A Nonlinear Filter for Real-Time Joint Tracking and Recognition

This paper presents an approach to non-cooperative aircraft identification that uses non-linear filtering to fuse target kinematic state measurements and target signature measurements. This is applicable to sensors whose signature measurements are sensitive to the sensor-target orientation such as high range resolution radar, synthetic aperture radar, ladar and electo-optical imagers. Fusion is achieved by constructing the joint density for the target’s kinematic state and its class, conditioned on the data. The marginal density for the target class is obtained by integrating out the kinematic variables in this joint density, enabling target recognition. This process is inherently non-Gaussian due to non-linear target dynamics and the presence of multiple modes in the signature densities. To model these non-linearities the time evolution of the joint density between measurements is determined by solving the Fokker-Planck equation using the Alternating Direction Implicit method, a fast finite difference partial differential equation solver. As measurements become available, they are used to update the joint kinematic/class density using Bayes rule and densities derived from physical sensor models. This preserves the non-Gaussian features of the joint density including feedback between aspect, kinematics and class. In a test problem fusing simulated position measurements with high range resolution radar signatures this reduces time-to-classify compared with a Maximum Likelihood Classifier that does not use the position measurements.