Context- and Terrain-Aware Gait Analysis and Activity Recognition

The global population is ageing, causing an increase in the number of people living with age-related gait-affecting conditions that increase risk of falling. Current methods of gait analysis enable healthcare professionals to evaluate a person’s fall risk and inform rehabilitation. However, gait analysis in unable to reproduce the real-world contexts in which people walk, limiting its capacity as a tool for diagnosis and prognosis with respect to falls. Wearable sensors show promise in enabling gait data to be captured outside the laboratory, but the context-labelling of this data is necessary due to the dependency of gait on walking activity and terrain. Whilst the field of Human Activity Recognition (HAR) provides successful methods of determining walking activity, recent studies have highlighted a lack of consideration for terrain variation among HAR datasets.
This work aims to produce a prototype automatic gait analysis system capable of collecting gait data and labelling it with the context — that is, the activity and terrain, in which the data was collected. Particularly, this work places a focus on producing the first dataset which enables high-accuracy terrain classification using wearable sensors to take gait analysis out of the laboratory and into the real world.
To achieve this aim, a comprehensive background and literature review is established, which finds that technologies that address the underlying spatio-temporal gait parameters have the widest reach and can be applied to both healthy and gait-impaired individuals alike. Following this background, the thesis comprises material from four research papers submitted to various journals and conferences, the first of which is a systematic review which explores the differing healthcare and technological approaches to fall prevention and highlights a lack of real-world data in many fall-related technologies, limiting the generalisation of proposed systems. Then, an exploratory analysis into existing HAR datasets and methods for achieving high activity classification accuracies finds that Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs) enable the highest classification accuracies, whilst Inertial Measurement Units (IMUs) are the highest performing sensor types for activity classification. A novel sensor
system is then designed, comprising IMU sensors, Force Sensing Resistor (FSR) insoles, and novel LiDAR and colour sensor implementations, which are investigated for their use in terrain recognition. This sensor system is used to collect the Context-Aware Human Activity Recognition (CAHAR) dataset, which features 7.8 hours of gait data collected from 20 participants, who each perform 38 combinations of 11 unique activities on 9 different indoor and outdoor terrains. This dataset enables novel investigations into the effects of terrain on gait, and the detectability of these changes using wearable sensors. Analysing this novel dataset shows that both the activity and terrain in which gait data was collected can be classified using SVMs with 96% accuracy for an optimised set of sensors featuring just IMU and colour sensor data only. Overall, this thesis makes major contributions towards the fields of HAR and fall research through the identification of interdisciplinary research gaps in the fall literature, a bias-reduced analysis of existing HAR datasets, the development of a novel sensor system for both HAR and terrain classification, the collection of the first multi-terrain HAR dataset, and the demonstration of the feasibility of high-accuracy terrain classification using wearable sensors. Each of these contributions help to construct a foundation for remote gait analysis systems that can determine the full context in which gait data was captured outside the laboratory.