Dynamic Artificial Potential Fields for Autonomous Camera Control
Paolo Burelli and Arnav Jhala
Center For Computer Games Research
IT University Of Copenhagen
pabu@itu.dk, arjh@itu.dk
Abstract
Camera control in real-time interactive 3D applications is a
challenging problem. Developing a generalized system able
to produce high quality visual results and smooth camera
movements in dynamic environments remains an open prob-
lem in the research community. In this paper, we describe the
implementation and evaluation of Artificial Potential Fields
for automatic camera placement. We first describe the re-
casting of the frame composition problem as a solution to a
two particles suspended in an Artificial Potential Field. We
demonstrate the application of this technique to control both
camera location and camera aim direction. We show that
this technique can be successfully used to solve both camera
animation and frame composition tasks in dynamic environ-
ments in real-time through an example implemented on an
existing 3D game engine.
Introduction
In interactive 3D applications, such as games, a virtual cam-
era represents the point-of-view of the user. Therefore, cam-
era control is one of the fundamental aspects of user’s in-
teraction with the virtual world. The quantity and quality
of information perceived by the user in a 3D environment is
dependent on the camera control employed by the system.
An autonomous camera control system is a tool that per-
mits the application designer to control these aspects without
pre-scripting the camera parameters at design time. Such a
system enables the designer to define a set of visual require-
ments and automatically configures the camera view accord-
ing to these requirements at run-time.
Several techniques for autonomous camera control have
been proposed in the past (for a comprehensive overview see
(Christie and Olivier 2006)) to address this problem. Most
of these techniques either focus on frame composition, or
on camera motion. While there are efficient approaches for
solving one of these problems at a time, these approaches
are not effective for both composition and movement at the
same time.
We propose a system that employs Artificial Potential
Fields to perform both camera animation and frame com-
position tasks within the same camera framework. Artifi-
cial Potential Fields (APFs) is a well-known technique in the
Copyright c© 2009, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
field of robotics used for robot navigation in dynamic envi-
ronments; it has been applied to camera control (Beckhaus,
Ritter, and Strothotte 2000) to address camera path-planning
in exploratory tasks within virtual environments.
Our system extends this work by modelling the camera
movement and camera orientation parameters, both with
multiple APFs. integrating frame composition concepts to
both camera positionning and orientation. To do this we
first propose a model to recast frame composition require-
ments into an Artificial Potential Fields representation and
then we propose the adoption of multiple potential fields to
control both camera position and camera orientation. We
show that this technique permits the development of an au-
tonomous camera system capable of controlling the camera
in dynamic 3D environments according to frame composi-
tion requirements with real-time performances.
In the remainder of the paper, we first describe our APF
representation of geometric constraint satisfaction. Then,
we show how the proposed system is able to deal with over-
constrained frame requirements (no solution for given con-
straints) and under-constrained frame requirements (multi-
ple camera configurations as solutions to the given con-
straints). We show how this enables the designer to choose
whether camera is fully controlled by the system or can by
partially controlled by the user. Finally, a working example
implementation of the system is described.
Related work
The problem of automatically control the camera in virtual
camera in 3D environments has received significant atten-
tion from the research community (Christie and Olivier
2006). Earliest approaches (Ware and Osborne 1990;
Blinn 1988; Gleicher and Witkin 1992) focused on the map-
ping between the degrees of freedom (DOF) for input de-
vices to 3D camera movement.
Direct control of the several camera’s DOFs showed to be
often problematic for the user (Drucker and Zeltzer 1994) so
researchers started to investigate how to automatically place
and configure the camera. Christie and Olivier (Christie and
Olivier 2006) classify different approaches into three main
categories according to the the modelling algorithm: alge-
braic systems, reactive systems and generalized approaches.
One of the first examples of algebraic systems, was de-
veloped by Blinn already in 1988; it was am automatic cam-
8
Proceedings of the Fifth Artificial Intelligence for Interactive Digital Entertainment Conference

era control system for planets visualization in a space sim-
ulation at NASA (Blinn 1988). Blinn’s work has been fol-
lowed by many others trying to produce more flexible au-
tonomous camera system and to integrate aspects like cam-
era motion and frame composition (Arijon 1991). Gleicher
and Witkin (Gleicher and Witkin 1992) introduced through-
the-lens camera control technique, a reactive technique in-
spired by visual-servoing which permits the user to manip-
ulate the camera by controlling and constrainting projected
images parameters.
As through-the-lens based systems are computationally
efficient, they are ideal for tasks such as object tracking in
interactive applications. Their aim, however, is to maintain
specific image features (i.e. keep an object in the center of
the screen) and require a preliminary camera initialization.
Generalized approaches have modelled camera control as
a constraint satisfaction or optimization problem. These ap-
proaches require the designer to define a set of required
frame properties which are then modelled either as an ob-
jective function to be maximized by the solver or as a set of
constraints that the camera configuration must satisfy.
Optimization based systems, like Halper and Olivier’s
CAMPLAN (Halper and Olivier 2000), always find the best
camera configuration for the given requirements but their
computational cost is high. On the other hand, constraint
fulfilling systems (Jardillier and Langue`nou 1998) are much
more efficient but may not return any result if there is not
configuration respecting all the frame requirement.
Bares (Bares and Lester 1999) addressed the issue by
identifying conflicting constraints and produce multiple
camera configurations corresponding to the minimum num-
ber of non-conflicting subsets. Bourne (Bourne and Sat-
tar 2005) extended Bares’s solution by adding a weight
property to each constraint to define a relaxation prior-
ity. A third set of generalized approaches (Pickering 2002;
Christie and Normand 2005; Burelli et al. 2008), combines
constraints satisfaction to select feasible volumes (therefore
reduce the size of the search space) and optimization to find
the best camera configuration within these spaces.
These approaches on finding the best camera configura-
tion to obtain a specified shot, other researchers have in-
stead investigated the problem of camera animation. Beck-
haus(Beckhaus, Ritter, and Strothotte 2000) employed Arti-
ficial Potential Fields to generate collision-free camera paths
through a virtual environment, Xiao (Xiao and Hubbold
1998) employed force fields to simulate gravity in camera
motion and help the camera controller to avoid collisions.
We extend Beckhaus’s work and apply Artificial Poten-
tial Fields (Khatib 1986), a well established technique for
robot motion planning, to camera control addressing both
camera movement and frame composition problems. Frame
constraints, typical of the generalized approaches, are mod-
elled into potential fields and these are used to compute both
camera position and camera orientation.
CamOn
CamOn is an autonomous camera system capable of gener-
ating smooth camera animations and solving camera com-
position tasks. The system iteratively animates the camera
to converge to an optimal camera position, at each iteration
CamOn takes the current camera position, frame description
and scene description as input, and returns a new camera
position as output.
Frame description requires identification of the frame sub-
jects, definition of subject importance and definition of com-
position rules. Composition defines disposition of visual el-
ements in an image (Arijon 1991); following the model pro-
posed by Bares (Bares et al. 2000) we have translated these
composition rules into soft constraints.
Current version of the system supports three constraints:
• Visibility
This constraint defines what fraction of the subject must
be visible in the frame, fraction is defined as a number
between 0.0 if subject is not visible to 1.0 if subject is
fully visible.
• ProjectionSize
This constraint defines the size of subject in the frame,
size is defined as the quotient between frame height or
width and the relative longest side of the subject’s pro-
jected bounding box.
• ViewAngle
This constraint defines the angle from which the camera
should shoot the subject, angle is defined using spherical
coordinates.
Each of these constraints for each subject and the scene
geometry is modelled into the system using Artificial Poten-
tial Fields. Camera position and look-at point are modelled
as particles moving along the potential field so created this
way.
Artificial Potential Fields
Artificial Potential Fields (Khatib 1986) is an iterative tech-
nique commonly adopted in the area of robotics for con-
trolling the navigation of robots in dynamic environments.
Robots are modelled as particles moving in a field of po-
tentials attracted by low potential areas; the position to
be reached generates an attractive force (a low potential
zone) and obstacles generate repulsive forces (high potential
zones). At each iteration the particle moves along the force
resulting from the sum of all repulsive (obstacle avoidance)
and attractive (goals) forces influencing current particle po-
sition; the particle continues to move until it reaches a stable
state.
CamOn system employs Artificial Potential Fields to
move camera and solve frame composition tasks: obsta-
cle avoidance is modelled using repulsive forces and frame
composition is obtained by translating frame constraints into
forces affecting both the position and the look-at point of the
camera.
Each frame constraint produces one force attracting or re-
pulsing camera position and one force attracting or repulsing
camera look-at point, the system treats these two aspect of
the camera as two different particles moving into two differ-
ent potential fields. An example of the two potential fields
created by a frame constraint (the target sphere has to be
fully visible in the projected frame) can be seen in Figure
9

Figure 1: Example APF produced by a visibility constraint
on a sphere (from left to right: sphere, position APF, orien-
tation APF)
1, the two potential fields shown are a sample of the 3D
field measured along the horizontal plane passing through
the sphere centre. The particle is attracted by the low po-
tential (white) areas. The two potential fields are linked by
constraint satisfaction values as it is explained more in depth
in the next section. At each iteration CamOn moves both
camera’s position and orientation; this makes constraint sat-
isfaction values change concurrently and the fields dynami-
cally change during camera movement according to the new
particles positions.
CamOn handles over-constrained and under-constrained
tasks effectively. When there is no camera configuration
that can satisfy all the constraints imposed to the camera
(over constrained task) the system converges to the nearest
configuration which optimizes frame requirements (giving
higher priority to more important subjects). When more than
one configuration potentially satisfies frame requirements
the system converges to one of the feasible configurations
but lets the camera freedom of movement within the limits
imposed by frame constraints. This feature permits to let the
user to control the camera without loosing the possibility to
impose some constraints on camera position, orientation and
movement.
Recasting frame constraints into potential fields
Translating frame constraints into potential fields requires
identification of position and orientation goals correspond-
ing to each frame constraint.
Figure 2: Example APF produced by a two front-shot frame
description (from left to right: spheres, position APF, orien-
tation APF)
Ideal camera positions and orientations for each constraint
are modelled as low potential zones; any other part of the
solution space has a potential proportional (the exact relation
can vary from constraint to constraint) to the distance from
the ideal position and to the constraint satisfaction of the
corresponding camera configuration.
The forces imposed by each constraint are defined as fol-
low:
• Visibility
Assuming α as the current visible fraction value and αexp
the desired one, A as the camera position and B as subject
position, the force Vpos applied to the camera position is:

Vpos(

B) =
{
(
B− A)norm(α−αexp)
(1−αexp)
α < αexp
0 α ≥ αexp
(1)
Assuming α as the current visible fraction value and αexp
the desired one, A as the camera view direction and B as
subject position relative to the camera, the force Vorient
applied to the camera look-at point is:

Vorient(

B) =
⎧
⎨
⎩
(
B− A)norm(α−αexp)
1−αexp
α < αexp
(
A− B)norm(αexp−α)
αexp
α ≥ αexp
(2)
• ProjectionSize (only position)
Assuming α as the current projection size value and αexp
the desired one, A as the camera position and B as subject
position, the force Vpos applied to the camera position is:

Δ =
{
(

B −

A) α < αexp
(

A −

B) α ≥ αexp

Vpos(

B) =
⎧
⎨
⎩

Δnorm(
|

Δ|(α−αexp)
1−αexp
)
2
α < αexp

Δnorm(
|

Δ|(αexp−α)
αexp
)
2
α ≥ αexp
(3)
• ViewAngle (only position)
Assuming α and θ as the current view angle of the camera
and αexp and θexp the desired ones, A as the camera po-
sition, B as subject position, F as the normalized subject
front vector and Frot(α,θ) as the desired view direction
(front vector rotated by α and θ), the force Vpos applied
to the camera position is:
λ =
1
2
(
|αexp − α|
π
+
|θexp − θ|
2π
)

Aexp = |

A|

Frot(α,θ)

Vpos(

B) = (

Aexp −

A)norm(λ|

Aexp −

A|)
2 (4)
Force value at each point is described as a linear combina-
tion of scalar weighted forces, where each force corresponds
to a frame constraint and each scalar value to the relative
subject importance; the resulting values define the gradients
of the potential field.
Figure 2 shows the potential field associated to a frame
composed by two spheres with equal importance value both
with Visibility set to 1, ProjectionSize set to 1 and ViewAn-
gle set to (0,0).
10

(a) Camera facing another
robot
(b) Camera orients towards
the target
(c) Camera approaches the
target maximizing size
(d) Camera aligns to the
right view angle
Figure 3: Camera converging to a right-profile shot of a robot (listing 1)
Implementation and Experimental Results
CamOn is currently implemented as a library of classes for
camera control related functions in Ogre3D 1. The library
includes functions for setting camera parameters, setting
constraints, and automatically computing the motion of the
camera. A set of desired camera parameters (location, ori-
entation, and field-of-view) identifies a camera profile and
it is defined as a set of subjects (each with an importance
value) and a set of viewing constraints applied to each sub-
ject. Viewing constraints take the form of desired composi-
tion parameters like primary object, object occlusion, etc. At
each frame the system computes the current camera position
and orientation velocities which are then used to compute
the new camera configuration; the two velocities are com-
puted at each iteration with a 4th order Runge-Kutta. The li-
brary facilitates change of the number of iterations per frame
and the Runge-Kutta speed factor to tune camera speed.
A small demo application has been developed to test Ca-
mOn behaviour and performance in a game-like 3D envi-
ronment. The application shows a non-interactive intro se-
quence showing the main character (a green ninja) and the
enemies he has to avoid (4 steady robots and 2 walking
robots). The user can watch the default camera animation
(a predetermined sequence of camera profiles) or can inter-
actively decide to select a specific camera profile for com-
posing a shot of any of the characters in the scene. Camera
position and aim direction can be modified by the user con-
tinuously during the demo and the freedom of control de-
pends on the currently loaded camera profile. The camera
definition file included in the demo can also be modified to
test the camera animations corresponding to different con-
straint values and weights.
In the next sections we show how the camera moves and
satisfies a simple camera profile with one subject and how
camera deals with moving subjects and over-constrained re-
quirements. We’ll explain CamOn behaviour with under-
constrained frame requirements and finally we show sys-
tem’s run-time performance.
Camera movement
Figure 3 shows how CamOn converges to a right-profile shot
of a target robot starting from another shot (figure 3(a)); the
1Object-Oriented Graphics Rendering Engine -
http://www.ogre3d.org
camera requires approximately 3 seconds to reach the steady
configuration (figure 3(d)).
Visibility is the first constraint to be satisfied (figure 3(b)),
meanwhile camera approaches the target subject to satisfy
ProjectionSize (figure 3(c)) constraint and finally moves to
the stable configuration to satisfy also the ViewAngle con-
straint (figure 3(d)).
The sequence depends on the function used to model dif-
ferent constraint forces and especially on the size of their
area of interest.
Listing 1: Right-Profile shot camera profile
<s u b j e c t name=”Robot1 ” impo r t a n c e =” 1 . 0 ”>
<v i s i b i l i t y v i s i b l e F r a c t i o n =” 1 . 0 ” />
<viewAngle h o r i z o n t a l =” 100 ” v e r t i c a l =” 45 ” />
<p r o j e c t i o n S i z e s i z e =” 1 . 0 ” />
</ s u b j e c t>
The time required by the camera to reach the optimal con-
figuration, within performance limits, depends primarily on
the chosen animation speed and the starting camera position.
Performance limits are directly connected to the inte-
gration method used to search the optimal configuration
through the potential fields. Increasing the convergence
speed increases also computational error (in the fourth-order
RungeKutta method the total accumulated error has order
h
4 where h is the convergence speed factor). On the other
hand, increasing the number of iterations per frame leads to
a performance drop, so the speed limit depends on the trade-
off between camera placement accuracy and system perfor-
mance.
Over-constrained profile
Figure 4 shows how CamOn deals with the camera require-
ments defined in listing 2 in a dynamic scene. The cam-
era system is instructed to generate and maintain a three-
quarter front shot of the two walking robots (4(a)), however,
as robots during their patrol face opposite directions (4(b)),
the problem becomes over-constrained and no camera con-
figuration can satisfy the requirements anymore.
Listing 2: Three-quarter front two-shot profile
<s u b j e c t name=” WalkingRobot1 ” impo r t a n c e =” 1 . 0 ”>
<v i s i b i l i t y v i s i b l e F r a c t i o n =” 1 . 0 ” />
<viewAngle h o r i z o n t a l =”−30” v e r t i c a l =” 10 ” />
<p r o j e c t i o n S i z e s u b j e c t =”RobotW1” s i z e =” 1 . 0 ” />
</ s u b j e c t>
11

(a) Robots are distant but
face the same direction
(b) One robot changes sud-
denly walking direction
(c) Robots get near (d) Robots get distant again
facing opposite directions
Figure 4: Camera searching an optimal configuration for a three-quarter front shot of two walking robots (listing 2); a 2D
sample of the corresponding position potential field is shown below each frame with camera represented as a small sphere
<s u b j e c t name=WalkingRobot2 impo r t a n c e =” 1 . 0 ”>
<v i s i b i l i t y v i s i b l e F r a c t i o n =” 1 . 0 ” />
<viewAngle v e r t i c a l =”−30” v e r t i c a l =10/>
<p r o j e c t i o n S i z e s i z e =” 1 . 0 ” />
</ s u b j e c t>
The camera starts searching for a new stable configura-
tion while the potential fields change dynamically according
to the robots movement; figure 4(c) shows how camera fully
satisfies the Visibility and ProjectionSize constraints and fig-
ure 4(d) shows the camera satisfying only the visibility con-
straints due to the distance between the two robots.
Under-constrained profile
Listing 3 shows a profile of the same two robots in figure
4 which doesn’t specify any preference about view angle.
This is an example of under-constrained composition prob-
lem since there are several camera configurations that satisfy
frame requirements.
Listing 3: Under-constrained two-shot profile
<s u b j e c t name=” WalkingRobot1 ” impo r t a n c e =” 1 . 0 ”>
<v i s i b i l i t y v i s i b l e F r a c t i o n =” 1 . 0 ” />
<p r o j e c t i o n S i z e s u b j e c t =”RobotW1” s i z e =” 1 . 0 ” />
</ s u b j e c t>
<s u b j e c t name=WalkingRobot2 impo r t a n c e =” 1 . 0 ”>
<v i s i b i l i t y v i s i b l e F r a c t i o n =” 1 . 0 ” />
<p r o j e c t i o n S i z e s i z e =” 1 . 0 ” />
</ s u b j e c t>
In this case the camera control system converges from any
starting configuration to the nearest configuration satisfying
all the constraints; the user can move both camera position
and orientation while the system continuously limits camera
movements to keep constraints satisfaction.
Performance
The experiment was run on a Apple Macbook equipped with
an Intel Core 2 Duo (2.0GHz) and 4Gb of RAM (DDR3 1.07
GHz) running MacOS 10.5.6; the demo was implemented
using Ogre3D version 1.6 (Einhort).
Table 1 shows average execution time required to com-
pute the force related to each implemented constraint
(ViewAngle and ProjectionSize constraint don’t generate
any orientation force). CamOn computes at each frame the
sum of the forces on the current particle position for both
camera location and orientation; this is done 4 times due to
the 4th order Runge-Kutta integration.
Constraint Average time (ms)
ProjectionSize (Position) 0.08647
ViewAngle (Position) 0.02401
Visibility (Position) 0.04477
Visibility (Orientation) 0.04350
Table 1: Constraints execution times collected in approxi-
mately 10,000 runs
The shot shown in figure 3, with one subject and three
constraints, requires approximately 0.86 milliseconds per
frame; even with multiple subjects and multiple constraints
per subject (a shot with 10 subjects would not take more
than 9 ms per frame), the system maintains real-time perfor-
mance.
Conclusions and future work
This paper describes CamOn, an autonomous camera con-
trol system base on Artificial Potential Fields. The system
solves camera motion and frame composition problems by
extending Beckhaus work on Artificial Potential Fields with
frame constraint based forces.
Two different potential fields are generated from scene’s
geometry and frame constraints; camera position and cam-
era look-at point are modelled as particles moving through
the two potential fields. As potential fields don’t need to be
precomputed and can be calculated locally at each iteration
step, CamOn performances are independent from scene size.
12

Further, the camera converges to a solution from any initial
configuration and can converge to a solution automatically
even if the user takes over control in the middle of the con-
vergence process.
While CamOn satisfies all the four major requirements
mentioned by Bourne (Bourne and Sattar 2005), viz. auton-
omy, reactivity, real-time performance, and dynamic envi-
ronment handling, there are many aspects that can be still
improved. Cinematic camera movement principles could be
incorporated to influence camera motion. CamOn’s current
implementation only supports basic object view constraints,
more composition constraints, like occlusion or in-frame po-
sition, could be introduced to match the expressiveness of
state-of-art composition systems such as the one by Bares
et. al. (2000). Our current implementation of Artificial Po-
tential Fields suffers of local minimum problem which could
occur in complex environments like mazes and lead to un-
desirable local minima with occluded views.
This issue might be solved using multiple particles placed
near the attractive points and making the camera follow the
one converging to the best value. A Global Optimization al-
gorithm could also be run parellelly to to Artificial Potential
Fields based algorithm to tune the forces according to the
global optimum as soon as this is found.
This paper demonstrates the use of APFs for camera com-
position and movement in real-time interactive 3D virtual
environments. Initial results of this implementation are en-
couraging. One of the long-term goals of this implementa-
tion is to provide designers with editing interfaces for defin-
ing APF profiles for different objects in the environment.
Rather than modifying constraint values between 0 and 1,
the colored visual representation of a potential field might
be more accessible to designers. Potential fields for collision
of the camera with dynamic world objects can also be auto-
matically extracted from geometry information available in
the spatial data-structure of the game engine.
References
Arijon, D. 1991. Grammar of the Film Language. Silman-
James Press LA.
Bares, W. H., and Lester, J. C. 1999. Intelligent multi-shot
visualization interfaces for dynamic 3d worlds. In IUI ’99,
119–126. ACM.
Bares, W.; McDermott, S.; Boudreaux, C.; and Thainimit,
S. 2000. Virtual 3d camera composition from frame con-
straints. In MULTIMEDIA ’00, 177–186. ACM.
Beckhaus, S.; Ritter, F.; and Strothotte, T. 2000. Cubi-
calpath - dynamic potential fields for guided exploration in
virtual environments. In PG ’00. IEEE Computer Society.
Blinn, J. 1988. Where am i? what am i looking at? IEEE
Comput. Graph. Appl. 8(4):76–81.
Bourne, O., and Sattar, A. 2005. Applying constraint
weighting to autonomous camera control. In AIIDE ’05,
3–8.
Burelli, P.; Di Gaspero, L.; Ermetici, A.; and Ranon, R.
2008. Virtual camera composition with particle swarm op-
timization. In Smart Graphics, 130–141. Springer-Verlag.
Christie, M., and Normand, J.-M. 2005. A semantic space
partitioning approach to virtual camera composition. Com-
puter Graphics Forum 24(3):247–256.
Christie, M., and Olivier, P. 2006. Camera Control in Com-
puter Graphics. 89–113. Eurographics Association.
Drucker, S. M., and Zeltzer, D. 1994. Intelligent camera
control in a virtual environment. In Graphics Interface 94,
190–199.
Gleicher, M., and Witkin, A. 1992. Through-the-lens cam-
era control. In Computer Graphics, 331–340.
Halper, N., and Olivier, P. 2000. Camplan: A camera plan-
ning agent. In Smart Graphics 2000 AAAI Spring Sympo-
sium, 92–100.
Jardillier, F., and Langue`nou, E. 1998. Screen-space con-
straints for camera movements: the virtual cameraman.
Computer Graphics Forum 17(3):175–186.
Khatib, O. 1986. Real-time obstacle avoidance for manip-
ulators and mobile robots. Int. J. Rob. Res. 5(1):90–98.
Pickering, J. 2002. Intelligent Camera Planning for Com-
puter Graphics. Ph.D. Dissertation, University of York.
Ware, C., and Osborne, S. 1990. Exploration and virtual
camera control in virtual three dimensional environments.
SIGGRAPH 24(2):175–183.
Xiao, D., and Hubbold, R. 1998. Navigation guided by ar-
tificial force fields. In CHI ’98, 179–186. Addison Wesley.
13