Incorporation of Features in Multistatic Active Sonar Tracking

This document contains the majority of the research on multistatic feature-aided tracking that I have done in my graduate student career at the University of Washington. It contains an overview of sonar and the measurements that an active sonar system generates. It also gives an overview of a tracker based on joint probabilistic data association (JPDA) which is the basis for the research on integrating features into tracking. Several methods for integrating features are compared: integration into JPDA itself at two places, integration into track management, and simply rejecting contacts that appear to be clutter. The methods were tested on the TNO benchmark dataset, showing that integrating the features into track management performed the best, resulting in increased accuracy, fewer "spurs'' coming off of target tracks, and decreased track fragmentation.;In addition, the use of tracking information to improve classification was explored. By using a tracker to predict the aspect of the target at the current time, contacts can be classified based on their aspect-dependent features, target strength and Doppler. The results of this were interesting for two reasons: a high average accuracy can be obtained by using the aspect estimate along with the uncertainty from the prediction, and that only using the prediction (no uncertainty) always performed worse than using no information at all.;This document also describes the development of two preprocessing techniques (posterior distribution preprocessing and likelihood-based clustering) that allow the combination of measurements that come from different sources, which can then be tracked by a standard JPDA-based tracker. This is especially key for multistatic sonar, as the preprocessing techniques allow a tracker to track very dim targets (Probability of Detection of approximately 0.1) in high clutter scenarios (44 clutter contacts per receiver).;The posterior distribution preprocessing technique is extremely flexible and can fuse extremely different types of measurements (IR and HD video data, imaging sonar and HD video data, multistatic sonar). It allows for the appropriate modeling of the measurement noise, resulting in a system that can be applied to many types of data. In addition, this work describes how the preprocessing step can be modified to incorporate any additional feature data.;The likelihood-based clustering technique works well on multistatic sonar data, and allows for the incorporation of any features when calculating the similarity between contacts. This is especially useful for aspect-dependent features, such as Doppler or amplitude. The clustering step is followed by a fusion step that allows for the estimation of target heading or velocity if the appropriate features are used. Using the preprocessing step results in a tracking system that has improved performance, especially on dim targets in a large amount of clutter.