Adaptive system noise covariance for performance enhancement of Kalman filter-based algorithms

Several designs of Kalman filters and the interacting multiple models algorithm were used in real tracking tasks involving high dynamic targets. The data were obtained through the joint effort of the defense departments of Canada and the US. Their performance, measured in terms of positional deviation and the number of track losses, are rather unsatisfactory even though they perform particularly well when using simulated data. To identify the reasons behind, we compared and analyzed the differences between the model assumptions behind the design of these Kalman filters and the model required for accurate tracking of these targets. In this paper, we discussed our findings. Moreover, based on the characteristics of real tracking data, we present an alternative methodology for measuring the effectiveness of various Kalman filter based trackers in stressful environmental. It can also be used to explain the well known characteristics of Kalman filter. A lower bound for the deviation, obtained from this equation, shows that deviation could be too large to manage if noise bandwidth is as high as the real data instead of a pre-assumed magnitude. Instead of having to redesign many existing Kalman filters to suit for stressful environment, we developed a design-independent module that can be added to different types of Kalman filters based trackers to enhance their performance in the tracking high dynamic targets. The module is called adaptive systems noise covariance estimation. It is not only safe (i.e. almost no negative effect) but it can sometimes even double the performance of trackers in stressful environment.