This is an implementation of the Gaussian mixture probability hypothesis density filter (GM-PHD) described in:
B.-N. Vo, W.-K. Ma, "The Gaussian Mixture Probability Hypothesis Density Filter", IEEE Transactions on Signal Processing, Vol 54, No. 11, November 2006, pp4091-4104.
The submission includes a Matlab implementation of the linear Kalman filter GM-PHD filter and the extended Kalman filter GM-PHD filter algorithms described by Vo & Ma, as well as one of the simulated problems described in their paper. A few modifications were made from Vo & Ma's algorithm but they are for technical reasons and don't change the overall structure of the filter.
The GM-PHD filter is a means of estimating the number and positions of targets in measurement data. Its advantages include a representation of target position uncertainty (using a covariance matrix) as well as target existence uncertainty (using a weight) and the absence of data association in the update step. This implementation is quite heavily commented and will probably be helpful to people trying to learn about GM-PHD filtering, but the paper by Vo & Ma is essential to understand what is really going on.
The simulator creates noisy measurements of two moving targets in a cluttered environment, with a third target being created approximately halfway through the simulation. It's a fairly simple problem but it is effective in demonstrating filter performance. Vo & Ma use it to demonstrate the linear GM-PHD filter but I use it for the EKF version as well, rather than coding up the EKF simulator that they describe.
Ba-Ngu Vo has kindly allowed me to include his implementation of the Optimal Subpattern Assignment (OSPA) metric proposed by Schuhmacher et al in:
Schuhmacher, D.; Ba-Tuong Vo; Ba-Ngu Vo, "A Consistent Metric for Performance Evaluation of Multi-Object Filters," Signal Processing, IEEE Transactions on , vol.56, no.8, pp.3447,3457, Aug. 2008.
This isn't necessary for the GM-PHD filter to work but provides a nice way of visualising and quantifying performance. If people want to tweak this code, the metric is a nice way to see how well things are performing and it is fairly easy to modify this implementation to perform multiple test runs with the same data but different filter parameters and compare the metrics.
Read the README.txt, or just start with GM_PHD_Filter.m and work from there. |