Learning algorithms for micro-doppler radar based detection, classification and imaging of humans in indoor environments

Abstract

Indoor radars have been researched and developed to detect and monitor humans for applications that range from law enforcement, security and surveillance purposes to ubiquitous sensing applications such as smart home or occupancy detection, assisted living, and, bio-medical applications. These radars are typically phase synchronized to obtain Doppler information of human motions. Humans are non-rigid bodies. The movements of the arms and legs of the humans modulate the carrier frequency of the radar signal giving rise to micro-Doppler features in the radar returns. Research in micro-Dopplers has focused primarily on classifying different human activities, distinguishing between armed and unarmed personnel, and anomaly detection (such as fall detection). There are, however, some limitations associated with the current state-of-the-art methods using low complexity continuous wave (CW) micro-Doppler radars. This dissertation investigates signal processing, and machine learning assisted solutions for advancing the current state of the art techniques for mainly three tasks- target detection, target classification, and target imaging with indoor micro-Doppler radar. Firstly, Doppler sensors are capable of detecting only a single dynamic target. However, indoor environments typically consist of multiple movers - humans, fans, and loudspeakers. When these multiple targets move simultaneously in the propagation channel, their radar backscatter interferes, resulting in distorted micro-Doppler signatures and poor classification accuracies. This limitation may be partially overcome at the cost of increased hardware complexity, but this would offset the inherent advantages of low cost, portable Doppler sensors. Instead, I focused on investigating signal processing solutions to detect multiple simultaneously moving targets. I presented a supervised dictionary learning api proach to represent our micro-Doppler data. Since the resulting representations or dictionaries are customized or fine-tuned to the underlying data - as opposed to data independent transforms such as Fourier or wavelets - I hypothesize that they will have greater success in actually resolving the micro-Doppler signals. Superposed radar returns from multiple targets are resolved into individual components based on their sparse representations. Secondly, in the current works, the training and testing of micro-Doppler signatures for classification have been carried out in the identical system and environmental conditions. However, these conditions may often be necessarily violated in real-world scenarios. For instance, situations may arise where the propagation channel or the presence of interference sources in the test site will permit only specific frequency bands of radar operation. These bands may differ from those used previously while training. In this dissertation, I examine the data-driven signal processing algorithms that demonstrate versatility in handling diversity in test and training data in real-life scenarios. I use customized dictionaries learned from micro-Doppler radar data gathered at different carrier frequencies to obtain sparse representations which are highly discriminative and characterize the target motion as opposed to the sensor parameters. These features are subsequently used for classifying test data from another distinct carrier frequency. Our experimental results show that the dictionary learning algorithms are capable of extracting meaningful representations of the micro-Dopplers despite the rich frequency diversity in the data. Thirdly, there has been very limited research effort in imaging targets based on micro-Doppler radar returns due to the considerable variations that may exist in the indoor propagation environment. For instance, if the radar is deployed in through-wall settings, walls being dispersive and in-homogeneous mediums may introduce considerable distortions such as attenuation, delay and multipath to the radar returns, resulting in distorted radar images. In this dissertation, I focus on mitigating wall effects using a machine learning-based solution- denoising autoencoders- that does not require prior information of the wall parameters or room geometry. Instead, the method relies on the availability of a large volume of training radar data gathered in through-wall conditions and the corresponding clean data captured in line-of-sight conditions. I have validated the performance of the denoising solution for both static and dynamic human subjects. In each of these cases, the signal processing and machine learning algorithms ii are trained to handle diversity in human motion characteristics, radar system parameters, and propagation channel conditions. However, the performances of these machine learning algorithms are tied to a large volume of high-quality training data. Therefore, I gathered a highly curated data set of simulated and measured human micro-Dopplers in both line-of-sight conditions and through-wall conditions. I have also presented a computationally efficient method to model radar micro-Dopplers in indoor conditions by integrating the stochastic finite-difference time-domain (sFDTD) technique with the primitive based scattering center model of human radar returns. It captures diversity in the propagation environment using a single simulation.