Recognition of human activities can supplement many diverse areas of study. It has been utilised to track exercise habits, analyse gait, detect falls, and quantify rehabilitation. Excellent predictive accuracy has been achieved in most of these areas, using the full power of deep learning methods, without time or computational constraints placed on the system. In this thesis, it is proposed that simplified classical machine learning methods can be sufficient to provide an accurate system for many classification tasks with realistic and scalable computational overhead. Data is provided from a smartphone application-based recording system with smartwatch data included and then tested using a series of classical machine learning methods. Studies utilising a combination of these devices are sparse within the literature. The dataset used includes 30 participants, performing at least a minute of 6 different activities, as well as a personalised dataset from 1 participant, performing at least 3 minutes of 21 activities. The outcomes are analysed concerning their performance, and their ability to reduce the overall computational overhead of the system. Employing such an approach, 98.9% accuracy can be achieved in activity classification, 96.2% in gender classification, and 95.6% in participant classification using a reduced set of original features. This is comparable to, and sometimes equal to, deep learning and classical machine learning methods which have performed similarly on such tasks within the literature. Compared to the literature, the datasets used contain the six most common activities classified and show equal performance to the work in the literature which attempts to classify these activities. The models used (Gaussian Naïve Bayes, Decision Tree, Random Forest, Support Vector Machine, and Multi-layer Perceptron) are present in the literature, and provide a direct comparison. The classification of gender is novel, and only some literature identifying participants exists. In the second dataset involving a single participant, complex activities are shown to be classified using a reduced set of features with high accuracy, even across orientation changes. A simple support vector classifier is shown to be capable of above 96.8% accuracy with only 20 features, and more complex models are capable of above 99% accuracy in this dataset. Other works have attempted to classify complex activities which are similar to those used in this dataset, but none have managed to achieve quite so impressive performance.
Analysing Motion Classification Tasks Using a Smartwatch Smartphone Pair
Lunn, L. (Author). 2022
Student thesis: Master's Thesis › Master of Philosophy