A Wearable Interface for Recognition and Assistance

Abstract

Human activity recognition (HAR) and gait phase recognition are central to the development of intelligent assistive devices but remain constrained by challenges in real-time inference, robustness to noise, sensor minimization, and generalization across diverse populations. Conventional ankle-foot orthoses (AFOs), though widely used in clinical assistance for dorsiflexion support, restrict the range of motion, while existing active AFOs often fall short due to unreliable phase detection and limited subject generalization.

This thesis addresses these gaps by introducing a Bayesian framework for robust gait phase recognition, integrated with phase-dependent hybrid control. A compact HAR model is developed using wearable inertial and goniometric sensors, achieving high accuracy with ultra-fast inference. To support broader research, the Bilateral Lower Limb Neuromechanical Signals (BLISS) dataset is introduced, containing synchronized bilateral EMG, inertial, motion capture, and force data from both healthy and impaired individuals.

Building on these resources, a minimal-sensor gait phase recognition algorithm—PHRASE (Probabilistic Heuristic Recognition Algorithm for Sequential Events)—is proposed, attaining 98% accuracy and low-latency performance using only two IMUs within a robot operating system (ROS) framework. A proof-of-concept active AFO prototype further demonstrates how accurate real-time phase detection can enable phase-dependent dorsiflexion assistance, validated on unseen participants.

This work advances HAR and gait phase recognition methodologies by demonstrating robust, efficient, and reproducible solutions, laying a foundation for future applications in assistance and locomotor activity monitoring.

Date of Award	12 Oct 2025
Original language	English
Awarding Institution	University of Bath
Sponsors	Egyptian Cultural Affairs and Missions Sector
Supervisor	Uriel Martinez Hernandez (Supervisor) & Ioannis Georgilas (Supervisor)