| Radar offers unique advantages over other sensors for the detection of humans, such as remote operation during virtually all weather and lighting conditions. However, humans are difficult targets to detect because they have a small radar cross section (RCS) and move with a low velocity. Thus, humans often fall below the minimum detectable velocity (MDV) of Ground Moving Target Indication (GMTI) radars and are easily masked by ground clutter. Most current radar-based human detection systems employ some type of linear-phase matched filtering as part of the detector, such as Doppler processing or the Adaptive Matched Filter (AMF) test. Multi-channel systems also employ space-time adaptive processing (STAP) to suppress clutter and maximize output signal-to-interference-plus-noise ratio (SINR). However, the phase history of human targets is highly nonlinear, and the resulting phase mismatch causes significant SINR losses in the detector itself, degrading the human detection performance.;In fact, the nonlinearity of the human phase history is not arbitrary, but caused by the complexity of human motion. The periodic motion of each body part, especially that of the arms and legs, makes the human target return distinct and unique, distinguishable from that of even other animals, such as dogs. Thus, while many characteristics, such as the speed, trajectory, and size of a potential human target, are unknown, the uniqueness of human gait can be used to specify the structure of the expected target return. This knowledge can then be used to derive a matched filter more closely matched to the sought target's return.;In this thesis, two algorithms exploiting human modeling and gait analysis are proposed: a parameter estimation-based optimized non-linear phase (ONLP) detector, and a dictionary search-based enhanced optimized non-linear phase (EnONLP) detector.;First, the design of the ONLP detector for single-channel radar systems is considered. As the strongest component of the human return comes from the torso reflection, a sinusoidal model is employed to approximate the expected human return. Maximum likelihood estimates (MLEs) of unknown geometry and model parameters are obtained to maximize the likelihood ratio and resulting matched filter output. Comparisons of the Cramer-Rao bounds (CRB) with the variance of the parameter estimates show that at about 5 dB the MLEs meet the CRB, and that there is about 5 dB of modeling error due from the ONLP approximation. Performance of linear phase FFT-based matched filters is compared to that of the proposed ONLP detector, as well as to the ideal "clairvoyant" detector using Receiver Operating Characteristic (ROC) curves. Results show that the proposed ONLP detector consistently outperforms conventional linear-phase filters. Improved output SNR is achieved, thereby significantly increasing the probability of detection attained.;The impact of clutter is addressed by extending the parameter-estimation based ONLP detector to multi-channel systems, which enable the application of clutter mitigation techniques, and ONLP performance in clutter is considered. Although relative to the Gaussian noise only case performance was degraded, the ONLP algorithm continued to outperform linear-phase matched filters in highly cluttered environments as well.;Second, a dictionary-search based enhanced ONLP (EnONLP) detector is proposed that searches over a dictionary, or database, of potential human responses generated for each possible combination of parameters values in the human model. Now, the complete human model is exploited, not just the torso response. An orthogonal matching pursuit (OMP) algorithm is used to search the dictionary for the optimal linear combination of dictionary entries that represent the clutter cancelled data. The EnONLP detector thus offers a framework not only for extracting additional information on the features of a single detected human target, but can also be used to detect and extract features of multiple targets. ROC curves comparing the performance of conventional STAP, ONLP, and EnONLP in clutter show that the best performance is achieved by the EnONLP algorithm, while both proposed algorithms outperform STAP. |
