon, we have to place S at every point of 6 and then
compute the common intersection as shown in Fig. 14a. Unlike the convex case,
placing S only at the two end points of the line segment i would result into a
region which is a superset of S 8 B (Fig. 14b).
We now try to compute S 8 B through our decomposition boundary tracing
technique. But the fundamental question is: Will that same technique developed for
the convex-convex case work? It needs no explanation to state that for determining
the common intersection of the instances of S, we have to identify, in every
direction, the left-most parts of those instances. Therefore, we have to examine
whether the D-tracing technique ensures choosing the left-most parts always. To
simplify the problem, we examine the situation in two parts.
Part 1. For the Convex parts of S. As long as B is convex (as is the case here)
and we are concerned with the convex parts of S (i.e., on the edges of S incident
only on the convex vertices), the principle of formation of D does not need any
modification. Because the task remains exactly the same as before-from each of
the e-bunch corresponding to the convex edge e of S, choose the Le-seg.
Part 2. For the Nonconvex parts of S. But the problem arises for the nonconvex
parts of S, i.e., for the edges that are incident to its concave vertices. The problem is
due to relative supporting line. Note that any convex edge e of S is a global
supporting line of S. Therefore, the whole of the common intersected region (i.e.,
the region of our interest) will lie at the left of the corresponding e-bunch, and the
Le-seg of that e-bunch will be the “sole” contributing edge of the boundary of the
 A SOLUTION OF POLYGON CONTAINMENT 25

(a)

S placed on the two end The resulting common

pointsof 6 intersection

(b)

FIG. 14. Decomposition of a simple polygon by a line segment.

intersected region in that particular direction. On the other hand, any nonconvex
edge e’ of S is a relative supporting line of S within a small neighborhood.
Therefore, unlike the convex case, only a part of the intersected region, and not the
whole, is supposed to lie at the left of the corresponding e’-bunch. And clearly the
corresponding Le’-seg “cannot” be the sole contributing edge in that direction. Fig.
14a demonstrates this situation, where we see that not only the Le’-segs, but all the
segments in each of the two e’-bunches contribute to the intersected boundary at
the vicinity of the concave vertex u. We again may look into this problem in two
parts.
Part A. When B is relative supporting line at o. Let us assume that the line
segment B in our problem is a relative supporting line at the concave vertex u. It is
a typical situation that we have chosen in our example in Fig. 14a. Certainly, in this
case too, the Le’-segs, from each of the two bunches corresponding to the two
nonconvex edges incident on u, contribute to the boundary of the common intersec-
tion. But in addition to these two segments, “some parts” of the other segments of
the two bunches also contribute. But can we characterize those “some parts”? Since
S is placed at every point of B, the vertex u is the only contributing point at any
instant. And the sum total cont+bution is the locus or instances of u as S is placed
continuously at every point of B. And that locus is nothing but the line segment B
itself. This situation is clearly demonstrated in Fig. 14a.
Now if we follow the D-tracing technique, we shall also arrive at the same result.
The resulting D for our example and the enclosed positive region T (which
conforms to S 8 B) is shown in Fig. 15.
26 PIJUSH K. GHOSH

T= SOB
Tracing D

FIG. 15. The decomposition boundary tracing for the example given in Fig. 14a

Note. One significant difference between the nature of the tracings in the previous
convex-convex case and this simple-convex case must be noted. In the former only
the edges of S were effectively contributing to the boundary of the T region, WV&rile
in the latter we find an edge, say ab in Fig. 15, which appears to be an edge of B, is
also contributing. The edge ub is, in fact, the instances or the locus of the concave
vertex u.
Part B. when B is not a relative supporting line at u. If the line segment B is
not a supporting line at the concave vertex, the situation turns out to be much
simpler. The L-e’-seg for a nonconvex edge e’ of S becomes the “sole” contributing
factor in that particular direction, and the D-tracing technique does not need any
modification. A typical such situation is shown in Fig. 16.

$0
rx /:
‘.
v” ,’ I
‘\
‘\ v’,,’ I
;,

V V

I‘
62 Le, - seg
$
1,
El f-----+.-----
et ~
1
S S place on 6
(a) (b)

T=S@B

(d)

FIG. 16. The decomposition boundary tracing D when B is not a relative supporting liue at the
concave vertex v.
 A SOLUTION OF POLYGON CONTAINMENT 27

(using easy rule)

(b) (cl
FIG. 17. Decomposition of a simple polygon S by a convex polygon B.

From the above discussion we can, therefore, conclude that the D-tracing tech-
nique, described in Section 4.2.1.4, is applicable even if S is a simple polygon and B
is a line segment.
4.2.2.3. The generalization: Decomposition by a convex polygon B. Let us now
come back to our original problem. We claim, from the study of the above
subproblem, that the D-tracing technique does not need any modification even if B
is a convex polygon instead of a l@e segment. In the first place, recall the fact that
placing S on the boundary edges aB of B alone is sufficient, according to Theorem 1.
We can now argue in the following way. Since B is a convex polygon, the
technique of tracing the part of D corresponding to the convex parts (say, corre-
sponding to the edges e,, e2, etc. in Fig. 17) remains exactly similar to that of the
convex-convex case. Corresponding to the nonconvex parts (i.e., a concave vertex u
and its incident edges, say, es, e4 in fig. 17) the boundary of the common
intersection contains the Le’-segs from the nonconvex e’-bunches (i.e., Le,-seg,
L-e,-seg in Fig. 17), plus it contains the left-most part of the locus of the vertex u as
it moves along 86. The vertex u and thereby its locus comes in between those two
L.e’-segs. That essentially means that the edge(s) of i whose slope(s) in the reverse
direction falls in between the slopes of es and e4 must also be included in the
boundary of the common intersection, in addition to the Le,-seg and L.ecseg. Our
tracing technique also does exactly the same thing. (A different proof of this part is
given in Ghosh [3], by exploiting the fact that a nonconvex part of a polygon can be
described through the convex hole H. Therefore, its contribution to the boundary o,f
the common intersection can be determined from the boundary of the sum H 63 B.
To keep the paper self-complete the proof is not included since it uses the concept
of boundary tracing for Minkowski addition of polygons.) Once we form the D
polygon, we can determine the positive region T enclosed by D. T is the required
S 8 B.
4.2.2.4. Complexity analysis for the Algorithm A3. The simple-convex algorithm,
like the convex-convex case, has also two basic steps:
Step 1. Formation of D which is a self-crossing polygon.
Step 2. Determination of the positive region T enclosed by D.
28 PIJUSH K. GHOSH

NE-1 line segments

4
I’ \ (contributed by 6 )

I
(0 Le-seg contributed byS)

FIG. 18. A typical sickle.

Let us first find out the maximum number of edges in D. Let that number be N.
N = number of edges contributed by S + number of edges contributed by B
= N, + (number of vertices of S x max numbers of edges
of B contributing at any vertex)
= Ns + N, x ( NB - 1)
= N, - NB.
Next we have to find out the maximum number of self-intersections between
these N edges. To do this, we consider these N edges as a collection of Ns number
of sickles: each of these sickles consists of NB number_of line segments, of wh.icb
( NB - 1) segments are contributed by the edges of the B polygon (shown dotted in
Fig. 18) and the rest, say 1, is a Le-seg contributed by one of the edges of S (shown
boldface in Fig. 18).
Note that each such sickle is a conuex chain, since B is convex and the line
segment 1 is an anti-supporting line of B. Since we have N, such sickles, therefore,
the maximum number of self-intersections among them could be of the order of Ni.
If the edges of S and B are given in slope-sorted order, step 1 can be carried out
in O(N), i.e., in O(Ns + NB) time. Step 2 can be executed by using the method
suggested by Bently and Ottman [ll] for reporting all K intersections of an
arbitrary set of N line segments that takes O((N + K)log N) time. (More recently,
Chazelle [ll] reported an 0( N log2 N/loglog N + K) algorithm for the same
problem). Therefore, step 2 takes O(( Ni + NsN,)log N,N,) time in our case (or,
0( N,N, log’ N,N,/loglog N,N, + N:), if one chooses to use Chazelle’s method).
(Note that Algorithm A3 could be made more efficient by not forming the
complete D-polygon explicitly, but computing only those parts of D that may
contribute to the positive region T. For example, in case of the convex vertices of S,
the edges of i need not be traced explicitly as is shown in Al. However, we shall
not concern ourselves with those finer details in this paper.)
5. A UNIFIED APPROACH AND ITS DEMONSTRATIONS

5.1. The Approach

The previous discussions bring out the fact that, in determining S 8 B the major
task is to form the decomposition boundary tracing D. Once we are able to form the
D polygon, we can use any standard algorithm to compute the boundary of the
positive region T that is enclosed by D. T is the required S 0 B.
But does the D-tracing technique work for all kinds of polygons? If we can show
that it indeed works, then we arrive at a unified approach to solve all 2D
 A SOLUTION OF POLYGON CONTAINMENT 29

(a) A part of S polygon lb) Corresponding anti- (c) Formation of the

and rotating supporting supporting points tracing D
lines at 0

FIG. 19. Analysis of the decomposition boundary tracing D.

decompositions. For that purpose, we must analyse lirst the steps we follow in the
D-tracing technique.
First, any edge of a polygon can be characterized as the intersection of the
polygon and the supporting line of the polygon parallel to that edge. Therefore,
according to our notations, an edge of a polygon is the supporting points at the
polygon with respect to the supporting line parallel to that edge. (According to our
notion of oriented line, “parallel” means “similarly oriented” also.)
Now recall that a D polygon consists of Le-tegs corresponding to each edge e of
S, and r’-edges corresponding to the edges of B which keep continuity between two
adjacent Le-segs. Any Le-seg, we knoy, is the edge e of S translated to its
corresponding anti-supporting point at B (refer to Section 4.2.1.3). This is equiva-
lent of uecroriufb adding the edge e to its corresponding anti-supporting point at B.
In other words, this is equivale?t of vectorially adding the supporting points at S to
the anti-supporting point at B with respect to the supporting line parallel to the
edge e of S.
Second, in forming D, we move along the edges of 6 to reach from the
anti-supporting point for one edge of S to the anti-supporting point for the next
edge of S. This step is equivalent to assuming that the supporting line parallel to
one edge of S is continuously changing its direction (i.e., rotating continuously) to
finally become parallel to the next edge of S, and at every stage the supporting point
at S (it is, in fact, the vertex of S where the two edges meet) and the corresponding
anti-supporting points at B are being vectorially added. (This continuous movement
of the supporting line ensures continuity of the D polygon.) Figure 19 might clarify
the idea better. In Fig. 19a we-show how the supporting line is gradually rotating
from the direction of the edge e, to the edge e2. At all times the supporting point a,t
S remains fixed at its vertex u, but the corresponding anti-supporting point at B
(Fig. 19b) changes from c to a along the edge CLL For every direction of the
supporting line, we vecto@lly add the supporting point at S to the corresponding
anti-supporting point at B which finally forms an r’-edge (Fig. 19~);
(Our formulation of the exact rule for traversing the edges of the B polygon must
be clear to the readers at this point. The easy rule had also worked since we finally
discarded the negative regions enclosed by the D polygon,)
30 PIJUSH K. GHOSH

The above analysis could be summarized in the following form. The tracing D is
essentially the vector sum of the supporting points at S and the corresponding
anti-supporting points at i in every possible direction. But does the tracing D thus
formed ensure correct S 8 B in all situations? Yes, it does. Because for determining
S 8 B, we have to first identify, in every direction,” the left-most parts of the
instances of S when placed on the boundary points of B. The supporting point(s) at
S in any given direction is %ctually its boundary point(s) in” that direction. The
anti-supporting point(s) at B is the left-most point(s) of B in that direction.
Therefore, vector addition of the two will surely give us the left-most part of the
instances of S in that given direction.
We can now abstract out our approach to decomposition in terms of two
propositions stated below.
PROPOSITION 1 (supporting line proposition of 2D decomposition). Given two
plane $gures S and B in R2, the decomposition boundary tracing D of S 8 B is the set
of all points s + (;b) (i.e., vectorial addition of the points s and -b), such that
s E S and -b E B are the supporting point(s) and corresponding anti-supporting
point(s) respectively with respect to a supporting line in a given direction, and the
direction of the supporting line varies continuously from 0 to 360” covering all
directions in the plane.
Note that varying the direction of a supporting line from 0 to 360’ is equivalent
to varying it continuously along the boundary of S. However, varying it in the latter
way has certain computational advantages, and therefore, has been adopted by us.
PROPOSITION 2 (proposition on decomposition boundary tracing). In R2, S 8 B
is the positive region T enclosed by the decomposition bounakty tracing D of S 8 B.
(Note. Complete algebraic proofs of these two propositions can be obtained by
making use of Shephard’s theorem on Minkowski addition [4, 31 and the result
stated in Eq. (5).
We consider these two propositions as the fundamental algorithmic concepts of
2D Minkowski decomposition, since they state the underlying principle behind all
our algorithms for 2-dimensional decomposition. However, this does not mean that
the efficient algorithms for all kinds of S and B will be exactly the same. To make
an algorithm efficient, one must exploit the characteristics of the input figures.
Therefore, the actual computational procedure for decomposition varies depending
on the types of S and B. We already demonstrated, for example, that in the case of
the convex-convex decomposition, there is no need to form the decomposition
boundary tracing explicitly, although the underlying principle remains the same. In
the next section we present two more algorithms to establish this very point.
Moreover, these two algorithms will bring our study on 2-dimensional decomposi-
tion to a certain level of completion.
5.2. Two More Algorithms
5.2.1. Simple-Simple Case: Algorithm A4
Let S and B both be simple polygons as shown in Fig. 20a. the task is to compute
S 8 B by making use of the previous two propositions.
In tracing the D polygon we face no problem for the convex edges of S, where we
can assume decomposition by conv( B) instead by B (see Theorem 3). But the same
 A SOLUTION OF POLYGON CONTAINMENT 31

Lx3 0
k
3 '1

*--- FTJ
d
r
---- +a--

2j
(a)

r&edge

(b) Tracing D, (cl T,= SOB,

ri-edge
Id) Tracing D2 (e) T2=SOB2

r--F- -._+i-=j
r-------’ f
L.-.-.,.-. J I
(f) SOB =tSOB,) ncsoe,) (g) SOB

FIG. 20. Decomposition of a simple polygon by a simple polygon.

assumption is not applicable for the nonconvex edges of S. Now, since B is a simple
polygon in this case, it may happen that for some nonconvex edgeJs) of S, there may
be more than one “disconnected” anti-supporting point(s) at B. See our example
figure whert, for the supporting line parallel to the edge e4 of S, both the vertices a
and d of B are the corresponding anti-supporting points. Clearly, a and d are
disconnected points. According to Proposition 1 if we add { e4} @ {d, a } to form
D, the tracing D becomes disconnected at this point. Although it can be proved that
even such a disconnected tracing will eventually generate the correct S 8 B, it is
definitely preferable for various reasons to generate a completely connected D
polygon.
32 PIJUSH K. GHOSH

To achieve this aim, a number of remedies could be suggested. We suggest the

following method. (It must be mentioned that our technique does not guarantee the
best optimal algorithm.) The basic idea is to consider only one anti-supporting point
at each time, to perform decomposition operation as many times as there are
multiple anti-supporting points to determine the decomposed regions, and finally to
take the intersections of all those intermediary decomposed regions to obtain the
actual S 8 B.
In our example figure (Fig. 20a), we have already stated that the edge e4 has two
anti-supporting points a and d at B. Therefore, the decomposition operation has to
be performed twice. The first time we assume that e4 has only one anti-supporting
point, say a, and determine the tracing D, (Fig. 20b) and subsequently the
corresponding positive region T1 (Fig. 20~). The second time we choose d as the
only anti-supporting point of e4. The corresponding D2 and T2 are determined
(shown in Figs. 20d and e, respectively). Finally we take the intersection of T, and
T2 to obtain the required S 8 B.
The above procedure essentially means that we are assuming

B,=B’.d (consider B without the vertex d )

B,=B\a (consider B without the vertex a )

And we are claiming that

S 8 B = (S 0 B,) n (S 8 B,).

This is a valid claim, since for any general sets S, B,, and B, in Rd, the following
relation holds:

S 0 (B, u B2) = (S 8 B,) n (S 8 B2).

And we have chosen B, and B, in such a way that

B = B, u B,.

Note again that the algorithm stated above is only a conceptual model. In actual
practice, it may not be necessary to compute all the (S ~3 Bi) regions explicitly. It is
possible to determine beforehand, by examining the nature of the multiple support-
ing points, whether some of the (S 8 B;)‘s are proper subsets of the others.
However, we avoid going into those details in this paper.

5.2.2. Smooth-Smooth Case: Algorithm AS

The task of Algorithm A5 is to compute S 0 B when both S and B have smooth
boundary curves (Fig. 21).
Note that for a smooth boundary curve, the supporting line and the directed
tangent line become synonymous. Therefore, the task of the formation of D in this
case reduces to the following sequence of operations. For every point s of the
boundary of S, find the direction of the t?gent line and then determine the
corresponding point ( - b) at the boundary of B, where the tangent line is oppositely
 A SOLUTION OF POLYGON CONTAINMENT 33

-b

FIG. 21. Decomposition of two regions bounded by smooth curves.

directed; add s vectorially with (-b), i.e., s + (- b). The method is depicted in
Fig. 21b.
We may go further. It is often possible to obtain an algebraic solution of D in this
case. Let the boundary curves of S and B be represented respectively in the
parametric forms

aB= [B,(u),B,(u)],
as= [S,(U),
S,(U)],
where u and u are two scalar quantities and are closed intervals on the u-axis and
u-axis, respectively. We assume that both B and S are oriented in the sense
corresponding to an increase in the parameters u and u, respectively. Clearly,

ai = [kh), &.d] = [-Bx(u), -B,(u)].

Let s E S for :ome u, i.e., s = as(u). To obtain the corresponding anti-tangent

point (-b) at B, we have to solve the following equation for U:

[2$]=[2$]
Let the solution for u be u = f(u), where “f ” denotes some function. Therefore,
the analytical expression for the D-curve is

To obtain more details in this regard readers may refer to Ghosh [2].

6. A FEW CONCLUDING REMARKS

In this paper we have introduced three important concepts:
1. We established the direct relationships between the Minkowski operations
and the containment, spatial planning, and other related problems.
2. We reformulated Minkowski decomposition (as well as addition) problem,
which is intrinsically a geometric problem, as a set theoretic problem and then used
34 PIJUSH K. GHOSH

the set theoretic tools to reduce the computational complexity of the problem. This
set theoretic reformulation also allows us to go beyond convexity as well as beyond
2-dimensional.
3. We introduced the notion of decomposition boundary tracing that eventu-
ally helped us in proposing a unified theory for Minkowski decomposition algo-
rithms in R2.
Over and above those concepts, the paper also suggested (indirectly) a number of
new directions that seem promising for future research. We may explicitly list here a
few of them:
(a) The way we developed the theoretical framework and the 2D algorithms
clearly allows us to go beyond 2dimensional. First, the set theoretic results (Section
4.1) that simplify the geometric problem are applicable for general sets in Rd.
Second, Propositions 1 and 2, which are the basis for all the 2D algorithms can be
extended to higher dimensions simply by generalizing the concept of an oriented
supporting line. The general notion is that of a supporting hyperplane with an
orientation such that the direction of the outer normal of the supporting hyperplane
matches that of the corresponding surface of the object.
We face one problem here. The problem is in choosing the order of traversal of
the surfaces of S and B. In forming the D-polygon we traversed the edges of S and
B in such a way that at every stage of augmentation, the tracing is connected. The
choice of the traversal order was easy for us, since the vertices and the edges of a
planar polygon are intrinsically ordered. On the other hand, the vertices, edges, and
faces of a 3-dimensional polyhedron are not ordered accordingly. The problem
could be overcome by defining some order of traversal to maintain connectedness of
the tracing at each stage. Obviously, that will lead to more complex data structures
and algorithms.
(b) The problem was investigated by allowing only translation of the polygon
-but not rotation. It seems that efficient algorithms ahowing both translation and
rotation may be devised by formulating the problem completely in terms of
supporting points and decomposition boundary tracing. (Some investigations in this
direction by the author have indeed produced some encouraging results.)
(c) It also appears that the boundary tracing technique could be used effec-
tively in solving polygon union and intersection problems. We must note that some
work in this direction was once initiated by Guibas et al. [5].
(d) After living with the subject for a long time we strongly feel that the
significance of Minkowski operators is tremendous and need further investigation.
Not only their importance is being felt in solving spatial planning, containment, and
other related problems, but also in the areas like image processing, geometric
modelling, biological form description and analysis, graphic arts and typesetting,
etc. [13, 3, 7, 21. Besides their impact on classical problems, the efficient computation
of Minkowski operations give rise to theoretical and algorithmic problems of their
own and can significantly enrich the discipline of computational geometry.
ACKNOWLEDGMENTS
I would like to express my gratitude to Professor R. Narasimhan, Dr. S. Ramani, Dr. S. P. Mudur, Dr.
S. K. Ghosh, Ms. L. Parida, Mr. T. M. Vijayaraman, Mr. P. R. Chandrasekban, and particularly to Mr.
 A SOLUTION OF POLYGON CONTAINMENT 35

P. Sadanandan for their help at various stages of preparation of the manuscript. And, finally, I thank the
referees of this paper for constructive suggestions.

REFERENCES
1. B. S. Baker et ul., Polygon containment under translation, J. Algorithms 7, No. 4, 1986.
2. P. K. Ghosh and S. P. Mudur, The brush-trajectory approach to figure specification: Some
algebraic-solutions, ACM Trans. Graphics 3,1984.
3. P. K. Ghosh, Computational Theoretic Framework for Shape Representation and Analysis Using
Minkowski Addition and Decomposition Operators, Ph.D. thesis, Tata Institute of Fundamental
Research, Bombay, September 1986.
4. B. Grunbaum, Convex Polytopes, Interscience, London, 1967.
5. L. J. Guibas et al., A kinetic framework for computational geometry, IEEE 24th Annual Symposium
on Foundations of Computer Science, 1983.
6. R. M. Haralick, et al., Image analysis using mathematical morphology, IEEE Pattern Anal. Mach.
Intell. PAMI-9, No. 4, 1987.
7. D. E. Knuth, Mathematical typography, Bull. Amer. Math. Sot. 1,1979.
8. D. T. Lee and F. P. Preparata, An improved algorithm for the rectangle enclosure problem,
J. Algorithms 3, No. 3, 1982.
9. T. Lozano-Perez, Spatial planning: A configuration space approach, IEEE Trans. Comput. C-32, No.
2, 1983.
10. G. Matheron, Random Sets and Integral Geometry, Wiley, New York, 1975.
11. F. P. Preparata and M. I. Shamos, Computational Geometry: An Introduction, Springer-Verlag, New
York, 1985.
12. J. Terra, Image Analysis and Mathematical Morphology, Academic Press, London, 1982.
13. J. Serra, Introduction to mathematical morphology, Comput. Vision Graphics Image Process. 35, No.
3, 1986.
14. J. T. Schwartz, Finding the minimum distance between two convex polygons, Inform. Process. L,ett.,
13, No. 4, 1981.
15. V. Vaishnavi and D. Wood, Data structures for the rectangle containment and enclosure problems,
Comput. Graphics Image Process. 13, 1980.