EPSRC Industrial CASE Award
(with Neil Cade, Selex Galileo, Luton)
Collective Tracking without Data Association
Summary
The project is concerned with a disruptive approach to data processing with wide application but is particularly aligned
to the autonomous systems priority in the Defence and Aerospace sector, where sense has to be made (without
human intervention) of sparse data from multiple sensors.
Target tracking is generally seen a necessary part of sensor systems in such applications as crime prevention, defence and
border surveillance. The sensors are used to provide
measurements at time intervals and the tracking task is to estimate its
location and dynamics. If there are multiple targets then the idea is usually extended by assigning a single-target tracking
filter to each target and using a data association technique to assign the correct measurement to each filter. Although conceptually
simple, the variable numbers of targets, missing measurements and the exponential proliferation of associations
make this approach impractical in all but the simplest of problems. Given the technical difficulty, the question arises as
to whether tracking each target is necessary or even desirable: a fleet of ships, a crowd of football supporters or convoy
of Lorries might better be described by their collective dynamics. The proposed research takes as its starting point the
impracticality of associating data with individual targets and addresses the problem of what properties of collectives can
be tracked without association and the algorithmic approaches that might be used.
The idea of tracking without association is radical in itself and therefore requires at least the identification of a specific
mathematical direction to be followed. Thus the proposed project will consider a random-set approach, where, by design,
report-to-track association is not observable. This idea has been the basis for such association-freemulti-target filtering algorithms
as the particle-PHD filter and Gaussian mixture PHD filter. These PHD approaches have been applied to tracking
of vehicles, passive radar targets, mines, objects in images sequences and time-varying numbers of speakers. Typically,
good tracking performance been demonstated at much lower computational cost than with conventional
methods. Despite
the approach being originally devised to address non-conventional group tracking problems there have been no practical
implementations of these techniques in such domains. This proposal is therefore concerned with exploring the application
of the random-set (PHD) approach to understanding passive acoustic and EM sensor data (core SELEX S&AS business) with Daniel Clark at Heriot Watt who is the leading UK researcher in random-set multi-object tracking. These specific
data sources are rather different from the video data processing currently addressed by such PHD filtering approaches.
More significantly, the data sources of interest to SELEX S&AS are intrinsically difficult to track and the methods to be
explored may well offer the only solutions to some of these tracking problems.