Target localisation and tracking in a UWB radar network

Abstract

For both security and analytics, much research has gone into person tracking already. As a result, many
different state of the art technologies exist. However, in darkness or without a direct line of sight, much
less technologies are capable of this. The choices become especially limited when the setup needs to
be portable.
A method for person localisation and tracking is implemented. This method consists of a localisation
part, which works with any range-based detection method. Least square estimation is used to determine
the location from the radar detections. With two or more people, it is mathematically impossible to
distinguish which locations are correct, if only the current measurement is taken into account.
Thus, the first problem to be solved is connecting ranges to targets. This is done using target association.
After this is done, one-dimensional tracking can track people at lower computational cost.
The tracking is both in one dimension (per-radar) and in two dimensions. The Hungarian algorithm
is used for keeping track of people using a Kalman filter. The Kalman filter considers the predicted
next location and the measured next location, and makes a best guess. A neural network was used for
the optimisation of location-specific noise parameters, something that has not been done before in this
context. Single person tracking and two person tracking works as expected. The tracking is relatively
cheap in terms of computational complexity. While the tracking has no limits on the maximum number
of people present, the localisation gets increasingly difficult with a complexity of O (n^n). Detecting
the correct peaks is a non-trivial problem because of multi-path reflections. In combination with UWB
radar detections, single and dual person tracking in a room is achieved. More people can be handled by
the tracking algorithm, which is detection-method-agnostic, but not by the localisation. There is some
room for improvement in the dual and triple-person case. However, going further than this is currently
unfeasible, because of the many reflections that occur. Furthermore, the large amount of possible person
locations also has an effect. This is a problem that scales with O (n^n) where n is the amount of
targets.