Impact Point Prediction and Track Fusion for Heterogeneous Sensors

In this dissertation, two topics in the area of estimation, target tracking and data fusion are investigated: impact point prediction (IPP) and track fusion for heterogeneous sensors. The first topic of interest is the problem of estimating the state of thrusting/ballistic projectiles for the end purpose of IPP. For an IPP algorithm using very short time observations from a single active or passive sensor, we face two major challenges. First, for a thrusting/ballistic projectile, different phases of its trajectory require different mathematical models to describe the corresponding physical behaviors --- there is uncertainty as to which motion model is in effect. Secondly, in practical situations, for a trajectory with unknown drag coefficient and unknown thrust, it is difficult to separate between them given a very short observation time --- there is an ambiguity in their estimation. Utilizing state-of-the-art techniques, we propose an IPP algorithm using state estimation approach (multiple interacting multiple model estimators) with a single active sensor and an IPP algorithm using parameter estimation approach (maximum likelihood estimator) with a single passive sensor. These are further extended under more realistic and complicated environments, i.e., in the presence of wind and with no launch point information, respectively. Secondly, track fusion for heterogeneous sensors will be presented. For distributed multisensor tracking systems, fusing local track estimates from multiple sensors into system tracks achieve better estimation performance than single sensor tracking. However, if one has a heterogeneous mix of sensors --- active (with full position measurements) and passive (with line of sight measurements) --- they use different system models in different state spaces. Compared with homogeneous track-to-track fusion that assumes the same system model for different sensors, the heterogeneous case poses two major challenges. The first one is that we have to fuse estimates from different state spaces (related by a nonlinear transformation). The second is the estimation errors are dependent due to the "common process noise effect" and there is no known way to capture the ``common" part exactly. The linear minimum mean squared error (LMMSE) and the maximum likelihood (ML) approaches are developed for this nonlinear case and compared with the corresponding centralized measurement tracker/fuser (CTF).