See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3193582

Complete Solution Classification for the Perspective-Three-Point Problem

Article in IEEE Transactions on Pattern Analysis and Machine Intelligence · September 2003
DOI: 10.1109/TPAMI.2003.1217599 · Source: IEEE Xplore

CITATIONS READS
885 4,860

4 authors, including:

Xiao-Shan Gao
Chinese Academy of Sciences
219 PUBLICATIONS 4,302 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Difference Algebra View project

Symbolic computation in information theory View project

All content following this page was uploaded by Xiao-Shan Gao on 03 August 2013.

The user has requested enhancement of the downloaded file.

930 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO. 8, AUGUST 2003

Complete Solution Classification for the

Perspective-Three-Point Problem
Xiao-Shan Gao, Member, IEEE, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng

Abstract—In this paper, we use two approaches to solve the Perspective-Three-Point (P3P) problem: the algebraic approach and the
geometric approach. In the algebraic approach, we use Wu-Ritt’s zero decomposition algorithm to give a complete triangular
decomposition for the P3P equation system. This decomposition provides the first complete analytical solution to the P 3P problem. We
also give a complete solution classification for the P3P equation system, i.e., we give explicit criteria for the P3P problem to have one, two,
three, and four solutions. Combining the analytical solutions with the criteria, we provide an algorithm, CASSC, which may be used to find
complete and robust numerical solutions to the P3P problem. In the geometric approach, we give some pure geometric criteria for the
number of real physical solutions.

Index Terms—Perspective-Three-Point problem, pose determination, analytical solutions, solution classification, geometric criteria,
Wu-Ritt’s zero decomposition method.

1 INTRODUCTION

T Perspective-n-Point (PnP) problem is originated from

HE
camera calibration [1], [2], [3], [4]. Also known as pose
estimation, it is to determine the position and orientation of
Bolles [2] presented the RANSAC algorithm. They have
noticed that there are at most four possible solutions to the
P3P equation system. Hung et al. [8] presented an algorithm
the camera with respect to a scene object from n correspon- for computing the 3D coordinates of the perspective center
dent points. It concerns many important fields such as relative to the camera frame. In 1991, Haralick et al. [9]
computer animation [5], computer vision [3], automation, reviewed the major direct solutions up to 1991, including
image analysis, and automated cartography [2], photogram- six algorithms given by Grunert (1841), Finsterwalder
metry [6], robotics [1], and model-based machine vision (1903), Merritt (1949), Fischler and Bolles (1981), Linnain-
system [7], etc. Fischler and Bolles [2] summarized the maa et al. (1988), and Grafarrend et al. (1989), respectively.
problem as follows: They also give the analytical solution for the P3P problem
“Given the relative spatial locations of n control points, and with resultant computation. DeMenthon and Davis [10],
given the angle to every pair of control points from an [11] showed that by using approximations to the perspec-
additional point called the Center of Perspective (CP ), find tive, simpler computational solutions can be obtained.
the lengths of the line segments joining CP to each of the Quan and Lan [4] reduced the problem to a new quartic
control points.”
equation with Sylvester resultant and proposed a linear
The study of the P nP problem mainly consists of two aspects: algebra algorithm to solve the PnP problem.
One of the important research directions on the
1.Design fast and stable algorithms that can be used to P 3P problem is its multisolution phenomenon. Fischler
find all or some of the solutions of the P nP problem. and Bolles [2] presented some examples of multisolutions of
2. Give a classification for the solutions of the the P3P problem. In 1986, Wolfe [7] pointed out that the six
P nP problem, i.e., give the conditions under which permutations of the three control points combined with
the problem has one, two, three or four solutions. four-solution possibility can produce 24 possible camera-
There are many results for the first problem and the second triangle configurations consistent with a single perspective
problem is still open. The aim of this paper is to give a view [6], [7]. Yuan [6] gave a necessary condition for the
complete and effective solution to the above two problems existence of the solution for first time. In 1991, Wolfe and
for the P3P problem. Jones [12] gave a geometric explanation to this multisolution
The P3P problem is the smallest subset of control points phenomenon in the image plane under the assumption of
that yields a finite number of solutions. In 1981, Fischler and “canonical view.”
In 1997, Su et al. [5] applied Wu-Ritt’s zero decomposi-
tion method to find the main solution branch and some
. X.-S. Gao and J. Tang are with the Institute of System Science, AMSS nondegenerate branches for the P3P problem. But a
Academia Sinica, Beijing 100080, China.
complete decomposition was not given. In [13], they used
E-mail: {xgao, jtang}@mmrc.iss.ac.cn.
. X.-R. Hou is with the Institute of Computational Mathematics, NingBo the Sturm sequence to give some conditions to adjudicate
University Ningbo, 315211, China. E-mail: [email protected]. the number of solutions. In 1998, Yang [14] gave partial
. H.-F. Cheng is with the Department of Mathematics, Pennsylvania State solution classifications of the P3P problem under some
University, University Park, PA 16802.
nondegenerate conditions.
Manuscript received 6 July 2001; revised 29 Apr. 2002; accepted 5 Oct. 2002.
Recommended for acceptance by R. Kumar.
The P3P problem is the most basic case of the
For information on obtaining reprints of this article, please send e-mail to: PnP problems. All other PnP (n > 3) problems include the
[email protected], and reference IEEECS Log Number 114489. P3P problem as a special case. Therefore, a complete study of
0162-8828/03/$17.00 ß 2003 IEEE Published by the IEEE Computer Society
GAO ET AL.: COMPLETE SOLUTION CLASSIFICATION FOR THE PERSPECTIVE-THREE-POINT PROBLEM 931

this problem is desirable. This paper is an effort toward this

goal. We use two approaches to solve the P3P problem: the
algebraic approach and the geometric approach. In the
algebraic approach, we apply Wu-Ritt’s zero decomposition
algorithm [15], [16], [17] to find a complete zero decomposi-
tion for the P3P equation system. This decomposition
provides the first complete analytical solution to the
P3P problem. Based on this decomposition, we give a
complete solution classification to the P3P equation system
for the first time, i.e., we give explicit criteria for the
P3P problem to have one, two, three, or four solutions. The Fig. 1. The P3P problem.
procedure of obtaining this classification consists of the most
difficult part of this paper. With these criteria, we introduce A set of solutions for X; Y ; Z is called a set of physical
the concept of stable and critical values for the input solutions if the following “reality conditions” are satisfied.
parametric values. If a set of values is stable, then a small These conditions are assumed through out the paper.
variation of these values will give the same number of 8
solutions. Therefore, for a set of stable values, we may use the >
> X > 0; Y > 0; Z > 0; a0 > 0; b0 > 0; c0 > 0; a0 þ b0 > c0 ; a0 þ c0 > b0 ; b0 þc0 > a0
<
0 < ; ; < ; 0 < þ þ < 2 
usual floating-point calculations to enhance the computation ð2Þ
speed; and for a set of critical values, we may use high- >
> þ > ; þ > ; þ >
: 2 2 2
I0 ¼ p þ q þ r ÿ pqr ÿ 1 6¼ 0 ðPoints P ;A;B;C are not coÿplanar ½5ŠÞ:
precision computation tools [18] to provide more robust
computation. To simplify the equation system, let
pffiffi pffiffiffiffiffi pffiffiffiffi X ¼ xZ;
Combining the analytical solutions with the criteria, we Y ¼ yZ; jABj ¼ vZ; jBCj ¼ avZ; jACj ¼ bvZ. Since
provide an algorithm, CASSC (Complete Analytical Solution Z ¼ jP Cj 6¼ 0, we obtain the following equivalent equa-
with the assistance of Solution Classification), which can be tion system:
used to find complete and robust numerical solutions to the 8
P3P problem. Our experimental results support this assertion. < y2 þ 1 ÿ yp ÿ av ¼ 0
In the geometric approach, we consider the three x2 þ 1 ÿ xq ÿ bv ¼ 0 ð3Þ
: 2
perspective angles separately. Then, the locus of the center x þ y2 ÿ xyr ÿ v ¼ 0:
of perspective point for each angle is a toroid and the center of Since jrj < 2, we have v ¼ x2 þ y2 ÿ p
perspective is the intersection of three toroids. In this way, we ffiffi > 0. Thus, Z can be
xyr
uniquely determined by Z ¼ jABj= v. Eliminating v from
give some pure geometric criteria for the number of solutions (3), we have
of the P3P problem. One interesting result is “the P3P problem 
can have only one solution if the three perspective angles are p1 ¼ ð1 ÿ aÞy2 ÿ ax2 ÿ py þ arxy þ 1 ¼ 0
ðESÞ
obtuse.” This kind of criteria is much simpler than the p2 ¼ ð1 ÿ bÞx2 ÿ by2 ÿ qx þ brxy þ 1 ¼ 0
algebraic one and gives some insight into the multisolution which has the same number of physical solutions with (1).
phenomenon. On the other hand, since the field of view of Now, the P3P problem is reduced to finding the positive
most cameras is much less than 90 degrees, this result does solutions of two quadratic equations. As a consequence, we
not have much practical value. In any case, to find the obtain the following result: The P3P problem has either an
solutions we must use the algebraic computation approach. infinite number of solutions or at most four physical solutions. This
The rest of the paper is organized as follows: In Section 2,
result was known before only for the “main part” of the
we present the zero decomposition for the P3P equation
P3P problem.
system. In Section 3, we present the solution classification.
In Section 4, we present the CASSC algorithm and the 2.2 Zero Structure for the P3P Equation System
experimental results. In Section 5, we present the geometric Wu-Ritt’s zero decomposition method [15], [16], [17] is a
approach. In Section 6, we present the conclusions. general method to solve systems of algebraic equations. It
may be used to represent the zero set of a polynomial
equation system as the union of zero sets of equations in
2 ZERO STRUCTURE FOR THE P3P EQUATION triangular form, that is, equation systems like
SYSTEM
2.1 Simplification of the Equation System f1 ðu; x1 Þ ¼ 0; f2 ðu; x1 ; x2 Þ ¼ 0; . . . ; fp ðu; x1 ; . . . ; xp Þ ¼ 0;
Let P be the Center of Perspective, and A; B; C the control where the u could be considered as a set of parameters
points. Let jP Aj ¼ X; jP Bj ¼ Y ; jP Cj ¼ Z; ¼ ffBP C; and the x are the variables to be determined. As shown in
¼ ffAP C; ¼ ffAP B; p ¼ 2 cos ; q ¼ 2 cos ; r ¼ 2 cos ; [15], solutions for an equation system in triangular form
jABj ¼ c0 ; jBCj ¼ a0 ; jACj ¼ b0 . From triangles P BC, P AC, are well-determined. For instance, the solution of an
and P AB, we obtain the P3P equation system (Fig. 1): equation system in triangular form can be easily reduced
8 2 to the solution of univariate equations. For a polynomial
< Y 2 þ Z 2 ÿ Y Zp ÿ a0 ¼ 0
2 set P S and a polynomial I, let ZeroðP SÞ be the set
Z 2 þ X2 ÿ XZq ÿ b0 ¼ 0 ð1Þ
: 2 2 of solutions of the equation system P S ¼ 0, and Zero
X þ Y 2 ÿ XY r ÿ c0 ¼ 0:
ðP S=IÞ ¼ ZeroðP SÞ ÿ ZeroðIÞ.
932 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO. 8, AUGUST 2003

Among the “reality conditions” listed in (2), I0 6¼ 0 could TABLE 1

be used to simplify the computation. Therefore, we consider The Maximal Number of Solutions for Each Component
ZeroðES=I0 Þ. Using Wu-Ritt’s zero decomposition method
[15], we decompose ZeroðES=I0 Þ into 10 disjoint components:

[
10
ZeroðES=I0 Þ ¼ Ci : ðDEÞ
i¼1

In the above formula, Ci ¼ ZeroðT Si =Ti Þ; i ¼ 1;


; 9 and 3. From the experimental results in Section 4, we can see
C10 ¼ ZeroðT S10 =T10 Þ [ ZeroðT S11 =T11 Þ, where T Si are poly- that the above decomposition provides a complete and
nomial equations in triangular form and Ti are polynomials. robust way to find the solutions to the P3P problem.
T Si and Ti may be found in Appendix A.
Among the 10 components, ZeroðT S1 =T1 Þ is called the 3 COMPLETE SOLUTION CLASSIFICATION FOR THE
main component of the P3P equation system, which is of the P3P EQUATION SYSTEM
following form: For polynomials fðxÞ and gðxÞ, let Vf be the number of
 real solutions of fðxÞ, and Vf ðg > 0Þ the number of real
f ¼ a0 x4 þ a1 x3 þ a2 x2 þ a3 x þ a4 ¼ 0
ðTS 1 Þ solutions of fðxÞ such that g > 0. If fðxÞ has n real
g ¼ b0 y ÿ b1 ¼ 0: ðn;jÞ
solutions, then let Cf ðg > 0Þ denote the conditions that
The coefficients ai and bj may be found in Appendix A. This make fðxÞ having j real solutions such that g > 0. The
component has been obtained in [2], [4], [5], [9]. All other following lemmas will be used in this section.
components could be considered as degenerate cases. This is
Lemma 1 (Descartes’ Rule of Sign [19], [16]). Let f ¼
because, in these cases, the parameters a; b; p; q; r must Pn i
i¼k ai x ; ðan ak 6¼ 0Þ be a polynomial with real numbers as
satisfy certain algebraic relations. In other words, we are
coefficients. Then, the number of positive roots of f is less than
considering special cases of the problem.
or congruent to the number of sign changes in the sequence of
Compared to the main component, the “degenerate” cases
coefficients an ; . . . ; ak ð mod 2Þ.
are less possible to occur. But, they are still important due to
the following reasons: Solutions satisfying some degenerate Lemma 2 [16]. Let fðxÞ, gðxÞ be two polynomials and f of degree
n. Let
conditions, such as a þ b ÿ 1 ¼ 0 (in T S4 ) meaning that ffACB
is a right angle, may occur quite often if points A; B; C are from Y
n

man-made structures like buildings, where many right angles rðT Þ ¼ resultantðf; g ÿ T ; xÞ ¼ c ðgðxi Þ ÿ T Þ;
i¼1
exist. In the general case, the degenerate cases, such as T S2 ,
could be complicated and have no clear geometric meaning. where xi ; i ¼ 1; . . . ; n are the roots of fðxÞ ¼ 0. If all the
Therefore, it is difficult to tell when it will occur. solutions of fðxÞ ¼ 0 are real, then Vf ðg > 0Þ ¼ Vr ðT > 0Þ.
Table 1 gives the maximal number of solutions for each
component. Let fi ðxÞ; gi ðx; yÞ be the first two polynomials in T Si .
Since Ci and Cj ði 6¼ jÞ are disjoint, to solve the equation In components T Si ; i ¼ 5; 6; 7; 10, fi ðxÞ, and gi ðx; yÞ are
numerically, for a set of specific values of the parameters linear in x and y, respectively. In these cases, each component
p; q; r; a; b, either can have only one positive solution and it is trivial to give the
conditions for them to have such solutions.
ZeroðP S=I0 Þ ¼ ZeroðT Sk =Tk Þ In components T Si ; i ¼ 3; 4; 8; 9; 11, fi ðxÞ and gi ðx; yÞ are
if k satisfying 1  k  9 and Tk ðp; q; r; a; bÞ 6¼ 0, or either linear or quadratic in x and y, respectively. We will
treat these cases in Section 3.1.
ZeroðP S=I0 Þ ¼ ZeroðT S10 =T10 Þ [ ZeroðT S11 =T11 Þ In component T S2 , f2 ðxÞ is a cubic equation and g2 ðx; yÞ
is linear in y. We will treat this case in Section 3.2.
if T10 ðp; q; r; a; bÞ ¼ T11 ðp; q; r; a; bÞ 6¼ 0. Since the polynomials
In component T S1 , f1 ðxÞ is a quartic equation and g1 ðx; yÞ
in T Si are of degree  4, the solution of the P3P problem is
is linear in y. We will treat this case in Section 3.3.
reduced to the solution of equation systems in triangular
form and, hence, to the solution of univariate equations of 3.1 The Quadratic Cases
degree  4. The quadratic cases may have three forms: 1) fi ðxÞ is
From this decomposition, we have the following quadratic and gi ðx; yÞ is linear in y, 2) fi ðxÞ is linear and
observations: gi ðx; yÞ is quadratic in y, and 3) fi ðxÞ and gi ðx; yÞ are
quadratic in x and y, respectively. All of them can be treated
1. Since the solutions for each triangular set are well-
similarly. We will take T S9 as an illustrative example. The
determined, this decomposition provides a complete
set of analytical solutions for the P3P problem. equation system is
2. From Table 1 and the analysis following the table, it is 8
< f9 ¼ ðÿ1 þ a þ bÞx2 þ ðÿqa þ qÞx ÿ 1 þ a ÿ b ¼ 0;
easy to see that there are at most four distinct solutions
g ¼ ðÿ1 þ a þ bÞy2 ÿ 1 ÿ a þ qxa þ b ¼ 0; ð4Þ
under the reality condition (2). Notice that this result : 9
p ¼ 0; r ¼ 0; x > 0; y > 0; a þ b ÿ 1 6¼ 0:
was proven previously only for the main component.
GAO ET AL.: COMPLETE SOLUTION CLASSIFICATION FOR THE PERSPECTIVE-THREE-POINT PROBLEM 933

The number of solutions for the above system is the same as 1. Equation (4) has two physical solutions iff one of the
the following equation system: following statements holds:
8
< f9 ¼ ðÿ1 þ a þ bÞx2 þ ðÿqa þ qÞx ÿ 1 þ a ÿ b ¼ 0; 1:1: ½ p ¼ r ¼ 0; a þ b ÿ 1 > 0; q > 0; D13 > 0;
1 < 0;
2 < 0;
3 < 0Š
g ¼ ðÿ1 þ a þ bÞð1 þ a ÿ qxa ÿ bÞ > 0; ð5Þ 1:2: ½ p ¼ r ¼ 0; a þ b ÿ 1 < 0; q > 0; D13 > 0;
1 > 0;
2 < 0;
3 > 0Š:
:
p ¼ 0; r ¼ 0; a þ b ÿ 1 6¼ 0; x > 0:
We first assume that 2. Equation (4) has one physical solution iff one of the
following statements holds:
resultantðf9 ; g; xÞ 6¼ 0;
2:1: ½ p ¼ r ¼ 0; a ¼ 1; q þ b ÿ 2 < 0Š

resultantðf9 ; x; xÞ ¼ a ÿ b ÿ 1 6¼ 0; 2:2: ½ p ¼ r ¼ 0; a ÿ b þ 1 ¼ 0; q 6¼ 0Š

and 2:3: ½ p ¼ r ¼ 0; a ÿ b ÿ 1 ¼ 0; q > 0; q2 a < 4Š

2:4: ½ p ¼ r ¼ 0; a þ b ÿ 1 > 0; D13 ¼ 0;
1 < 0Š

resultantðg; x; xÞ ¼ ða þ b ÿ 1Þða ÿ b þ 1Þ 6¼ 0:
2:5: ½ p ¼ r ¼ 0; a þ b ÿ 1 < 0; D13 ¼ 0;
1 > 0Š
Let 2:6: ½ p ¼ r ¼ 0; a þ b ÿ 1 > 0;
2 ¼ 0;
1 < 0Š

r11 ðT Þ ¼ resultantðf9 ; g ÿ T ; xÞ ¼ r110 T 2 þ r111 T þ r112 ; 2:7: ½ p ¼ r ¼ 0; a þ b ÿ1 < 0;
2 ¼ 0;
1 > 0Š

2:8: ½ p ¼ r ¼ q ¼ 0; ða þ b ÿ 1Þ ða ÿ b ÿ 1Þ < 0; ða þ b ÿ 1Þ ða ÿ b þ 1Þ > 0Š

r12 ðT Þ ¼ resultantðf9 ; xg ÿ T ; xÞ ¼ r120 T 2 þ r121 T þ r122 : 2:9: ½ p ¼ r ¼ 0; q > 0; a ÿ b ÿ 1 > 0 ;D13 > 0;
1 < 0;
2 > 0;
3 6¼ 0Š

2:10: ½ p ¼ r ¼ 0; q > 0; a ÿ b ÿ 1 > 0; D13 > 0;
1 > 0;
2 < 0;
3 6¼ 0Š

By computation, we obtain the Sylvester-Habicht sequences
[21] of ðf9 ; diffðf9 ; xÞÞ which are denoted by D11 , D12 , and 2:11: ½ p ¼ r ¼ 0; q > 0; a þ b ÿ 1 > 0; D13 > 0;
1 < 0;
2 < 0;
3 6¼ 0Š

D13 (the discriminant), respectively. So, f9 has two real 2:12: ½ p ¼ r ¼ 0; q > 0; a þ b ÿ 1 > 0; D13 > 0;
1 > 0;
2 < 0;
3 6¼ 0Š;

solutions iff D13 > 0 and one real solution iff D13 ¼ 0. where
Since resultantðf9 ; g; xÞ 6¼ 0, resultantðf9 ; x; xÞ 6¼ 0, and
resultantðg; x; xÞÞ 6¼ 0, we have D13 ¼ q2 ða ÿ 1Þ2 ÿ 4ða þ b ÿ 1Þða ÿ b ÿ 1Þ;
8
1 ¼ q2 aða ÿ 1Þ ÿ 2ða ÿ b þ 1Þða þ b ÿ 1Þ;
>
> Vf ¼ Vf9 ðx > 0Þ þ Vf9 ðx < 0Þ;
< 9
2 ¼ q2 a ÿ ða ÿ b þ 1Þ2 ;
Vf9 ðx < 0Þ ¼ Vf9 ðx < 0; g > 0Þ þ Vf9 ðx < 0; g < 0Þ;
ð6Þ
>
> V ðg > 0Þ ¼ Vf9 ðx > 0; g > 0Þ þ Vf9 ðx < 0; g > 0Þ;
3 ¼ q2 aða ÿ 1Þ2 ÿ ða þ b ÿ 1Þðb ÿ 3ab ÿ 1 ÿ 2a þ 3a2 Þ:
: f9
Vf9 ðxg > 0Þ ¼ Vf9 ðx > 0; g > 0Þ þ Vf9 ðx < 0; g < 0Þ:
From the above equations, we obtain the following formula Proof. We know that (4) has two physical solutions iff
for the number of physical solutions: Vf9 ðx > 0; g > 0Þ ¼ 2. From (7), this is possible iff

Vf9 ðx > 0; g > 0Þ ¼ Vf9 ðx > 0Þ ¼ 2; Vf9 ðg > 0Þ ¼ 2; Vf9 ðxg > 0Þ ¼ 2; Vf9 ¼ 2:
1ÿ
ð7Þ
Vf9 ðx > 0Þ þ Vf9 ðg > 0Þ þ Vf9 ðxg > 0Þ ÿ Vf9 : The first part of the Theorem 3 follows directly from these
2 conditions. For the second part, there are four cases.
By Lemmas 1 and 2, we have the following results: Suppose resultantðf9 ; g; xÞ 6¼ 0, resultantðf9 ; x; xÞ 6¼ 0,
Vf9 ðx > 0Þ ¼ the number of sign changes in coefficients of f9 ðxÞ; and resultantðg; x; xÞ 6¼ 0. In this case, (4) has one physical
Vf9 ðg > 0Þ ¼ the number of sign changes in coefficients of r11 ðT Þ; solution iff Vf9 ðx > 0; g > 0Þ ¼ 1. From (7), (4) has one
Vf9 ðxg > 0Þ ¼ the number of sign changes in coefficients of r12 ðT Þ: physical solution iff one of the following conditions holds:

If resultantðf9 ; g; xÞ ¼ 0, (5) becomes: Vf9 ðx > 0Þ ¼ 2; Vf9 ðg > 0Þ ¼ 2; Vf9 ðxg > 0Þ ¼ 0; Vf9 ¼ 2;
8 Vf9 ðx > 0Þ ¼ 2; Vf9 ðg > 0Þ ¼ 1; Vf9 ðxg > 0Þ ¼ 1; Vf9 ¼ 2;
>
> f ¼ ÿaqða þ b ÿ 1Þx ÿ q2 a þ q2 a2 ÿ 2b ÿ a2 þ 1 þ b2 ¼ 0;
<9 Vf9 ðx > 0Þ ¼ 2; Vf9 ðg > 0Þ ¼ 0; Vf9 ðxg > 0Þ ¼ 2; Vf9 ¼ 2;
g ¼ ðÿ1 þ a þ bÞð1 þ a ÿ qxa ÿ bÞ > 0;
>
> p ¼ 0; r ¼ 0; Vf9 ðx > 0Þ ¼ 1; Vf9 ðg > 0Þ ¼ 2; Vf9 ðxg > 0Þ ¼ 1; Vf9 ¼ 2;
:
a þ b ÿ 1 6¼ 0; x > 0 Vf9 ðx > 0Þ ¼ 1; Vf9 ðg > 0Þ ¼ 1; Vf9 ðxg > 0Þ ¼ 2; Vf9 ¼ 2;
ð8Þ Vf9 ðx > 0Þ ¼ 0; Vf9 ðg > 0Þ ¼ 2; Vf9 ðxg > 0Þ ¼ 2; Vf9 ¼ 2;
which may have one positive solution and the condition for Vf9 ðx > 0Þ ¼ 1; Vf9 ðg > 0Þ ¼ 1; Vf9 ðxg > 0Þ ¼ 1; Vf9 ¼ 1:
that is easy to obtain. For resultantðf9 ; x; xÞ ¼ a ÿ b ÿ 1 ¼ 0 Analyzing these conditions will lead to some of the
and resultantðg; x; xÞ ¼ ða þ b ÿ 1Þða ÿ b þ 1Þ ¼ 0, we can conditions in part two of Theorem 3. The other three
deal with them similarly.
cases are resultantðf9 ; g; xÞ ¼ 0, resultantðf9 ; x; xÞ ¼ 0,
Theorem 3. We have the following necessary and sufficient and resultantðg; x; xÞ ¼ 0, respectively. We can deal with
conditions to adjudicate the number of physical solutions them similarly. Combining the four cases will lead to the
Vf9 ðx > 0; g > 0Þ of ZeroðT S9 =T9 Þ. second part of Theorem 3. u
t
934 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO. 8, AUGUST 2003

h   i
3.2 The Cubic Case q21 ÿq21
3:1: D8 ¼ 0; D7 6¼ 0; < 0; g >0
For T S2 , we need to consider the polynomial system: q20 q20
h   i
8 h31 ÿh31
3:2: D8 ¼ 0; D7 6¼ 0; < 0; g >0
< f2 ¼ a5 x3 þ a6 x2 þ a7 x þ a8 ; h30 h30
g ¼ b2 y ÿ b3 ¼ 0; ð9Þ h   i
: 2 a2 ÿa2
x > 0; y > 0; a5 6¼ 0; b2 6¼ 0; 3:3: D8 ¼ 0; D7 ¼ 0; D6 6¼ 0; < 0; g >0
a1 3a1
h i
where the ai and bj could be found in Appendix A. The ð1;1Þ ð1;1Þ ð1;1Þ
3:4: D8 < 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ
number of solutions for the above system is the same as the h i
ð3;3Þ ð3;1Þ ð3;1Þ
following equation system: 3:5: D8 > 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ
h i
8 ð3;2Þ ð3;2Þ ð3;1Þ
< f2 ¼ a5 x3 þ a6 x2 þ a7 x þ a8 ; 3:6: D8 > 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ
g ¼ b2 b3 > 0 ð10Þ h i
ð3;2Þ ð3;1Þ ð3;2Þ
: 3:7: D8 > 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ
x > 0; a5 6¼ 0:
h i
ð3;1Þ ð3;3Þ ð3;1Þ
Let Resultantðf2 ; g; xÞ 6¼ 0, and Di the Sylvester-Habicht 3:8: D8 > 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ
h i
sequences of ðf2 ; diffðf2 ÞÞ, i ¼ 5; . . . ; 8 [21]. We know that ð3;1Þ ð3;2Þ ð3;2Þ
3:9: D8 > 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ
fðxÞ has three real solutions iff D8 > 0 [20]. Let h i
ð3;1Þ ð3;1Þ ð3;3Þ
3:10: D8 > 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ :
r7 ðT Þ ¼ resultantðf2 ; g ÿ T ; xÞ ¼ r70 T 3 þ r71 T 2 þ r72 T þ r73 ;
ðn;jÞ
The explicit expressions for all Cf may be found in
r8 ðT Þ ¼ resultantðf2 ; xg ÿ T ; xÞ ¼ r80 T 3 þ r81 T 2 þ r82 T þ r83 :
Appendix B.
If D8 > 0, then by Lemmas 1 and 2, we can give the conditions If Resultantðf2 ; g; xÞ ¼ 0, the problem becomes a quad-
for the equation system to have one, two, or three positive ratic case and can be treated similarly as in the preceding
solutions. If D8 < 0, f2 ðxÞ has only one real solution and two section.
complex solutions. By Descartes rule of sign, Vf2 ðx > 0Þ ¼ 1 3.3 The Quartic Case
iff the number of sign changes in coefficients of f2 ðxÞ is 1 or 3.
For T S1 , we need to count the number of solutions for the
Now, we consider the number of positive solutions of r7 ðT Þ
following system.
and r8 ðT Þ. Let x1 be real, x2 , x3 be complex. If gðx2 Þ is real, then 8
gðx3 Þ is also real, and the signs of gðx2 Þ and gðx3 Þ are the same. < f1 ¼ a0 x4 þ a1 x3 þ a2 x2 þ a3 x þ a4 ¼ 0;
So, by the Descartes rule of sign, Vrj ðT > 0Þ ¼ 1 iff the number g ¼ b0 y ÿ b1 ¼ 0; ð11Þ
: 1
x > 0; y > 0; a0 6¼ 0; b0 6¼ 0;
of sign changes in coefficients of rj ðxÞ is 1 or 3, j ¼ 7; 8: If
D8 ¼ 0, D7 6¼ 0, then there exists polynomial Q2 ðxÞ ¼ where ai and bi may be found in Appendix A. Since b0  0
q20 x þ q21 such that f2 ðxÞ ¼ cQ2 ðxÞH32 ðxÞ, where H3 ðxÞ ¼ (see Appendix A), it is equivalent to count the number of
h30 x þ h31 is the pseudoremainder of f2 with diffðf2 Þ for positive solutions for the following system
variable x. This case is easy to solve. If D8 ¼ 0, D7 ¼ 0, D6 6¼ 0, 8
< f1 ¼ a0 x4 þ a1 x3 þ a2 x2 þ a3 x þ a4
then f2 ðxÞ ¼ cH43 ðxÞ, where H4 ðxÞ ¼ 3a5 x þ a6 . b >0 ð12Þ
From the above discussion, we obtain the following result. : 1
x > 0; a0 6¼ 0; b0 6¼ 0:
Theorem 4. For T S2 , we have the following results:
We will first assume that Resultantðf1 ; b1 ; xÞ 6¼ 0 and
1. Equation (9) has three physical solutions iff the Resultantðf1 ðxÞ; x; xÞ 6¼ 0: Let f 0 ¼ Diffðf1 ; xÞ, and Di ;
following statements hold: Ai ; Bi ; Ci the Sylvester-Habicht sequences of ðf1 ; f 0 Þ,
h i
ð3;3Þ ð3;3Þ ð3;3Þ ðf1 ; xf 0 Þ, ðf1 ðxÞ; b1 f 0 Þ, and ðf1 ðxÞ; xb1 f 0 Þ, respectively, i ¼
D8 > 0; Cf2 ðx > 0Þ; Cf2 ðg > 0Þ; Cf2 ðxg > 0Þ :
0; ::; 4 [21]. First, let us assume D4 6¼ 0. It is known that
Vf1 ¼ 4 iff D2 > 0, D3 > 0, D4 > 0 which is denoted by
2. Equation (9) has two physical solutions iff one of the ð4Þ
Cf1 [20]. Let
following statements holds:
 ð3;3Þ ð3;2Þ ð3;2Þ
 Y
4
2:1: D8 > 0; Cf
2
ðx > 0Þ; Cf
2
ðg > 0Þ; Cf
2
ðxg > 0Þ r1 ðT Þ ¼ resultantðf1 ; b1 ÿ T ; xÞ ¼ a40 ðb1 ðxi Þ ÿ T Þ
 ð3;2Þ ð3;3Þ ð3;2Þ
 i¼1
2:2: D8 > 0; Cf ðx > 0Þ; Cf ðg > 0Þ; Cf ðxg > 0Þ
2 2 2
 ð3;2Þ ð3;2Þ ð3;2Þ
 ¼ r10 T 4 þ r11 T 3 þ r12 T 2 þ r13 T þ r14
2:3: D8 > 0; Cf ðx > 0Þ; Cf ðg > 0Þ; Cf ðxg > 0Þ
2 2 2
h     i Y
4
2:4:
h
D8 ¼ 0; D7 6¼ 0; h31 < 0; g
ÿh31
h30
q
> 0; q21 < 0; g
ÿq21
q20 >0 : r2 ðT Þ ¼ resultantðf1 ; xb1 ÿ T ; xÞ ¼ a40 ðxb1 ðxi Þ ÿ T Þ
30 20
i¼1
¼ r20 T 4 þ r21 T 3 þ r22 T 2 þ r23 T þ r24 ;
3. Equation (9) P S has one physical solution iff one of the
following statements holds: where fðxi Þ ¼ 0, i ¼ 1; 2; 3; 4.
GAO ET AL.: COMPLETE SOLUTION CLASSIFICATION FOR THE PERSPECTIVE-THREE-POINT PROBLEM 935

By the Lemmas 1 and 2, we have the following results: 2. Equation (11) has three physical solutions iff one of the
following statements holds:
Vf1 ðx > 0Þ ¼ the number of sign changes in coefficients of fðxÞ
h i
ð4;4Þ ð4;3Þ ð4;3Þ ð4Þ
Vf1 ðb1 > 0Þ ¼ the number of sign changes in coefficients of r1 ðT Þ 2:1: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
Vf1 ðxb1 > 0Þ ¼ the number of sign changes in coefficients of r2 ðT Þ:
h i
ð4;3Þ ð4;4Þ ð4;3Þ ð4Þ
2:2: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
By the complete discrimination method [22], Vf1 ¼ 2 iff one h i
ð4;3Þ ð4;3Þ ð4;3Þ ð4Þ
of the following conditions holds: 2:3: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1 :
8
>
> 1Þ D2 > 0; D3 > 0; D4 < 0
< 3. Equation (11) has two physical solutions iff one of the
2Þ D2 6¼ 0; D3 < 0; D4 < 0
ð13Þ following statements holds:
> 3Þ D2  0; D3 > 0; D4 > 0
>
: h i
4Þ D2 > 0; D3 ¼ 0; D4 > 0: ð2;2Þ ð2;2Þ ð2;2Þ ð2Þ
ð2Þ
3:1: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
We denote one of the conditions by Cf1 : It is clear that there h i
ð4;4Þ ð4;2Þ ð4;2Þ ð4Þ
are the following results: 3:2: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
h i
ð4;3Þ ð4;3Þ ð4;2Þ ð4Þ
1 3:3: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
Vf1 ðx > 0Þ ÿ Vf1 ðx < 0ÞÞ ¼ h i
h 2 i ð4;3Þ ð4;2Þ ð4;3Þ ð4Þ
3:4: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
signðA1 Þ þ signðA1 A2 Þ þ signðA2 A3 Þ þ signðA3 A4 Þ ; h i
ð4;2Þ ð4;4Þ ð4;2Þ ð4Þ
1 3:5: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
Vf1 ðb1 > 0Þ ÿ Vf1 ðb1 < 0Þ ¼ h i
h 2 i ð4;2Þ ð4;3Þ ð4;3Þ ð4Þ
3:6: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
signðB1 Þ þ signðB1 B2 Þ þ signðB2 B3 Þ þ signðB3 B4 Þ ; h i
ð4;2Þ ð4;2Þ ð4;4Þ ð4Þ
1 3:7: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1 :
Vf1 ðxb1 > 0Þ ÿ Vf1 ðxb1 < 0Þ ¼
h 2 i
signðC1 Þ þ signðC1 C2 Þ þ signðC2 C3 Þ þ signðC3 C4 Þ : 4. Equation (11) has one physical solution iff one of the
following statements holds:
Since Resultantðf1 ; b1 ; xÞ 6¼ 0; and Resultantðf1 ðxÞ; h i
ð2;2Þ ð2;1Þ ð2;1Þ ð2Þ
x; xÞ 6¼ 0, we have 4:1: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
h i
1 ð2;1Þ ð2;2Þ ð2;1Þ ð2Þ
Vf1 ðx > 0Þ ¼ 4:2: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
h 2 i h i
ð2;1Þ ð2;1Þ ð2;2Þ ð2Þ
Vf1 þ signðA1 Þ þ signðA1 A2 Þ þ signðA2 A3 Þþ signðA3 A4 Þ ; 4:3: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
h i
ð4;4Þ ð4;1Þ ð4;1Þ ð4Þ
1 4:4: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
Vf1 ðb1 > 0Þ ¼ h i
h 2 i ð4;3Þ ð4;2Þ ð4;1Þ ð4Þ
4:5: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
Vf1 þ signðB1 Þþ signðB1 B2 Þ þ signðB2 B3 Þþ signðB3 B4 Þ ; h i
ð4;3Þ ð4;1Þ ð4;2Þ ð4Þ
1 4:6: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
Vf1 ðxb1 > 0Þ ¼ h i
h 2 i ð4;2Þ ð4;3Þ ð4;1Þ ð4Þ
4:7: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
Vf1 þ signðC1 Þ þ signðC1 C2 Þ þ signðC2 C3 Þ þ signðC3 C4 Þ : h i
ð4;2Þ ð4;2Þ ð4;2Þ ð4Þ
4:8: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
For any two equations fðxÞ and gðxÞ, if ResultantðfðxÞ; h i
ð4;2Þ ð4;1Þ ð4;3Þ ð4Þ
gðxÞ; xÞ 6¼ 0, then 4:9: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
h i
ÿ
ð4;1Þ ð4;4Þ ð4;1Þ ð4Þ
VfðxÞ ¼ VfðxÞ gðxÞ > 0Þ þ VfðxÞ ðgðxÞ < 0 : 4:10: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
h i
ð4;1Þ ð4;3Þ ð4;2Þ ð4Þ
Then, we have 4:11: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
h i
ð4;1Þ ð4;2Þ ð4;3Þ ð4Þ
Vf1 ðx > 0; b1 > 0Þ ¼ 4:12: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1
1ÿ
h i
ð4;1Þ ð4;1Þ ð4;4Þ ð4Þ
Vf1 ðx > 0Þ þ Vf1 ðb1 > 0Þ þ Vf1 ðxb1 > 0Þ ÿ Vf1 : 4:13: Cf1 ðx > 0Þ; Cf1 ðb1 > 0Þ; Cf1 ðxb1 > 0Þ; Cf1 :
2
So, for a0 6¼ 0 and D4 6¼ 0, we can solve the equation system ðn;jÞ
The explicit expressions for all Cf may be found in
completely. Note that D4 is the discriminant for f1 . From the
Appendix B.
above discussion, we have proven.
If D4 ¼ 0 and D3 6¼ 0, then we know that there exists a
Theorem 5. For a0 6¼ 0, D4 6¼ 0, we have
polynomial Q1 ðxÞ ¼ q10 x2 þ q11 x þ q12 such that f1 ðxÞ ¼
1. Equation (11) has four physical solutions iff cQ1 ðxÞ H12 ðxÞ, where H1 ðxÞ ¼ D3 x þ D3 . Let DQ1 ¼ resultant
ð4;4Þ ð4;4Þ ð4;4Þ ð4Þ
Cf1 ðx > 0Þ, Cf1 ðb1 > 0Þ, Cf1 ðxb1 > 0Þ, and Cf1 ðQ1 ; diffðQ1 Þ; xÞ. If DQ1 > 0, then Q1 ðxÞ has two different real
hold. solutions. Let
936 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO. 8, AUGUST 2003
h ÿÿD
i
r3 ðT Þ ¼ resultantðQ1 ; b1 ÿ T ; xÞ ¼ r30 T 2 þ r31 T þ r32 ; 3:1:
D
D3 6¼ 0; D3 < 0; b1 3 >0
3 D3

r4 ðT Þ ¼ resultantðQ1 ; xb1 ÿ T ; xÞ ¼ r40 T 2 þ r41 T þ r42 : h ÿÿq
i

q
3:2: D3 6¼ 0; DQ1 ¼ 0; ÿ q11 > 0; b1 11
q10 >0
10
By Lemmas 1 and 2, we have: h ÿÿa
i
a
3:3: D3 ¼ 0; D2 ¼ 0; D1 6¼ 0; a1 < 0; b1 1
4a0 >0
0
VQ1 ðx > 0Þ ¼ the number of sign changes in coefficients of Q1 ðxÞ h ÿÿD
i
D
3:4: D3 ¼ 0; D2 6¼ 0; DH2 ¼ 0;D2 < 0; b1 2 >0
VQ1 ðb1 > 0Þ ¼ the number of sign changes in coefficients of r3 ðT Þ 2 D2
h i
VQ1 ðxb1 > 0Þ ¼ the number of sign changes in coefficients of r4 ðT Þ: 3:5: D3 ¼
ð2;2Þ
6 0; DQ1 > 0; CQ ðx > 0Þ; CQ
ð2;1Þ
ðb1 > 0Þ; CQ
ð2;1Þ
ðxb1 > 0Þ
1 1 1
h i
Similarly, we have ð2;1Þ ð2;2Þ ð2;1Þ
3:6: D3 ¼
6 0; DQ1 > 0; CQ ðx > 0Þ; CQ ðb1 > 0Þ; CQ ðxb1 > 0Þ
1 1 1
1 h i
VQ1 ðx > 0; b1 > 0Þ ¼ 3:7: D3 6¼ 0; DQ1 > 0; CQ
ð2;1Þ
ðx > 0Þ; CQ
ð2;1Þ
ðb1 > 0Þ; CQ
ð2;2Þ
ðxb1 > 0Þ
h 2 h
1 1 1
i

VQ1 ðx > 0Þ þ VQ1 ðb1 > 0Þ þ VQ1 ðxb1 > 0Þ ÿ VQ1 : 3:8: D3 ¼ 0; D2 6¼ 0; DH1 > 0; CH
ð2;2Þ
1
ðx > 0Þ; CH
ð2;1Þ
1
ðb1 > 0Þ; CH
ð2;1Þ
2
ðxb1 > 0Þ

ÿ
2 h i
If DQ1 ¼ 0, then Q1 ðxÞ ¼ q0 x þ 2qq1110 . 3:9: D3 ¼ 0; D2 6¼ 0; DH1 > 0; CH
ð2;1Þ
1
ðx > 0Þ; CH
ð2;2Þ
1
ðb1 > 0Þ; CH
ð2;1Þ
2
ðxb1 > 0Þ

For D4 ¼ 0 and D3 ¼ 0, D2 6¼ 0, let H1 ðxÞ ¼ D2 x2 þ h i

ð2;1Þ ð2;1Þ ð2;2Þ
3:10: D3 ¼ 0; D2 ¼
6 0; DH1 > 0; CH ðx > 0Þ; CH ðb1 > 0Þ; CH ðxb1 > 0Þ :
D x þ D . Then, f ðxÞ ¼ cH 2 ðxÞ: If D > 0, then V ¼ 2. Let
1 1 2
2 2 1 1 H1 H1
ðn;jÞ
r5 ðT Þ ¼ resultantðH1 ; b1 ÿ T ; xÞ ¼ r50 T 2 þ r51 T þ r52 ; The explicit expressions for all Cf may be found in
Appendix B.
r6 ðT Þ ¼ resultantðH1 ; xb1 ÿ T ; xÞ ¼ r60 T 2 þ r61 T þ r62 : If Resultantðf1 ; b1 ; xÞ ¼ 0, then the equation system
Similar to the above discussion, we can solve VH1 ðx > 0; becomes the following form
8
b1 > 0Þ. If D4 ¼ 0, D3 ¼ 0, D2 ¼ 0, D1 6¼ 0, then f1 ðxÞ ¼ cH24 , < f11 ¼ a10 x3 þ a11 x2 þ a12 x þ a13 ¼ 0;
b ¼ b10 x4 þ b11 x3 þ b12 x2 þ b13 x þ b14 > 0; ð14Þ
where H2 ðxÞ ¼ 4a0 x þ a1 . : 11
x>0
From the above discussion, we can obtain the following
theorem. which can be treated with the method in Section 3.2. We
may solve the case for Resultantðf1 ; x; xÞ ¼ 0 and
Theorem 6. For D4 ¼ 0, we have Resultantðf1 ; b1 ; xÞ ¼ 0 similarly.
1. Equation (11) has three physical solutions iff 3.4 A Special Case of the P3P Problem
"   Let us assume that a ¼ b ¼ 1 and r ¼ q. Since the formulas,
D3 D3
D3 ¼
6 0; ÿ > 0; b1 ÿ > 0; DQ1 in this case are quite simple, we may have an intuitive idea
D3 D3 about the distribution of the solutions. The P3P equation
#
system becomes
ð2;2Þ ð2;2Þ ð2;2Þ
> 0; CQ1 ðx > 0Þ; CQ1 ðb1 > 0Þ; CQ1 ðxb1 > 0Þ 
f ¼ ÿx2 ÿ py þ qxy þ 1
ð15Þ
g ¼ ÿy2 ÿ qx þ pxy þ 1:
holds;
2. Equation (11) has two physical solutions iff one of the Using Wu-Ritt’s method, this equation system has the
following statements holds: following two components:
h ÿ q
D ÿ D
i 
2:1:
q
D3 6¼ 0; DQ1 ¼ 0; q11 < 0;b1 ÿq11 ; D3 < 0; b1 ÿ D3 f1 ¼ x2 ÿ qx þ p ÿ 1
10 10 3 3 ð16Þ
h i f2 ¼ y ÿ 1
ð2;2Þ ð2;2Þ ð2;2Þ
2:2: D3 ¼
6 0; DQ1 > 0; CQ ðx > 0Þ; CQ ðb1 > 0Þ; CQ ðxb1 > 0Þ
h
1 1 1
i 
ð2;2Þ ð2;2Þ ð2;2Þ g1 ¼ ðÿ1 þ q2 Þx2 þ ðÿq ÿ qpÞx þ 1 þ p
2:3: D3 ¼ 0; D2 6¼ 0; DH1 > 0; CH ðx > 0Þ; CH ðb1 > 0Þ; CH ðxb1 > 0Þ ð17Þ
h
1 1 2
g2 ¼ y ÿ qx ÿ 1:
D
ÿ D


2:4: 6 0; DQ1 > 0; D3
D3 ¼ < 0; b1 ÿ D3 ;
3 3 It is clear that the number of positive solutions of (16) is
i determined by f1 ðxÞ ¼ 0. Notice that f1 ðxÞ is a quadratic
ð2;2Þ ð2;1Þ ð2;1Þ
> 0; CQ ðx > 0Þ;CQ ðb1 > 0Þ; CQ ðxb1 > 0Þ
h
1 1 1 equation in x, we have the following results:
D
ÿ D


2:5: D3 6¼ 0; DQ1 > 0; D3 < 0; b1 ÿ D3
3 3 . Equation system (16) has one positive solution iff,
i
ð2;1Þ
> 0; CQ
ð2;1Þ
ðx>0Þ; CQ
ð2;1Þ
ðb1 > 0Þ;CQ ðxb1 > 0Þ

1 1 1 q>0
h ÿ
1 4þq2 or p < 1:
2:6:
D
D3 6¼ 0; DQ1 > 0; D3 < 0; b1 ÿ D3
D 2p ¼ 4
3 3
i
ð2;1Þ ð2;1Þ ð2;2Þ
> 0; CQ
1
ðx > 0Þ; CQ
1
ðb1 > 0Þ; CQ
1
ðxb1 > 0Þ : . Equation system (16) has two positive solutions iff,

4 þ q2
3. Equation (11) has one physical solution iff one of the q > 0 and > p > 1:
4
following statements holds:
GAO ET AL.: COMPLETE SOLUTION CLASSIFICATION FOR THE PERSPECTIVE-THREE-POINT PROBLEM 937

Now, we will discuss (17). Let g ¼ qx þ 1.


R1 ðtÞ ¼ resultantðg1 ;gÿt;xÞ¼ðq2 ÿ1Þt2 þðq2 ÿq2 pÿ2Þtþðq2 ÿ1Þ
ð18Þ
R2 ðtÞ ¼ resultantðg1 ;xgÿt;xÞ¼R20 t2 þR21 tþR22 :

Here, R20 ¼ ðq ÿ 1Þ2 ðq þ 1Þ2 , R21 ¼ ÿqð1 þ pÞðq2 p ÿ 2q2 þ 3Þ,

R22 ¼ ðq ÿ 1Þðq þ 1Þð1 þ pÞ. Denote the discriminant of g1 by

. We have

¼ ð1 þ pÞðq2 p ÿ 3q2 þ 4Þ:

By Descartes rules of sign, (17) has two positive solutions iff
Fig. 2. Solution distribution for case a ¼ b ¼ 1; q ¼ r.
4
q > 1 and p > 3 ÿ 2 :
q Fig. 2 is the solution distribution diagram for this special
2

Otherwise, ES2 has no positive solution. case. L1 is p ¼ 4þq 2 4

4 , L2 is p ¼ q ÿ 2, and L3 is p ¼ 3 ÿ q2 .
We still need to consider the reality conditions (2): Table 2 shows where the solutions come from for each
region.
0< ; < , 0 < þ 2 < 2, and 2 > , which can be Here are some general observations from this diagram.
reduced to The P3P most probably will have one solution. The
probabilities to have two, three, and four solutions decrease
ÿ 2 < p < 2; -2 < q < 2; q2 ÿ 2 < p: in order. The P3P problem tends to have more solutions when
the three perspective angles are small. The most complicated
Combining the above conditions, we have the following
case occurs when the three perspective angles are almost
classification for the P3P problem: equal to the correspondent inner angles of triangle ABC.
1. Point P has four solutions, iff
4 EXPERIMENTAL RESULTS ON THE ALGEBRAIC
4 þ q2 4
2 > q > 1; > p > 1; and p > 3 ÿ : METHOD
4 q2
Two sets of formulas are given in Sections 2 and 3,
respectively. In this section, we will show how to combine
2. Point P has three solutions, iff them to obtain more robust numerical solutions.
( pffiffi  In Section 2, we give a set of complete analytical
1<q< 2 1<q<2 solutions to the P3P problem. Previous methods only give
or 2
3 ÿ q42 < p  1 p ¼ 4þq
4 : the main component, which, although covering most of the
cases, may not provide solutions in many cases. With these
formulas, numerical solutions to the P3P problems could be
3. Point P has two solutions, iff
found in all cases. The solution of the P3P problem is
( ( reduced to the computation of ZeroðT Si =Ti Þ. The main
0<q1 1<q<2
4þq2 or 4þq2 component ZeroðT S1 =T1 Þ has been studied in previous
1<p< 4 4 <p<2 work [9], [4], in which we need to solve a quartic equation.
( pffiffi
2<q<2 To solve cases ZeroðT Si :Ti Þ; i > 1, we need to solve
or equations with degrees  3. To solve linear and quadratic
1 < p  3 ÿ q42 and p > q2 ÿ 2:
equations is trivial. For the cubic and quartic equations, we
may write their roots aspformulas
ffiffiffiffiffiffi with radicals. Because
4. Point P has one solution, iff these formulas involve ÿ1, it is generally difficult to
( distinguish real roots from complex roots. In Section 3, we
 pffiffi pffiffi pffiffi pffiffi
ÿ 3 < q < 1 or 2  q < 3 1<q< 2 also give explicit formulas to determine the number of real
or positive roots. Combining these formulas with the solutions
q2 ÿ 2 < p < 1 q2 ÿ 2 < p  3 ÿ q42
in radical form gives us an efficient and stable method to
 pffiffi pffiffi  solve these equations.
0 < q  1 or 2q< 3 0<q1 Based on the criteria obtained in Section 3, we introduce
or or 2
p¼1 p ¼ 4þq
4 : the concept of stable and critical values for the parameters.
For a condition C in these criteria, let F ðCÞ be the set of  such

TABLE 2
Number of Solutions and the Component Providing the Solutions
938 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO. 8, AUGUST 2003

Fig. 3. Experimental results.

that one of the following formulas:  > 0,  < 0,  6¼ 0 occurs values. If they are stable, let M ¼ 16 (usual floating-point
in C. For instance, let C be Condition 1.1 in Theorem 3. Then, number); otherwise, let M ¼ 40.
F ðCÞ ¼ fa þ b ÿ 1; q; D13 ;
1 ;
2 ;
3 g: S4 Find all the solutions of T Sk ¼ 0 using high precision
numbers with M digits.
A set of values for the parameters a; b; p; q; r is said to be stable
S5 If the number of solutions obtained in S4 is the same as
for condition C if for each  2 F ðCÞ, jða; b; p; q; rÞj > ,
N, then these are the solutions. Otherwise, we need to
where  > 0 is a small number, say 1 percent of the value
select N “right” solutions from them. We first replace a
range of the parameters. Otherwise, it is critical. Basically
complex number u þ vi in the solutions with u if v is very
speaking, the critical values are those which will approxi-
small, say jvj < 10ÿ4 m. Then, we select the N largest
mately vanish the expressions in F ðCÞ. Therefore, the
positive solutions. According to Step S2, this is possible.
probability for the occurrence of the critical values is near
zero. The following experiments are done with Maple. The
If a set of values is stable, then a small variation of these first experiment is to show the stability of the criteria in
values will give the same number of solutions. On the Section 3. These formulas use arithmetic operations (+, -, *)
contrary, for a set of critical values, a small variation may lead only and are of moderate size. The computation will be
to changes in the number of solutions. Therefore, for a set of robust. Also, from the above analysis, only for critical
stable values, we may use usual floating-point calculations to values of the parameters, the computation will be unstable.
enhance the computation speed; and for a set of critical values, We also know that the probability for the occurrence of the
we may use high-precision computation tools provided by critical values is near zero. This observation gives another
symbolic computation software or by other special tools, like assurance that the computation is stable. The following
the exact geometric computation method [18]. experimental results support this statement.
From the above discussion, we propose the following the The parameters a, b, p, q, and r are randomly generated
CASSC algorithm. within some ranges by a random number generator in
1
Maple 7. We take a, b 2 ð10 10Þ, and p, q, r 2 ðÿ2 2Þ. One
CASSC Algorithm hundred trials are carried out and 100 sets of parameters are
Input: A set of values for parameters a; b; p; q; r. generated for each trial. For each set of parametric values,
Output: The physical solutions. two results are computed: one with the original parametric
values; the other with the parametric values perturbed by
S1 In this step, we will decide which of the 10 components in
random noises in certain level. In trial i, let ni be the number
(DE) will provide the solution. Let ESi be the set of
of the parametric values such that the two results are the
polynomials in T Si involving the parameters a; b; p; q; r
same and let kninÿnk (here, n ¼ 100) be the relative sets error.
only. Then, for a set of parameter values a; b; p; q; r, the
Fig. 3a gives the median, mean, and standard deviation of
solutions will be provided by Ck if ESk ða; b; p; q; rÞ ¼ 0 and
the relative errors w.r.t. varying noises. We observe that the
tk ¼ jTk ða; b; p; q; rÞj 6¼ 0. In practice, we use the criteria:
algorithm yields very graceful degradation with increasing
jP ða; b; p; q; rÞj < 10ÿ4 m for each P 2 ESk and tk > 10ÿ2 m,
noises and are, therefore, very stable.
where m ¼ maxða; b; p; q; rÞ. We will find the smallest k
The second experiment concerns the stability of the
such that the above conditions are satisfied. If no such k
exits, then there exists no solution. Otherwise, goto S2. whole CASSC algorithm. We choose three sets of parametric
values S1 ¼ fa ¼ 0:6; b ¼ 0:5; p ¼ r ¼ 0; q ¼ 0g, S2 ¼ fa ¼
S2 Determine the number of physical solutions with the 1; b ¼ 1; p ¼ 1:2; r ¼ 1:2; q ¼ 1:3g, and S3 ¼ fa ¼ 1:35; b ¼
criteria given in Section 3. Let N be the number and  the 1:65; p ¼ 1:2; r ¼ 1:0; q ¼ 1:67g. We choose n ¼ 100 random
criterion used to determine this number. If N ¼ 0, the numbers as noises nearing each of the seven numbers:
algorithm terminates. Otherwise goto S3. e1 ¼ 0:08, e2 ¼ 0:1, e3 ¼ 0:12, e4 ¼ 0:14, e5 ¼ 0:16, e6 ¼ 0:18,
S3 In this step, we will decide the digits of precisions used e7 ¼ 0:2 and compute the solutions for S1 , S2 , and S3 with
in the computation. With the criterion  obtained in our algorithm. Let ni be the number of parametric values
Step S2, we may determine whether a; b; p; q; r are stable near ei which give different number of solutions with the
GAO ET AL.: COMPLETE SOLUTION CLASSIFICATION FOR THE PERSPECTIVE-THREE-POINT PROBLEM 939

Fig. 4. Existence conditions for point J. Fig. 5. Two cases for point H.

original parametric values, and knni k the relative sets error. 5.1 Determine CA ¼ SAB \ SAC
The experimental results are illustrated in Fig. 3b. It is easy Let ABdi (AB de ) denote the intersection of SAB and plane ABC
to check that S3 is stable and S1 ; S2 are critical. The
which is on the same (opposite) side of AB with point C. Since
experimental results strongly support the fact that our
0 0
definition of stable and critical values is meaningful. From the axes of symmetry for part-toroids SAB and SAC meet in
Fig. 3b, we see that even for a set of stable values (S3 ), the point A and point A is also on the part-toroids, from the shape
computation may be unstable. This is caused by the high of the part-toroid each branch of CA must pass through plane
noise level. The computation is stable for noises less that
ABC. That is, CA must meet with plane ABC. Curve CA
10 percent of the value range of the parameters, which is
intersects with plane ABC in at most four points: J ¼ AB de
quite reasonable.
We test our algorithm with a larger set of samples. For a \ACde ; H ¼ AB de \ ACdi ; K ¼ ABdi \ ACde , and I ¼ AB di \ AC
di .
set of solutions obtained with the algorithm, we substitute Please note that, in some cases (e.g., in Fig. 7), point A may not
them into (ES) and check whether the substituted values be on CA . In this case, point A is on the intersection curve of the
are zero or not. We take 100 sets of parameters randomly.
The maximal substituted value into ðESÞ is 0:3  10ÿ10 for parts of the toroids that are excluded by us.
the equation systems which is satisfactory. From now on, we also use A; B; C to denote the angles of
We also tested the speed for the CASSC algorithm, which ffA; ffB; ffC. We first give the existence conditions for points J,
should be fast for the following reasons. Steps S1, S2, and S3 H, K, and I.
only involve the evaluation of rational expressions of
moderate size. Step S4 is to solve univariate equations of . Point J exists if þ < A (Fig. 4). In Fig. 4,
degrees at most four. Step S5 is computationally trivial. We ffBJA ¼ , ffCJA ¼ , and ffBJC ¼ þ . If þ
test our algorithm with 100 randomly chosen samples. The is large enough, AB de and AC de will have no
average running time for Steps S1, S3, and S5 is almost zero; intersection point. If þ ¼ A, AB de is tangent to
the average running time for Step S2 is 0.011 second; the de at point A. If þ < A, the intersection of
AC
average running time for Step S4 is 0.013 second. The data is de and AC
AB de will exist.
collected with Maple V on a PC with a 2G CPU. Step S4, which . Point H. There are two cases. If B < , then point B
is to solve the quartic equations, is the most time consuming is outside of SAC and H1 exists if B < and þ
step. Note that this step is needed in most previous A < : In Fig. 5, ffBH1 A ¼ , ffCH1 A ¼ . To ensure
approaches to solving the P3P problem. The implementation the existence of H1 , must be greater than þ A. If
is based on the interpreter language of Maple for symbolic ¼ þ A, AC di is tangent to AB de at point A. If
computation which is known to be much slower than < B, then point B is inside SAC and H2 exists if
implementations with C languages for the tasks mentioned
< B and < þ A.
above. As a conclusion, the CASSC algorithm is quite fast.
. Point K. There are two cases. K1 exists if C < and
There is no problem to provide realtime solutions to the
þ A < : K2 exists if < C and < þ A.
P3P problem.
. Point I. There are two cases: I1 exists if B < ; C < ,
and þ þ A < 2 (Fig. 6). I2 exists if < B and
5 THE GEOMETRIC APPROACH < C (Fig. 6).
Let us consider the three conditions ¼ ffBP C, ¼ ffAP C,
and ¼ ffAP B separately. The set of all P satisfying
0
condition ffAP B ¼ is part of a toroid (part-toroid) SAB .
0 0
Similarly, we can define SAC and SBC . Because the three
part-toroids are symmetric with the plane ABC, we need
only consider what happens on one side of plane ABC. Let
0
SAB denote the half of SAB which is on one side of
plane ABC. We can similarly define SAC and SBC .
We divide the problem into two steps: First, we
determine the intersection curve CA of surfaces SAB and
SAC ; then, we determine the intersection of CA with SBC .
We have solved the first step completely. For the second
step, we have some partial results. Fig. 6. Existence conditions for point I.
940 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO. 8, AUGUST 2003

Fig. 7. CA consists of two curves.

We will now give a classification of CA by counting the Using (19) we can simplify this condition to
intersections of CA with plane ABC. Suppose that S; U; V are EðH2 Þ \ EðK2 Þ \ EðJÞ. That is,
points. We use EðSÞ (EðSÞ) to denote the existence (non-
existence) condition of point S. Notation S n ðU; V Þ means j ÿ j<A< þ ; B < ; andŠ < C:
that if S exists, then U and V will not exist. Notation S; U ) V In this case, CA consists of two spatial curves: one is
means that, if S and U exist, then V exists. From the results in from A to I2 and the other is from H2 to K2 . Since the
the preceding sections, we have the following conclusion:
detailed analysis is the same, we will omit it below.
8
> J n ðK1 ; H1 Þ . CA intersects plane ABC in two points. There are
>
> five subcases.
>
< H1 n ðJ; K1 ; H2 ; I2 Þ; H2 n ðH1 ; I1 Þ
K1 n ðJ; H1 ; K2 ; I2 Þ; K2 n ðK1 ; I1 Þ ð19Þ
>
> Case 1. The intersections are J; K2 (J; H2 ) if
>
> I 1 n ðI 2 ; K2 ; H2 Þ; I 2 n ðI 1 ; K 1 ; H1 Þ
: 8 08 1
K2 ; H2 ) I2 ; J; I2 ) H2 ; K2 : < þ <A < þ <A
B< @ <B A:
: :
. CA intersects plane ABC in four points. From the <C C<
above analysis, the four points must come from
In this case, CA consists of one spatial curve from
H1 ; H2 ; K1 ; K2 ; I1 ; I2 ; and J. Since H1 n ðH2 Þ; K1 nðK2 Þ
and I1 n ðI2 Þ, the fourth point must be J. From J n J to K2 (J to H2 ).
ðK1 ; H1 Þ; K2 and H2 must exist. Finally, from Case 2. The intersections are J; I1 if
K2 ; H2 ) I2 we get the fourth point I2 . So, the four þ < A; B < ; and C < :
points are J; H2 ; K2 ; I2 . Then, the condition of this case
should be EðJÞ \ EðH2 Þ \ EðK2 Þ \ EðI2 Þ \ EðH1 Þ \ In this case, CA consists of one spatial curve from
EðK1 Þ \ EðI1 Þ which is equivalent to EðJÞ \ EðI2 Þ by J to I1 .
(19). That is, Case 3. The intersections are H2 ; K1 (H1 ; K2 ) if
8 08 1
þ < A; < B; and < C: < þA< < þA<
<B @ B< A:
In this case, CA consists of two spatial curves: one is : :
from point J to I2 and the other is from H2 to K2 . Fig. 7 C< <C
shows the case in the ABC plane and the spatial case. In this case, CA consists of one spatial curve from
Note that, in this case, point A is not on the curves.
H2 to K1 (or H1 to K2 ).
. CA intersects plane ABC in three points. From
Case 4. The intersections are K1 ; I1 (H1 ; I1 ) if
J n ðK1 ; H1 Þ, we know that, if J exists, at lease one of
8 08 1
H2 and K2 should exist. Actually only one of H2 and K2 > þA< > þA<
>
< >
<
can exist. Otherwise, from K2 ; H2 ) I2 we know that B< B B< C
B C:
there will be four points! Then, we know that either H2 > C < @> C < A
>
: >
:
or K2 exists. From H2 n ðH1 ; I1 Þ and K2 n ðK1 ; I1 Þ, we þ þ A < 2 þ þ A < 2
know that I2 must exist. Since J; I2 ) H2 ; K2 , point J
must not exist. Since H1 n ðH2 Þ; K1n ðK2 Þ and I1 n ðI2 Þ, if In this case, CA consists of one spatial curve from
we assume that H1 exists, from H1 n ðJ; K1 ; H2 ; I2 Þ we K1 to I1 (H1 to I1 ).
know that the other two points are K2 and I1 . This Case 5. The intersections are K2 ; I2 (H2 ; I2 ) if
contradicts to K2 n ðK1 ; I1 Þ. Thus, H2 must exist. From 8 08 1
H2 n ðH1 ; I1 Þ, we know the other two points are K2 and < þA< < þA<
<B @ <B A:
I2 . The condition of this case should be : :
<C <C
EðH2 Þ \ EðK2 Þ \ EðI2 Þ
In this case, CA consists of one spatial curve from
\ EðJÞ \ EðH1 Þ \ EðK1 Þ \ EðI1 Þ: K2 to I2 (H2 to I2 ).
GAO ET AL.: COMPLETE SOLUTION CLASSIFICATION FOR THE PERSPECTIVE-THREE-POINT PROBLEM 941

one intersection point. If A < ; B < ; C < , from

Section 5.1, point I1 must exist. Similar to Lemma 8,
points A and I1 must be in different sides of SBC . Similar
to the proof of Lemma 7, a solution must exist. u
t

6 CONCLUSION
In this paper, we give a complete and robust algorithm
CASSC to find the numerical solutions for the P3P problem.
Fig. 8. A unique solution exists. This algorithm is based on two sets of formulas obtained by
us. The first is a set of complete analytical solutions to the
. CA intersects plane ABC in one point. We need to P3P problem. The second is a set of formulas to determine
consider two subcases. the number of real positive solutions to the P3P problem.
We also give partial geometric criteria for the number of
Case 1. The intersection is H2 (K2 ) if solutions of the P3P problem. This kind of results, like
8 08 1 Theorem 9, involves linear inequalities only and, hence, is
<j ÿ j < A < þ <j ÿ j < A < þ simpler and more intuitive than the algebraic approach. To
<B @ B< A:
: : find a complete geometric classification for the P3P problem
C< <C is a still challenging problem. There might be two
In this case, CA consists of one spatial curve from difficulties in doing so. The complete results reported in
A to H2 (A to K2 ). Section 5.1 are based on geometric intuition coming from a
Case 2. The intersection is I1 if dynamic geometry software: Geometry Expert [23].
Using Geometry Expert, we can see clearly how AB di and
j ÿ j<A< þ ; B < ; C < ; and þ þA de change when changing the six free parameters
AB
< 2: continuously. But for the 3D case, there is still no adequate
software to get an intuitive idea of how CA looks like. Also,
In this case, CA consists of one spatial curve from it is doubtful that the complete classification of the
A to I1 . P3P problem can be expressed with linear inequalities only.
5.2 Determine CA \ SBC
Determining the intersection of CA and SBC is much more APPENDIX A
difficult than determining CA . We will discuss the reason in The triangular sets in the zero decomposition for the
Section 6. Here, we will report some partial results. P3P problem.
 4
Lemma 7. The P3P problem has one or three solutions if CA 3 2
a0 x þa1 x þa2 x þa3 xþa4 ;
ðT S1 Þ
consists of one spatial curve and the two intersection points of b0 yÿb1 :
plane ABC and CA are not in the same side of SBC . 8
> 3 2
< a5 x þa6 x þa7 xþa8 ;
Proof. Since CA is a continuous spatial curve and the two b2 yÿb3 ; ðT S2 Þ
intersection points of plane ABC and CA are not in the >
: 2
a þðÿ2þ2bÿbr2 Þaÿ2bþb2 þ1:
same side of SBC , CA must intersect SBC for odd times. In 8 2 2
> ðr p ÿ4pqrþr2 q2 Þx2 þð4p2 qÿp2 r2 qÞxÿ4p2 þr2 p2 ;
addition, the maximum number of solutions is four, hence >
>
the problem has a unique solution or three solutions. t u < b4 yÿb5 ;
ðT S3 Þ
Lemma 8. If , ( ; ; ; ) are obtuse angles and > A >
> ðÿ4p2 þ4pqrþr2 p2 þr2 q2 ÿr3 pqÿ4q2 Þaþr2 p2 ÿ4pqrþ4q2 ;
>
:
( > B; > C), then the P3P problem can only have one or ðÿ4p2 þ4pqrþr2 p2 þr2 q2 ÿr3 pqÿ4q2 Þbþr2 q2 þ4p2 ÿ4pqr:
three solutions. 8 2 2 2 2
> ðp bþq bÿp Þx þðÿ4bqþp2 qÞxþ4bÿp2 ;
>
>
Proof. See Fig. 8. We have ffBI1 A ¼ > 2 , ffCI1 A ¼ > 2 . < pyþqxÿ2;
ðT S4 Þ
Point I1 is on the same side of BC with point A. >
>
> aþbÿ1;
According to the “reality condition,” we know that :
r:
þ þ < 2, which implies that point I1 is inside SBC . 8
Condition > A means that point A is in the outside of >
> qxÿ1;
>
>
SBC . Thus, the result follows from Lemma 7. u
t >
>
< pyÿ1;
ðp2 þq2 Þaÿq2 ; ðT S5 Þ
>
>
Theorem 9. Under the reality conditions (2), if , , and are >
> ðp2 þq2 Þbÿp2 ;
>
>
obtuse, then the P3P problem can only have one solution. :
r:
Furthermore if A < ; B < ; C < , then the P3P problem 8
has a unique solution. >
> qxÿ1;
>
>
>
>
Proof. From Lemma 8, we know that the problem will have < pyÿ1;
one or three solutions since , , and are obtuse and at ðp4 ÿ2p2 q2 þq4 Þaÿp2 q2 ÿq4 ; ðT S6 Þ
>
>
least one of A; B; C is acute. Since the three angles are all >
> ðp4 ÿ2p2 q2 þq4 Þbÿp2 q2 ÿp4 ;
>
>
obtuse, the three part-toroids and their intersection :
ðp2 þq2 Þrÿ4pq:
curves are concave. This implies that they can only have
942 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO. 8, AUGUST 2003
8 2
> ðp qÿ2prÞxÿp2 þr2 ; b4 ¼ r2 p2 ðrq2 þp2 rÿ4pqÞðp2 ÿpqrþr2 þq2 ÿ4Þ;
>
>
< pyÿ1; b5 ¼ r2 qððrp2 þrq2 ÿr2 pqÞxþpr2 ÿ4pÞððrp2 þrq2 ÿ4pqÞxþq2 pÿqp2 rþp3 Þ;
ðT S7 Þ
>
> ð4r2 þp2 q2 þp4 ÿr4 ÿp3 qrþpr3 qÿ4qprÞb I1 ¼ a0 ;I2 ¼ b0 ;I3 ¼ a5 ;I4 ¼ b2 ;I5 ¼ r;I6 ¼ rpÿ4pqþrq2 ;I7 ¼ p;I8 ¼ ðp2 þq2 Þbÿp2 :
>
:
þ2pr3 qÿ2p2 r2 þ2p3 qrÿp2 q2 r2 ÿp4 ÿr4 : T1 ¼ I0 I1 I2 ;T2 ¼ I0 I2 I3 I4 ;T3 ¼ I0 I2 I4 I5 I6 I7 ;T4 ¼ I0 I2 I4 I6 I7 I8 ;
8
>
> ð2pr3 qÿ2p2 r2 þ4br2 þp2 q2 bþp4 bÿr4 bþ2p3 qrÿp2 q2 r2 ÿp4 ÿr4 ÿ T5 ¼ I0 I2 I4 I6 I7 ;T6 ¼ I0 I2 I4 I5 I7 ;T7 ¼ I0 I2 I3 ;T8 ¼ T7 ;T9 ¼ I0 I1 ;T10 ¼ I0 I1 ;T11 ¼ I0 I1 :
>
>
>
> 3 3 2 2 3 2 2 2 2 3 2 4 4
< p qrbþr pbqÿ4qbprÞx þðÿq r pbþ2pq brþ2p r qþp q r þr qþqr b
þp4 qÿ2bpr3 þ3qr2 bp2 ÿ4r2 bqþ8rpbÿ2rbp3 ÿ2p3 q2 rÿ4bp2 qÿ2r3 pq2 Þx ðT S8 Þ
>
> APPENDIX B
>
> ÿp2 q2 r2 þ2pr3 qþ2p3 qrÿp4 ÿr4 ÿ2p2 r2 ÿ4qbprþq2 br2 þ4bp2 ;
>
> The explicit formulas for the conditions in the theorems in
:
ðÿqprþp2 þr2 Þyþpqxÿ2rxÿ2pþqr: Section 3 can be found at http://www.mmrc.iss.ac.cn/
8
>
>
ðÿ1þaþbÞx2 þðÿqaþqÞxÿ1þaÿb; ~xgao/paper/appendix.ps.
>
< 2
ðÿ1þaþbÞy ÿ1ÿaþqxaþb;
ðT S9 Þ
>
> p; ACKNOWLEDGMENTS
>
:
r: The authors would like to thank the anonymous referees for
8
>
> ð2prÿp2 qÞxÿr2 þp2 ; their valuable suggestions that led to the addition of Section 4
>
>
>
> and other improvements. This work is supported in part by
< pyÿ1;
Chinese National Science Foundation under an outstanding
ðÿpqr3 þr4 þrp3 qÿ4r2 ÿp2 q2 þ4rpqÿp4 Þaþp2 q2 ÿ4rpqþ4r2 ; ðT S10 Þ
>
> youth grant (No. 69725002) and by a National Key Basic
>
> ðÿpqr3 þr4 þrp3 qÿ4r2 ÿp2 q2 þ4rpqÿp4 Þbþ
>
> Research Project of China (No. G1998030600).
:
p4 þr4 þ2r2 p2 þp2 r2 q2 ÿ2rp3 qÿ2pqr3 :
8
>
> rxÿp;
REFERENCES
>
>
>
> 2 2 3 4 2 3 2 2 3 4
< ðÿp r þr qpÿr Þy þðp r ÿp r qþr pÞy [1] M.A. Abidi and T. Chandra, “A New Efficient and Direct Solution
ðÿpqr3 þr4 þrp3 qÿ4r2 ÿp2 q2 þ4rpqÿp4 Þaþp2 q2 ÿ4rpqþ4r2 ; ðT S11 Þ for Pose Estimation Using Quadrangular Targets: Algorithm and
>
> Evaluation,” IEEE Trans. Pattern Analysis and Machine Intelligence,
>
> ðÿpqr3 þr4 þrp3 qÿ4r2 ÿp2 q2 þ4rpqÿp4 Þbþ vol. 17, no. 5, pp. 534-538, May 1995.
>
>
: [2] M.A. Fishler and R.C. Bolles, “Random Sample Consensus: A
p4 þr4 þ2r2 p2 þp2 r2 q2 ÿ2rp3 qÿ2pqr3 :
Paradigm for Model Fitting with Applications to Image Analysis
and Automated Cartomated Cartography,” Comm. ACM, vol. 24,
no. 6, pp. 381-395, 1981.
Here, [3] R. Horaud, B. Conio, and O. Leboulleux, “An Analytic Solution
for the Perspective 4-Point Problem,” Computer Vision, Graphics,
2 2 2
a0 ¼ ÿ2bþb þa þ1ÿbr aþ2baÿ2a and Image Processing, vol. 47, pp. 33-44, 1989.
a1 ¼ ÿ2bqaÿ2a2 qþbr2 qaÿ2qþ2bqþ4aqþpbrþbrpaÿb2 rp [4] L. Quan and Z. Lan, “Linear N-Point Camera Pose Determina-
tion,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 21,
a2 ¼ q2 þb2 r2 ÿbp2 ÿqpbrþb2 p2 ÿbr2 aþ2ÿ2b2 ÿabrpqþ2a2 ÿ4aÿ2q2 aþq2 a2 Þ no. 8, pp. 774-780, Aug. 1999.
[5] C. Su, Y. Xu, H. Li, and S. Liu, “Application of Wu’s Method in
a3 ¼ ÿb2 rpþbrpaÿ2a2 qþqp2 bþ2bqaþ4aqþpbrÿ2bqÿ2q
Computer Animation,” Proc. Fifth Int’l Conf. CAD/CG, vol. 1,
a4 ¼ 1ÿ2aþ2bþb2 ÿbp2 þa2 ÿ2ba; pp. 211-215, 1997.
[6] J.S.C. Yuan, “A General Photogrammetric Method for Determin-
a5 ¼ ðaprþ2qaÿrpbþ2bqÿ2qÿar2 qþpr; ing Object Position and Orientation,” IEEE Trans. Robotics and
a6 ¼ ðÿ2q2 þr2 ÿ4þr2 q2 ÿpqrÞaþbr2 ÿp2 ÿbq2 þbp2 þ2q2 þ4ÿpqrÿ4b; Automation, vol. 5, no. 2, pp. 129-142, 1989.
[7] W.J. Wolfe, D. Mathis, C. Weber, and M. Magee, “The Perspective
a7 ¼ ð6qþprÿ2r2 qÞaþprÿ6qÿrpbþ2bqþqp2 ; View of Three Points,” IEEE Trans. Pattern Analysis and Machine
a8 ¼ 4ÿ4aÿp2 þar2 ;
Intelligence, vol. 13, no. 1, pp. 66-73, Jan. 1991.
[8] Y. Hung, P. Yeh, and D. Harwood, “Passive Ranging to Known
b0 ¼ bðp2 aÿp2 þbp2 þpqrÿqarpþar2 ÿr2 ÿbr2 Þ2 ; Planar Points Sets,” Proc. IEEE Int’l Conf. Robotics and Automation,
vol. 1, pp. 80-85, 1985.
b1 ¼ ðð1ÿaÿbÞx2 þðqaÿqÞxþ1ÿaþbÞðða2 r3 þ2br3 aÿbr5 aÿ2ar3 þr3 þb2 r3 ÿ2r3 bÞx3 þ
[9] R.M. Haralick, C. Lee, K. Ottenberg, and M. Nolle, “Analysis and
ðpr2 þpa2 r2 ÿ2br3 qaþ2r3 bqÿ2r3 qÿ2par2 ÿ2pr2 bþr4 pbþ4ar3 qþbqar5 ÿ2r3 a2 q Solutions of The Three Point Perspective Pose Estimation
Problem,” Proc. Int’l Conf. Computer Vision and Pattern Recognition,
þ2r2 pbaþb2 r2 pÿr4 pb2 Þx2 þðr3 q2 þr5 b2 þrp2 b2 ÿ4ar3 ÿ2ar3 q2 þr3 q2 a2 þ pp. 592-598, 1991.
2a2 r3 ÿ2b2 r3 ÿ2p2 brþ4par2 qþ2ap2 rbÿ2ar2 qbpÿ2p2 arþrp2 ÿbr5 aþ2pr2 bqþ [10] D. DeMenthon and L.S. Davis, “Exact and Approximate Solutions
of the Perspective-Three-Point Problem,” IEEE Trans. Pattern
rp2 a2 ÿ2pqr2 þ2r3 ÿ2r2 pa2 qÿr4 qbpÞxþ4ar3 qþpr2 q2 þ2p3 baÿ4par2 þ Analysis and Machine Intelligence, vol. 14, no. 11, pp. 1100-1105,
Nov. 1992.
ÿ2r3 bqÿ2p2 qrÿ2b2 r2 pþr4 pbþ2pa2 r2 ÿ2r3 a2 qÿ2p3 aþp3 a2 þ2pr2 þp3 þ2br3 qa
[11] D.F. DeMenthon and L.S. Davis, “Model-Based Object Pose in 25
þ2qp2 brþ4qarp2 ÿ2par2 q2 ÿ2p2 a2 rqþpa2 r2 q2 ÿ2r3 qÿ2p3 bþp3 b2 ÿ2p2 brqaÞ; Lines of Code,” Int’l J. Computer Vision, vol. 15, pp. 123-141, 1995.
[12] W.J. Wolfe and K. Jones, “Camera Calibration Using the
b2 ¼ bðÿ4ar3 þ4r3 þar5 ÿ2p3 qþ4rp2 ÿ6pqr2 ÿ4rp2 bÿ4p2 arþ6par2 qþ Perspective View of a Triangle,” Proc. SPIE Conf. Automation
2p2 rq2 þ2p2 ar3 þ2p3 bqþ2p3 qaþp4 arþp2 ar3 q2 ÿ2p2 rq2 aÿp2 rbq2 ÿ Inspection Measurement, vol. 730, pp. 47-50, 1986.
[13] C. Su, C.Y. Xu, H. Li, and S. Liu, “Necessary and Sufficient
2p3 ar2 qÿ2par4 qþ2pr2 bqÞ; Condition of Positive Root Number of P3P Problem,” Chinese J.
b3 ¼ ððÿ1þaþbÞx2 þðÿqaþqÞxÿ1þaÿbÞððÿpar3 þar4 qÿ2ar2 qÿ2r2 bqþ2r2 qÿpr3 þ
Computer Sciences, vol. 21, pp. 1084-1095, 1998 (in Chinese).
[14] L. Yang, “A Simplified Algorithm for Solution Classification of the
r3 bpÞx2 þðÿr2 p2 aþ2r3 paqþ4ar2 ÿr4 q2 aÿar4 ÿ2qarpþ2ar2 q2 ÿr4 bþr2 bq2 þr3 pq Perspective-three-point Problem,” MM-Preprints, MMRC, Acade-
mia Sinica, no. 17, pp. 135-145, 1998.
þ2pqrþ4r2 bÿ2qpbrÿ2r2 q2 ÿ4r2 Þxÿp3 arþ2ar4 qÿpar3 q2 þ2p2 ar2 qÿ2par3 þ2prq2 a
[15] W.T. Wu, Basic Principles of Mechanical Theorem Proving in
ÿ2p2 aqÿ6ar2 qþ4aprÿpr3 þ4pbrþprbq2 ÿ2r2 bqÿ2prq2 þ2qp2 ÿ2bqp2 ÿ4prþ6r2 qÞ; Geometries, vol. I: Part of Elementary Geometries, Beijing: Science
Press, (in Chinese), 1984, English version, Berlin: Springer, 1995.
GAO ET AL.: COMPLETE SOLUTION CLASSIFICATION FOR THE PERSPECTIVE-THREE-POINT PROBLEM 943

[16] B. Mishra, Algorithmic Algebra, pp. 297-381, Berlin: Springer, 1993. Xiao-Shan Gao received the PhD degree from
[17] D. Wang, “An Implementation of the Characteristic Set Method in the Chinese Academy of Sciences in 1988. He is
Maple,” Automated Practical Reasoning: Algebraic Approaches, a professor in the Institute of Systems Science,
J. Pfalzgraf and D. Wang, eds., pp. 187-201, New York: Springer, Academia Sinica. He has published more than 70
Wien, 1995. research papers and two monographs. His
[18] C.K. Yap, “Robust Geometric Computation,” Handbook of Discrete research interests include: automated reasoning,
and Computational Geometry, J.E. Goodman and J. O’Rourke, eds., symbolic computation, and intelligent CAD and
pp. 653-668, CRC Press, 1997. CAI (computer aided instruction). Web page:
[19] J. Bochnak, M. Coste, and M.F. Roy, Real Algebraic Geometry. http://www.mmrc.iss.ac.cn/~xgao. His a member
Berlin: Springer, 1998. of the IEEE and the IEEE Computer Society.
[20] D.S. Arnon, “Geometric Reasoning with Logic and Algebra,”
Artificial Intelligence, vol. 37, pp. 37-60, 1988.
[21] A.M. Cohen, H. Cuypers, and H. Sterk, Some Tapas of Computer Xiao-Rong Hou received the MS degree in
Algebra, pp. 121-167, Berlin: Springer, 1999. mathematics from the Chengdu Institute of
[22] L. Yang, J.Z. Zhang, and X.R. Hou, Non-Linear Equation System and Computer Applications, Academia Sinica. He is
Automated Theorem Proving, pp. 137-176, Shanghai: Shanghai Press now a professor of mathematics and computer
of Science, Technology, and Education, 1996. science at the University of Science and Tech-
[23] X.S. Gao, J.Z. Zhang, and S.C. Chou, Geometry Expert, Taipai, nology of China and at Ningbo University of
Taiwan: Nine Chapter Pub., 1998 (in Chinese). China. His research interests include symbolic
computation, real algebraic geometry, automated
reasoning, and intelligent software technology.

Jianliang Tang received the MS degree from

GuangXi Normal University of China in 2000.
Currently, he is finishing his PhD thesis at the
Institute of Systems Science, Academia Sinica.
His research interests include camera calibration
and symbolic computation.

Hang-Fei Cheng received the BS degree from the University of Science

and Technology of China in 1995 and the MS degree from the Institute
of Systems Science in 1998. He is now a PhD candidate at
Pennsylvania State University.

. For more information on this or any other computing topic,

please visit our Digital Library at http://computer.org/publications/dlib.

View publication stats