The Application of Millimetre-Wave Radars in Human Activity Recognition

Abstract

This thesis researches the application of the millimetre-wave (mmWave) radar in human activity recognition (HAR) systems. Initially, we introduce the fundamentals of the mmWave radar, different data types, and the implementation to capture the data. Furthermore, this thesis discusses the limitations of mmWave radar, various applications of mmWave radar in HAR, and our mmWave radar simulator.
Based on the former introduction, we present various mmWave systems with different aspects of HAR. Initially, we present a human detection and tracking system. We notice that the traditional side wall placed radar systems have the problems of false detection and low performance in some occlusion circumstances. Thus, we propose our ceiling-placed mmwave radar system with the novel tracking algorithm to increase the sensitivity and solve occlusion problems with less radar. Moreover, we present a health monitoring system, which can detect the posture of standing, sitting, and lying as well as detect the heart rate when people are lying down or walking with free body movement. Compared with most of related systems which require the person to lie down in a fixed position with one posture, our system can detect the heart rate when the person is lying with multiple postures and places. Furthermore, when the person is moving, our algorithm can overcome the interference from the body movement and extract the heart rate signal based on our novel phase extraction, heart rate decision, and heart tracking algorithms, while some related systems can merely detect the heart rate when the person is stable or moving with limited movement. Experiment shows our system can detect the heart rate when people are walking with free body movement with a low error rate. Finally, a voxelization algorithm for the point cloud from the mmWave radar system with the CNN is presented. We notice the traditional processing chain of mmWave radars can generate the unstable point cloud, which is ignored by most of the mmWave radar systems. With our novel voxelization and compression algorithms, our system can improve the stability of the point cloud, reconstruct the scene effectively, and improve the training efficiency of the neural network in terms of accuracy, processing time, and training time.

Date of Award	9 Jul 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Naim Dahnoun (Supervisor) & Martin J Cryan (Supervisor)