Advanced Human Activity Recognition and Identification Using Edge Processors and Millimeter-Wave Radar Technology

Abstract

Human Activity Recognition (HAR) has progressed significantly with various sensors and edge processors. Deploying HAR on edge devices like wearables requires fast processing and low energy use, as well as enhancing privacy by keeping data local. Meanwhile, researchers are shifting from vision-based methods to radar-based approaches for non-intrusive HAR, with Millimeter-Wave (mmWave) radars becoming popular for their high-resolution spatial data. This thesis provides a comprehensive review of HAR using different sensors and edge processors, focusing on deep learning and mmWave techniques, and covers the fundamentals of these methods that support HAR applications.

This thesis initially demonstrates the feasibility of deploying neural networks on various edge processors for HAR tasks, accompanied by a performance comparison. An adaptive neural network solution is proposed to accelerate the HAR inference process by a factor of 5, reduce energy consumption by approximately 80%, and maintain a comparable level of HAR accuracy (above 90%) compared to the baseline method on the UCI-HAR dataset.

Subsequently, this thesis investigates the application of mmWave technology for human tracking and fall detection. A framework that collaborates three mmWave radars in real time is proposed to track multiple individuals. It achieves an F1 score of 98.4% (a harmonic mean of precision and sensitivity) for tracking a single target, and 97.3% and 95.9% for tracking two and three targets, respectively. Furthermore, for human fall detection, the overall accuracy reaches 96.3%, with a sensitivity of 99.0% for detecting the fall category.

Building upon this framework, this thesis introduces the MmWave-based Human Identification Network (MMIDNet) to investigate indoor identification further. MMIDNet, a neural network architecture specifically designed for mmWave radar point cloud data, is able to classify the identities of the 12 individuals with an overall accuracy of 92.4%. With an extended observation duration from 3 seconds to 7 seconds, the accuracy surpasses 98%.

Date of Award	10 Dec 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Naim Dahnoun (Supervisor) & Jose Nunez-Yanez (Supervisor)