Track-before-detect for active sonar.

Abstract

The detection and tracking of underwater targets with active sonar is a challenging problem because of high acoustic clutter, fluctuating target returns and a relatively low measurement update rate. In this thesis, a Bayesian framework for the detection and tracking of underwater targets using active sonar is formulated. In general, Bayesian tracking algorithms are built on two statistical models: the target dynamics model and the measurement model. The target dynamics model describes the evolution of the target state with time and is almost always assumed to be a Markov process. The typical measurement model approximates the sensor image with a collection of discrete points at each frame and allows point measurement tracking to be performed. This thesis investigates alternative target and measurement models and considers their application to active sonar tracking. The Markov process commonly used for target modelling assumes that the state evolves without knowledge of its future destination. Random realisations of a Markov process can also display a large amount of variability and do not, in general, resemble realistic target trajectories. An alternative is the reciprocal process, which assumes conditioning on a known destination state. The first key contribution is the derivation and implementation of a Maximum Likelihood Sequence Estimator (MLSE) for a Hidden Reciprocal Process (HRP). The performance of the proposed algorithm is demonstrated in simulated scenarios and shown to give improved state estimation performance over Markov processes for scenarios featuring reciprocal targets. In point measurement tracking, reducing the sensor data to point detections results in the loss of valuable information. This method is generally sufficient for tracking high Signal-to-Noise Ratio (SNR) targets but can fail in the case of low SNR targets. The alternative to point measurement tracking is to provide the sensor intensity map, an image, as an input into the tracker. This paradigm is referred to as Track-Before-Detect (TkBD). This thesis will focus on a particular TkBD algorithm based on Expectation-Maximisation (EM) data association called the Histogram-Probabilistic Multi-Hypothesis Tracker (H-PMHT) as it handles multiple targets with low complexity. In the second key contribution, we demonstrate a Viterbi implementation of the H-PMHT algorithm, and show that it outperforms the Kalman Filter in the linear non-Gaussian case. A problem with H-PMHT is that it fails to model fluctuating target amplitude, which can degrade performance in realistic sensing conditions. The third key contribution addresses this by replacing the multinomial measurement model with a Poisson mixture process. The new Poisson mixture is shown to be consistent with the original H-PMHT modelling assumptions but it now allows for a randomly evolving mean target amplitude state with instantaneous fluctuations. This new TkBD algorithm is referred to as the Poisson H-PMHT. The Bayesian prior on the target state is also modified to ensure more robust performance. The fourth contribution is a novel TkBD algorithm based on the application of EM data association to a new measurement model that directly describes continuous valued intensity maps and avoids using an intermediate quantisation stage like the H-PMHT. This model is referred to as the Interpolated Poisson measurement model and is integrated into the Probabilistic Multi-Hypothesis Tracker (PMHT) framework to derive a TkBD algorithm for continuous data called the Interpolated Poisson-PMHT (IP-PMHT). The performance of the Poisson H-PMHT and IP-PMHT algorithms are verified through simulations and are shown to outperform the standard H-PMHT in terms of SNR estimation, particularly for scenarios featuring targets with highly fluctuating amplitude. The final key contribution is the application of several TkBD algorithms based on EM data association to the active sonar problem through a comparative study using trial data from an active towed array sonar. The TkBD algorithms are modified to incorporate changes in target appearance with received array bearing, and are shown to give improved SNR and state estimation performance compared with a conventional point measurement tracking algorithm. The thesis concludes by discussing the limitations of the proposed algorithms and possible avenues for future work.