Nonlinear Filtering for In-Surveillance and Out-of-Surveillance Tracking


UAV 

Overview

This work tackles the problem of target tracking in the nonlinear Bayesian filtering framework, where the state evolution is governed by the Fokker-Planck-Kolmogorov Equation. In nonlinear Bayesian filtering, the whole probability density function of the target state is maintained, rather than just the first- and second-order moments as in Kalman Filter-like approaches. This premise yields better prediction and estimation of the target states and provides invaluable flexibility for modeling purposes that will be exploited.

Two problems are studied. The first problem assumes the following scenario. In a given search area, an unmanned aerial vehicle (UAV) detects a mobile ground target using its own sensor(s). While the target is in-surveillance, the UAV makes several noisy observations on the target. These observations are inputted to a real-time nonlinear Bayesian filter, which produces optimal target state estimates. The optimality criterion is defined here in the minimum mean square error sense. Nonlinear Bayesian filters are inherently computationally exhaustive. To overcome this problem, an adaptive algorithm is derived to reduce remarkably the computational burden of this filter, while not degrading the accuracy of the state estimates.

The second problem assumes the following scenario. Some time after the ground target was in-surveillance, the same or another UAV views the target area again, but the target is not detected this time. There are three possibilities that could explain this discrepancy. (i) The first set of observations were not accurate enough, i.e. it was a false detection. (ii) The target is still there, but the second UAV could not detect it, i.e. it was a misdetection. (iii) The target has moved away. We will assume that the target has moved away. Since the target has moved away, where should we look for it? Traditional estimation filters based on historic kinematics information alone may not work well. The target kinematics information could be diluted quickly as the radius of possible target locations from that of the first set of observations gets bigger. However, the previous kinematics (target route history) at least provide a center location for future possible target locations. The proposed estimation algorithm employs the knowledge gained from the first set of observations and the concepts of hospitability maps (HMaps) and synthetic inclination maps (SIMaps) embedded into the optimal nonlinear Bayesian filter that was derived for the in-surveillance case. The HMap can be defined as a gridded spatial terrain-based map describing the effect of different terrain surfaces on the mobility and localization of the target objects. On the other hand, the SIMap describes the behavior of the target by capturing how the target favors certain regions within the search area, hence being synthetically inclined to move towards them.

Related Publications

  1. A nonlinear filter coupled with hospitability and synthetic inclination maps for in-surveillance and out-of-surveillance target tracking
    Z. Kassas and U. Ozguner
    IEEE Transactions on Systems, Man, and Cybernetics - Part C, Vol. 40, Issue 1, Jan. 2010, pp. 87-97

  2. Out-of-surveillance target state estimation: a combined hospitability and synthetic inclination approach
    Z. Kassas, U. Ozguner, and J. Layne
    IEEE Conference on Decision and Control, Dec. 14-17, 2004, Paradise Island, Bahamas, pp. 710-715

  3. A hospitability map approach for estimating a mobile target location
    J. Layne, M. Eilders, Z. Kassas, and U. Ozguner
    In Recent Developments in Cooperative Control and Optimization, Kluwer Academic Publishers, Nov. 2003, pp. 117-124