Radar Signal Estimation and Classification in Complex Environments

Radar systems are increasingly being deployed in environments where identifying targets of interest is complicated by complex propagation, complex scattering from clutter, or both. Recent radar sensing involving complex environments have a wide range of applications such as: autonomous vehicle collision avoidance, stand-off human health monitoring, and urban surveillance. This dissertation develops and demonstrates new radar signal processing methods for two important problems: human gait estimation and unmanned aerial system (UAS) detection, classification, and tracking. In each of these applications, radars must overcome challenges associated with operating in complex environments, whether they are due to multipath propagation and/or clutter which is difficult to distinguish from targets of interest.

Gait biometric identification in a scattering rich environment is a notoriously difficult problem because features or parameters useful for identification can be difficult to estimate with traditional methods due to multipath scattering. This dissertation shows in simulation that body part detection and micro-Doppler feature extraction can be performed using a wideband, high-resolution radar. The similarity between a pendulum model and a limb swinging is discussed and a real pendulum experiment is used to show proof of concept of how the bidirectional spectrum can be used to exploit multipath to help estimate physically meaningful features that be used for biometric identification.

Extreme clutter inhomogeneity in urban environments precludes unmanned aerial system detection and classification using conventional radar array processing methods. Conventional array processing methods typically rely on the availability of signal-free training data from ranges neighboring the range bin of interest. Due to clutter inhomogeneity, this signal-free data from neighboring range bins can not be relied upon in urban environments. This dissertation addresses this problem by tightly integrating a neural network (NN) classifier and blind source separation (BSS) to construct a signal-free covariance matrix using only data from each range bin of interest. It is demonstrated that adaptive beamforming for effective clutter suppression and improved small target detection can be achieved using this method. A technique for automatically labeling radar training and test data using video data is also presented.

For slow, small targets in the presence of other non-target movers, data association is arguably the most difficult part of the tracking problem. In traditional trackers, measurements are associated with existing tracks based on proximity based distance measures. This dissertation presents a recurrent neural network (RNN) based tracker which uses learned target dynamic models, a proximity based distance measure, and classification scores to associate measurements with target tracks. Classification scores for each radar measurement are generated by a NN trained using in situ "pattern-of-life" radar data. These scores together with the distances are then passed to RNN trackers with long-short-term-memory (LSTM), trained for each object class. Simulated and real data results indicate that classification scores significantly improves small UAS track holding times in urban environments.