Resource Efficient Machine Learning Techniques for Monitoring Repetitive Activities through Wearable Devices in Real-time

Abstract

Weight training activities have become an inseparable part of an athlete's training to reach their maximum performance. However, at the same time, weight training is considered as the fifth costliest sport in terms of injuries. Studies have shown that monitoring weight training performance is one of the most effective ways of reducing injuries caused by this type of exercise. Given the complexity of weight training exercises, it is a challenge for trainees to know whether they are performing their exercises correctly or not. Considering the fact that incorrect performance of a weight training exercise can result in life-long injuries, which may cost athletes their professional careers, it is of utmost importance to design systems that can detect the incorrect performance of weight training activities.

In this thesis, we motivate the importance of personalised monitoring of weight training performance using wearable devices. We show why current supervised approaches for monitoring weight training routines fail to address the needs of professional athletes. We discuss the emerging need for resource efficient machine learning techniques to monitor weight training activities in real-time, using a wearable device. We then present a novel workflow to detect weight training performance anomalies from observing only the correct performance of an exercise by the trainee. Our workflow motivates two fundamental questions to be addressed in the time series domain: 1- Identifying a trainee's weight training performance from the incoming stream of data generated from wearable devices efficiently and in real-time, 2 Analysing a trainee's weight training performance efficiently.

We address the first question of identifying the trainee's weight training performance using wearable devices by formally defining weight training activities as intervals of recurrence---short bursts of consecutive repeating signals---from the incoming time series data. We present an efficient, online, one-pass and real-time algorithm for finding and tracking intervals of recurrence in a time series data stream. We provide a detailed theoretical analysis of the behaviour of any interval of recurrence, and derive fundamental properties that can be used on real world data streams. We demonstrate the robustness of our method to variations in repetitions of the same pattern adjacent to each other.

We then advance our signal processing approach to monitoring weight training exercises by addressing the shape analogy of signals. Shape analogy is a technique where signals in the form of time series waveforms are compared in terms of how much they look alike. This concept has been applied for many years in geometry. Notably, none of the current techniques describe a time series as a geometric curve that is expressed by its relative location and form in space. To fill this gap, we introduce Shape-Sphere, a vector space where time series are presented as points on the surface of a sphere. We prove a pseudo-metric property for distances in the Shape-Sphere. We show how to describe the average shape of a time series set using the pseudo-metric property of the Shape-Sphere by deriving a centroid from the set. We demonstrate the effectiveness of the pseudo-metric property and its centroid in capturing the shape of a time series set, using two important machine learning techniques, namely: Nearest Centroid Classifiers and K-Means clustering, through 48 publicly available data sets. Our results show that Shape-Sphere significantly improves the efficiency of both techniques. Shape-Sphere improves the nearest centroid classification results when shape is the differentiating feature, while keeping the quality of clustering equivalent to current state-of-the-art techniques.

We subsequently design and develop LiftSmart: a novel smart wearable to detect, track and analyse weight training activities. LiftSmart is the first wearable for weight training that is based on unsupervised machine learning techniques designed in this thesis to eliminate reliance on labelled data. We developed LiftSmart with the ultimate goal of personalised monitoring of professional weight trainers. LiftSmart is tailored to the needs of individual professional weight trainers to monitor their performance by automatically adapting the standard performance of an exercise, which is set for each individual athlete.

In summary, in this thesis we design resource efficient machine learning techniques for monitoring weight training activities in real-time using a wearable device. We demonstrate the effectiveness of our technique in monitoring weight training activities in real-time by designing the first wearable device that automatically detects, tracks and provides feedback about any weight training activity that an athlete performs in a gym.