| Human motion recognition is one of the key issues in health care monitoring. Short-term health monitoring is typical for rehabilitation services, while long-term monitoring is common in residential care facilities for the elderly. Although this work gives particular attention to fall detection because of its significance within public health problems, the methods in this thesis can be easily applied to other types of motion.;In this work, the signals corresponding to human motions are captured by a radar system. Radar is a remote sensor that has been proven successful in human motion recognition. Radar signal returns, corresponding to human gross-motor activities, are nonstationary in nature. For these signals, the time-frequency domain is typically used, attributing to its ability to reveal velocities, accelerations, and higher-order Doppler terms of limbs and various human body parts in motion. The range is another important information which can be obtained from multi-frequency or wideband radar returns. It may be used to reveal human location vs. time, and with fine range resolution, it can tag the body main scatterers to their respective Doppler signatures. In general, velocity and range are considered fundamental to human motion classification.;This thesis consists of two parts. In the first part, new forms of time-frequency signal representations based on interpolation and compressive sensing are introduced. These representations provide proper signatures of the incomplete nonstationary data. Incomplete data, i.e., missing observations or random sampling can be due to various reasons, including the discard of noisy measurements, hardware simplification and logistical restrictions on data collection/storage. Once the proper time-frequency signature is obtained, the next step is the classification.;The issue of classification is investigated in the second part of the thesis. New deep learning schemes for classifying radar returns are described. Deep learning has emerged as the key part in the field of artificial intelligence due to its powerful brain-mimicking neural network structures. These complex structures allow an automated way of learning and capturing the intricate properties of the human motion signatures in different domains. Experimental results demonstrate that the proposed methods provide superior results compared to traditional techniques. Finally, the presented work could be used to pave the way for a successful real-time implementation of a radar-based human motion detector. |
