The Random Set Filtering Website

UK Mirror|   Australia Mirror

PHD Filter Multi-Target Tracking Demonstrations

Here we demonstrate some examples of tracking with Probability Hypothesis Density (PHD) filters on both simulated and real data to illustrate the ability of these algorithms to identify (detect) targets, initiate tracks and terminate them when targets leave the scene. The number of targets varies over time and the algorithms also estimate the number of targets. This is in the presence of clutter (false alarms), where there are many more false measurements than those generated by true targets. The tracking videos are grouped into different sections according to the implementation or application.

Gaussian mixture PHD Filter

GM-(C)PHD filter tracking |   Sonar |   Video |   Radar

Demonstrations

Filtering |   Tracking |   Jump Markov Model

Gaussian mixture PHD filter

Gaussian mixture PHD filter Estimation

The first example shows the intensity function across the state space (blue surface with peaks to indicate likely targets), measurements (red stars), true target positions (white crosses) and estimated target positions (green circles). In the Gaussian mixture implementation, we are able to maintain the track continuity by assigning unique labels to the Gaussian components (see the second example). (Video courtesy of Heriot-Watt University and the University of Melbourne).

The Gaussian mixture PHD/CPHD filter Multi-Target Tracker

One of the common misperceptions of the PHD filter is that it does not give you target tracks. Whilst this is true in the abstract formulation, in practice it is possible to maintain target continuity. This example shows the tracking of the targets between frames by labelling individual Gaussian components. This approach is also applicable to other Gaussian mixture multi-target intensity function filters such as the CPHD filter and is the foundation of track management schemes for them. More complex strategies for dealing with more challenging target resolution uncertainty have also been developed using this approach as a basis. The first practical implementation, which was on sonar data, is demonstrated in the sonar section. (Video courtesy of Heriot-Watt University and the University of Melbourne).

Jump Markov Model

This example shows the evolution of the tracking (left) against the measurements for a Linear Jump Markov Model with 3 models (See Pasha, Tuan and Vo 2006). (Video courtesy of the University of New South Wales and the University of Melbourne).
© Heriot-Watt University and the University of Melbourne 2008