Global Search for Occlusion Minimisation
in Virtual Camera Control
Paolo Burelli, Student Member, IEEE and Georgios N. Yannakakis, Member, IEEE
Abstract— This paper presents a fast and reliable global-
search approach to the problem of virtual camera positioning
when multiple objects that need to be within the reach of the
camera are fully occluded. For this purpose, a comparative anal-
ysis of global-search algorithms is presented for the problem of
maximising camera visibility across different tasks of varying
complexity and within different real-time windows. A custom-
designed genetic algorithm is compared to octree-based search
and random search and results showcase the advantages of the
genetic algorithm proposed with respect to efficiency, robustness
and computational effort.
I. INTRODUCTION
Camera control is a vital component of player experience
and enjoyability in games [23]. A camera in games provides
the player with a means for exploring the game world, getting
feedback on her actions, and updating the state of the game.
Given its critical importance in 3D virtual environments
camera control may provide means of challenge for and
justifiability of artificial intelligence. Camera settings for
games are usually predefined by designers and potential
camera visibility problems (e.g. in the case of occluded
objects) are tackled via simple, nevertheless, unjustifiable and
unrealistic solutions — i.e. the camera rapidly jumps to a
selected non-occluded position towards the closest object of
interest.
In this paper we propose a top-down global search ap-
proach to the problem of finding a non-occluded point when
in-game objects of interest are fully occluded. Occlusion oc-
curs when points (or objects) of interest that the camera needs
to look at are fully, or partially, hidden by objects or walls
of the designed scene geometry. Under these circumstances,
the camera controller is required to find a path towards
an occlusion-free point (ideally a fully non-occluded point)
within a realistic time frame. If the controller is unsuccessful
within a short time window (e.g. 200ms maximum), the
reliability of game camera control is challenged and any
immersion emerged by cinematographic game experience is
lost.
The problem of camera visibility is challenging because
(a) visibility heuristics are computationally expensive to cal-
culate in real-time, and (b) the generated visibility function
terrains are very rough for a search algorithm to explore.
In particular, in order to evaluate the visibility “goodness”
of a camera position, rays need to be casted towards the
object(s) of interest; the designer often has to sacrifice
Authors are with the Center for Computer Games Research, IT University
of Copenhagen, Rued Langgaards Vej 7, DK-2300 Copenhagen S, Denmark.
Emails: {pabu, yannakakis}@itu.dk
visibility accuracy due to the computational cost of visi-
bility estimation in real-time. Moreover, in the majority of
occlusion situations met in computer games the objects of
interest are much smaller than scene geometry features. This
suggests that objects are either occluded or non-occluded;
the situations where objects are partially occluded are rare.
On that basis, visibility fitness generates rough search spaces
the vast majority of which is covered by multimodal fitness
plateaus of near-zero gradient.
Gradient search [7] and local search (e.g. artificial potential
fields [9]) are bound to fail in such problems. Instead, robust
and efficient global search algorithms are required to allocate
fully, or partially, un-occluded camera positions. This paper
introduces a comparative study of search algorithms for solv-
ing the problem of visibility occlusion in camera control. A
custom-designed genetic algorithm (GA) for the investigated
problem, octree-based search [1] (best-first and depth-first)
and random search are evaluated in occlusion problems of
increasing complexity and their speed is evaluated across
different time windows. Algorithm performance is accessed
via the amount of object visibility of the generated solutions,
the number of times a global maximum is found and the time
it took the algorithm to find the maximum.
Results show that the GA proposed demonstrates robust-
ness and real-time efficiency across four dissimilar occlusion
case studies containing three objects of interest and varying
in complexity. It is also apparent that the GA approach
performs well consistently with respect to real-time perfor-
mance in all case studies examined. Random search performs
well in complex problems but it performs poorly in simpler
problems. Finally, octree-based search is outperformed by
the aforementioned algorithms and performs well in simple
problems only.
This paper is innovative in that it introduces an efficient
and reliable GA solution to the problem of full occlusion in
camera control; it examines complex case studies of multiple
(three) objects of interest; and it provides a comparative
analysis of search algorithms (including genetic and random
search) with respect to problem complexity and real-time
performance.
II. BACKGROUND
The problem of automatically controlling the camera in
virtual 3D environments has recently received significant
attention from the research community [11]. The majority
of the earliest approaches to camera control [24], [6], [13]
focus on the mapping between user input and the degrees
of freedom (DOF) of the camera in the 3D space. Direct