Extended target tracking using principal components

The increased resolution in today's radar systems enables tracking of small targets. However, tracking both small and large targets in a dense target scenario raises considerable challenges. The data association of tracks to measurement groups is highly dependent on good target extension models for filtering and likelihood computation. In our attempt to design a tracker for extended targets, we start by adopting the results from the technique referred to as random matrices, which enables us to separate the filtering into an extension and a kinematical part. We re-define the measurement model and discard the assumption of independent Gaussian-distributed plots. Instead we assume the principal components to be Gaussian distributed. Then, through a heuristic approach, we create a two-stage Kalman filter, where the first stage estimates the principal components, and the second stage estimates the centre of gravity, using the output from the first stage as measurement uncertainty. The advantage of having a Kalman filter with data-driven measurement noise over a standard Kalman filter is demonstrated using simulated data, where a significant improvement in kinematical accuracy is shown.