Centre de recherches mathématiques

La petite galerie du CRM

Comparison of Various Schema of Filter Adaptivity
for the
Tracking of Manoeuvring Targets

This page: 20 Kb. approx.

In their effort to develop a new generation of Command and Control System for naval and airborne applications, the R&D Department of Lockheed Martin Canada evaluated the tracking performances of different algorithms specifically designed to track manoeuvring targets. This challenging problem has generated a great deal of effort over the past several years leading to algorithms of an increasing complexity.

Figure 1 shows the block-diagram of a parallel filter algorithm (3CV-PAR) using three constant velocity motion models which only differs by the process noise. This algorithm incorporates a manoeuvre detection criteria in order to update the parameters or to modify the structure of the track filter. The limitation of such a filter is that the detection criteria is based on threshold rules that need to be satisfied a certain number of consecutive times prior to the filter’s being switched to a manoeuvre mode. The consequence is an unavoidable delay in the manoeuvre response of the track filter, which may lead to dramatic consequences such as target loss.

This problem is solved with track filters algorithms which use a mix of the state estimates of each expected target motion models. Such algorithms are called Interacting Multiple Models (IMM) algorithms. Figure 2 shows the block-diagram of a typical IMM algorithm using two motion models.

The respective tracking accuracy of the 3CV-PAR algorithm and an IMM including two motion models (constant velocity, constant acceleration : CVCA) was evaluated by performing 100 Monte-Carlo runs on a challenging multiple sensors scenario including various type of manoeuvres. The average covariance of the track (Figure 3) stays higher for the IMM than for the 3CV-PAR during the whole scenario indicating that the IMM algorithm does not become overconfident on its modeling of the target motion. This characteristics makes it faster in reacting to target experiencing moderate to evasive manoeuvres. As a consequence, the RMS error (Figure 4) of the IMM filter is lower during the manoeuvre than the RMSE of the 3CV-PAR filter, the tracking accuracy is better and the risk of losing the track during the manoeuvre has been lowered.

As a conclusion, the 3CV-PAR filter can perform adequately in tracking targets experiencing moderate manoeuvres. However, in some applications where targets are experiencing evasive manoeuvres specifically at high speed, tracking filters based on the Interacting Multiple Models algorithm are preferable.

For more information, please contact the author at: alexandre.jouan@lmco.com or visit his personal web page at the URL http://www.total.net/~jouan/home.htm

Alexandre Jouan

Fig.1
Figure 1. Block-diagram of the 3CV-PAR algorithm showing three motion models and the manoeuvre detector

Fig.2
Figure 2. Block-diagram of a two motion models IMM filter showing the track state (X) and covariance (P) update

Fig.3
Figure 3. Average track covariance (x) for the 3CV-PAR (plain line) and CVCA-IMM (dashed line)

Fig.4
Figure 4. Average RMSE (position x) for the 3CV-PAR (plain line) and the CVCA-IMM (dashed line)


Table des matières de la petite galerie du CRM

23 septembre 1998, webmaster@CRM.UMontreal.CA