Robust multi-sensor data fusion for distributed CFAR target detection

This thesis presents the design and experimental validation of a robust constant false alarm (CFAR) multiradar distributed detection system that operates in clutter whose distribution is unknown. In the system design, aligned radar detectors, each operating under unknown clutter or noise distribution, receives raw data that are not necessarily uncorrelated or stationary, processes them for clutter reduction and target enhancement, and evaluates a test statistic. Subject to a common set of decision and confidence level thresholds, a confidence level is evaluated and attached to each decision that is transmitted to the fusion through an error-free channel. The fusion employs a computationally simple fusion logic based on the individual sensor decisions and associated confidence levels. The fusion level is also independent of the underlying clutter process.; This thesis constitutes the first distributed predetection fusion design that involves end-to-end processes, has been designed to operate under absolute statistical uncertainty and has been validated theoretically as well as experimentally with field data.; The major contributions and novelties of this thesis are: (a) the development of a data structure and preprocessors for range-doppler or range-only processing; (b) the development of a CFAR peripheral test statistic that is asymptotically distribution-free and permits the networking of dissimilar sensors; (c) the use of a combination of peripheral sensor decisions and censored confidence levels to robustify the fusion decision and eliminate weaknesses of the Boolean fusion logic; and (d) extensive experimental validation of the design with collected field (not previously done elsewhere).; The theoretical performance analysis and Monte-Carlo simulations verify that the system exhibits the desired characteristics of CFAR operation, robustness, insensitivity to power fluctuations and fault-tolerance. The confidence level concept employed is a hybrid between Likelihood Ratio Quantization and the binary confidence level and is based on boundaries that are selected independently of the underlying clutter processes. The system is tested with experimental data ranging from target-in-clear to target-in-heavy-clutter conditions and represents the first major end-to-end experimental validation of a distributed detection system.