Seeking Optimum System Settings for Physical Activity Recognition on Smartwatches

TL;DR

This study investigates optimal system configurations for smartwatch-based physical activity recognition, comparing personal and impersonal models through extensive analysis of features and classifiers to enhance recognition accuracy.

Contribution

It provides the first large-scale exploration of smartwatch-based PAR, identifying optimal feature sets and classifiers for both personal and impersonal models.

Findings

Optimal feature and classifier combinations identified

Personal models outperform impersonal models in accuracy

Validation confirms effectiveness of proposed settings

Abstract

Physical activity recognition (PAR) using wearable devices can provide valued information regarding an individual's degree of functional ability and lifestyle. In this regards, smartphone-based physical activity recognition is a well-studied area. Research on smartwatch-based PAR, on the other hand, is still in its infancy. Through a large-scale exploratory study, this work aims to investigate the smartwatch-based PAR domain. A detailed analysis of various feature banks and classification methods are carried out to find the optimum system settings for the best performance of any smartwatch-based PAR system for both personal and impersonal models. To further validate our hypothesis for both personal (The classifier is built using the data only from one specific user) and impersonal (The classifier is built using the data from every user except the one under study) models, we tested single subject validation process for smartwatch-based activity recognition.