4

Collision Avoidance
The UAV may be assigned either to map or to avoid obstacles. Here we limit
the discussion to avoiding obstacles, as mapping is beyond the scope of the
present book. An ‘obstacle’ will be defined as any object in the environment
that the UAV has to avoid. These can be moving or fixed. Fixed obstacles
could be objects such a buildings, hills, forests, etc. that will intercept the
normal flight path of the UAV. They may also be clear areas that the UAV
may not fly into. These would include controlled airspace or dangerous areas.
Moving obstacles constitute other aircraft or UAVs that occupy the same
airspace and whose trajectories will intercept the UAV flight path. These
moving obstacles could be part of the UAV group that would be expected
to cooperate with the UAV, but they could also be non-cooperating vehicles
that are flying through the airspace occupied by the UAV.
Obstacle avoidance will normally be in the form of a software program
within the UAV combined with a set of sensors that monitor the UAV’s
environment for fixed or moving obstacles. If detected, the software will then
modify the UAV path if a future collision is detected.
A collision between two objects occurs when they try to occupy the same
point at the same time. As the collision needs to be avoided, a prediction
mechanism is required to provide enough lead time to take evasive action. As
stated, there are two categories of obstacle: fixed and moving. The first case of
fixed obstacles deals with path planning of a single UAV in an environment
of static obstacles, where most, if not all, of the obstacles’ locations and
sizes are known in advance. In the second case of moving obstacles, two
possibilities arise: (i) negotiating static obstacles with more than one UAV,

Cooperative Path Planning of Unmanned Aerial Vehicles

Antonios Tsourdos, Brian White and and Madhavan Shanmugavel
© 2011 John Wiley & Sons, Ltd. ISBN: 978-0-470-74129-0
82 Cooperative Path Planning of Unmanned Aerial Vehicles

where each UAV has to avoid collision with other UAVs (inter-collision
avoidance between UAVs); or (ii) negotiating dynamic obstacles with one
or more UAVs, where each UAV has to avoid non-cooperating vehicles in
addition to inter-collision avoidance.
The case of static known obstacles is normally handled by global path
planners such as Voronoi diagrams, cell decomposition, the establishment
of potential fields, or global optimization techniques. The approach taken
in this book is to build on the path forms already used for path planning,
such as Dubins and PH paths, and to adapt them to negotiate obstacles. This
problem is relatively easy to handle, as the obstacles are fixed with respect
to place and time, whereas in the case of dynamic obstacles, the trajectories
need to be estimated prior to avoidance path planning. However, in both
cases, the presence of obstacles causes not only re-planning of paths, but also
the management of communications, task assignment, resource allocation
and other functions, depending on the mission. These issues are important,
but are beyond the scope of this book. Suffice it to say that the use of efficient
path planning is a core requirement for these issues.
The basic equation of path planning (6.4) for an obstacle-rich environment
is now modified to
r(q) 
Ps −→ Pf , safe . (4.1)

This effectively states that the path from the start pose Ps to the finish pose
Pf should be safe and so will avoid all fixed and moving obstacles. It will
be assumed that all fixed obstacles have known positions and all mobile
obstacles will have known position and velocity (speed and direction).
When encountering an obstacle, the direction of manoeuvre in 2D is
restricted to manoeuvring to either the left or the right of the obstacle, as
shown in Figure 4.1. However, the number of choices is greater in three
dimensions. For example, if we extend the circle into a sphere, the number
of choices is infinite in that any path that manoeuvres around the sphere is a
valid path.
Apart from the locations and mobility of the obstacles, the other factors
that make the problem complex are the size and shape of the obstacles.
Research on the avoidance of convex polygonal obstacles is ongoing in
the area of computational geometry, and will not be dealt with in this
book, although this problem is very similar to the approach taken here.
Producing curvature-constrained paths among the polygonal obstacles is
NP-hard (non-deterministic polynomial-time hard) (Agarwal et al. 2002; Reif
and Wang 1998). For the treatment in this book, it will be assumed that the
Collision Avoidance 83

Figure 4.1 Obstacle avoidance in 2D

obstacles are surrounded by a safety circle or sphere, which determines the

safe boundary of the obstacle for path planning purposes.
This chapter deals with collision avoidance of mapped obstacles, where
the obstacles are static and known a priori. In the first case, that of static
obstacles, two main approaches are used to re-plan the path in 2D. The
first is again by analytical geometry, where the obstacle is avoided either by
increasing the curvature of the path or by creating an intermediate waypoint
and creating a path with several segments. The curvature approach is based
on the fact that the curvature and the torsion determine the path in space. In
the case of 2D Dubins and clothoid composite paths, the start and finish radii
are increased to meet the safety conditions using equation (2.13b). A safe
2D PH path is produced by increasing the boundary curvatures: equations
(2.95) and (3.48a), and equations (2.96) and (3.48b). The offset PH path is
used to represent the safety boundary on either side of the flight path.
Differential geometry is also presented to extend the Dubins paths to include
the obstacles in the path structure.
The next chapter discusses dynamic obstacle avoidance and resolu-
tion conflicts, where differential geometry is used to predict the mobile
obstacle trajectory and to develop geometries that produce trajectories to
avoid collision.

4.1 Research into Obstacle Avoidance

The obstacle avoidance problem is closely associated with path planning

because the presence of obstacles usually results in the re-planning of paths.
Hence, a brief review of approaches to obstacle avoidance is useful here.
84 Cooperative Path Planning of Unmanned Aerial Vehicles

The configuration space principle illustrated in Lozano-Pérez (1983) for

robotics applications reduces the size of the robots to point masses and
compensates for this reduction by increasing the size of the obstacles. This
approach enables efficient use of algorithms such as Voronoi diagrams (Beard
et al. 2002; Bortoff 2000; Li et al. 2002; McLain 2000), visibility diagrams and
probabilistic methods to model the given map of landmarks and obstacle
locations into a searchable database (see Chapter 1). The databases are graphs
and trees forming a network of nodes and edges isolating the obstacles. A
search algorithm is used to connect the start and finish points through the
nodes of the network. The A∗ algorithm is used in Yang and Zhao (2004)
to find the path in the midst of known obstacles and conflicts. However,
the resultant path may not be flyable, as the curvature constraint on paths
is not taken into account. A hybrid approach in which a route planned for
the known obstacles is deformed to handle the unknown obstacles is used
in Lamiraux et al. (2004). A similar method of deforming a planned path by
varying its curvature is seen in Shanmugavel et al. (2005). However, a reactive
method, responding to sensor feedback, is essential to handle fast-moving
dynamic obstacles.
Obstacle avoidance methods have been studied using potential fields,
where attractive and repulsive forces, respectively, are generated by the goal
and obstacles (Kim and Khosla 1992). The potential field is then used to avoid
the obstacles in path planning of multiple UAVs by flying down the valleys
of the potential field. The obstacles are assumed to be circular in shape, with
a safety circle surrounding them (Eun and Bang 2006).
An analytical and discrete optimization approach is used in Zabarankin
et al. (2002) for optimal risk path generation in 2D space, with a constraint on
path length. Mixed integer linear programming (MILP) is used in Richards
and How (2002) for collision avoidance. The resulting path from both of these
approaches does not take into account the curvature constraint, and so extra
work has to be done to produce a flyable path. Avoidance circles are used for
obstacle avoidance of multiple mobile robots in Fujimori et al. (2002), where
two robots cooperate within a specified area, while others wait outside the
area. However, what is feasible for mobile robots, where the robot can stop
and perform pivot turns, is not feasible for UAV applications, where the
UAV cannot stop and where they have curvature constraints imposed on
their paths.
Path planning with obstacle avoidance in 3D is more complicated. The com-
plication arises because there are infinitely more directions for manoeuvres
for both the UAV and the obstacle. Also, other constraints, especially gener-
ating the shortest path length, is more complicated in 3D space. An octree
Collision Avoidance 85

representation of 3D space has been proposed in Kitamura et al. (1995, 1996),

where the space is divided into searchable regions or cells and the collision-
free path is generated by applying potential fields to each cell of the octree.
The obstacle avoidance problem has been studied in different applications:
to minimize the risk of aircraft detection by radars, sensors or surface-to-air
missiles (SAMs) (Chan and Foddy 1985; Hebert et al. 2001; Vian and More
1989; Zabarankin et al. 2002), to minimize the risk of submarine detection by
sensors (Washburn 1990), or to maximize the probability of target detection
by a searcher (Assaf and Sharlin-Bilitzky 1994; Benkoski et al. 1991; Eagle and
Yee 1990; Koopman 1980; Stone 1975; Thomas and Eagle 1995; Washburn
1983). Another application area in this research field is the use of UAVs
in civilian airspace, where a collision avoidance algorithm equivalent to
that available in commercial aircraft is required. The existing system uses
the ‘Traffic alert and Collision Avoidance System’ (TCAS), which issues
resolution advisories (RAs) based on the position and altitude data from
transponders aboard the aircraft and from the ground sensors, if available.
The safety aspects of using UAVs in civilian airspace is discussed in Dowek
et al. (2001) and Zeitlin and McLaughlin (2007). The generation of cooperative
trajectories for air traffic is discussed in Bicchi and Pallottino (2000). This
is an active research area and will need to be developed to a point where
such algorithms are robust and accepted before UAVs can fly in controlled
airspace with civilian aircraft.

4.2 Obstacle Avoidance for Mapped Obstacles

In order to manoeuvre in an environment that is cluttered with obstacles,

the UAV needs to be able to compute trajectories that will circumnavigate
these obstacles and still arrive at the final pose and position. Hence there
is a requirement to compute composite Dubins trajectories that include
avoidance manoeuvres.
Consider the simple case of a single obstacle that intersects the straight-
line segment of the standard Dubins trajectory, as shown in Figure 4.2. The
Dubins trajectory is now made up of a five-segment trajectory consisting of
initial, intermediate and final arc manoeuvres, together with two straight-line
segments. The initial and final arc manoeuvres are modified to make the line
segments tangent to the obstacle circle. There are also two possible solutions:
a positive and negative rotation arc around the obstacle. Hence two avoidance
trajectories are required to be calculated for every obstacle. Two possibilities
occur for this case: the obstacle intersects the straight-line segment of the