User and Period Independent Transportation Mode Detection for People with Mobility Disabilities

Abstract

Transportation mode detection (TMD) is an important research area in human activity recognition. It can be more advantageous to people with mobility disabilities, as it can improve their mobility and accessibility by providing a better understanding on their mobility patterns. Such improvement is important to enhance their quality of life and social inclusion. Although previous TMD studies for people without mobility disabilities exhibited high performance, the performance may decrease when the technique is applied to data from a new user or period. The performance of the new data would be more important for people with disabilities because collecting large data from them is difficult. This study investigated the impact of user and period differences on the performance of TMD for wheelchair users (wTMD) and created a novel method to address these challenges. Our main findings are (1) intraclass variability, including user and period differences, degraded the wTMD performance from 94.28% to 62.85%; (2) the multi-DenseNet with a soft voting ensemble provided an 76.49% accuracy to data from different users and period, which was 17.17 percentage points higher than the baseline methods. We expect that our understanding of intraclass variability on wTMD and the newly suggested detection methods will aid in designing more generalized wTMD models.