Real-time human activity recognition via mmWave radar

Abstract

Human Activity Recognition (HAR) has attracted significant attention in various domains, such as health monitoring, sports assistance, and human-computer interaction. Contactless HAR has gained promise as it eliminates the need for cumbersome wearable devices. Among the contactless HAR techniques, millimeter-wave (mmWave) radar has emerged as a promising direction owing to its independence from lighting conditions, penetration capabilities, and privacy-preserving nature. Although mmWave radar has demonstrated its advantages in HAR, there remains substantial room for improvement in real-time monitoring. In this thesis, we propose a real-time mmWave-based detection system for five human activities, incorporating lightweight CFAR-based data acquisition, a novel Doppler-Height feature map for human motion extraction, a CNN-LSTM classifier model, and an innovative sliding window dual-classifier real-time detection scheme. Our model achieves an impressive accuracy of 96.71% and demonstrates a 93.33% correct recognition rate for predefined actions in real-time, along with robustness against non-target motions in experiments.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Ling SHI (Supervisor)