Classification Algorithm Improvement for Physical Activity Recognition in Maritime Environments

Abstract

Human activity recognition using wearable sensors and classification methods provides valuable information for the assessment of user’s physical activity levels and for the development of more precise energy expenditure models, which can be used to proactively prevent cardiovascular diseases and obesity. The aim of this study was to evaluate how maritime environment and sea waves affect the performance of modern physical activity recognition methods, which has not yet been investigated. Two similar test suits were conducted on land and on a small yacht where subjects performed various activities, which were grouped into five different activity types of static, transitions, walking, running and jumping. Average activity type classification sensitivity with a decision tree classifier trained using land-based signals from one tri-axial accelerometer placed on lower back and leave-one-subject-out cross-validation scheme was 0.95 ± 0.01 while classifying the activities performed on land, but decreased to 0.81 ± 0.17 while classifying the activities on sea. An additional component produced by sea waves with a frequency of 0.3–0.8 Hz and a peak-to-peak amplitude of 2 m/s² was noted in sea-based signals. Additional filtration methods were developed with the aim to remove the effect of sea waves using the least amount of computational power in order to create a suitable solution for real-time activity classification. The results of this study can be used to develop more precise physical activity classification methods in maritime areas or other locations where background affects the accelerometer signals.