TY - GEN
T1 - Free-living Ambulatory Activity Classification
T2 - 9th IEEE RAS/EMBS International Conference on Biomedical Robotics and Biomechatronics, BioRob 2022
AU - Duong, Ton T.H.
AU - Musacchia, Leo
AU - Uher, David
AU - Montes, Jacqueline
AU - Zanotto, Damiano
N1 - Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - Wrist-worn and smartphone-embedded inertial sensors are among the most widely used sensing modalities for activity classification. Insoles instrumented with inertial and force sensors have also become available to researchers and the general public. However, little is known about how classification accuracy is affected by the combination of these three sensing modalities. This study compares the performances of 7 activity classification models, each relying on a combination of 3, 2, or a single sensing modality, under unstructured and free-living conditions. In each model, a genetic algorithm was applied for optimal feature selection, and multi-session leave-one-out cross-validation was used to evaluate model performance. Results for the unstructured condition indicated that the insole-embedded sensors can classify six common ambulatory activities with at least 95% accuracy when used alone or in combination with any of the other sensing modalities. In free-living conditions, sensor combinations that included the insole-embedded sensors demonstrated high levels of agreement with a silver-standard activity tracker. These results provide new insights into the feasibility of using instrumented insoles in combination with phone-embedded or wrist-worn sensors to enhance the accuracy of conventional methods for ambulatory activity classification.
AB - Wrist-worn and smartphone-embedded inertial sensors are among the most widely used sensing modalities for activity classification. Insoles instrumented with inertial and force sensors have also become available to researchers and the general public. However, little is known about how classification accuracy is affected by the combination of these three sensing modalities. This study compares the performances of 7 activity classification models, each relying on a combination of 3, 2, or a single sensing modality, under unstructured and free-living conditions. In each model, a genetic algorithm was applied for optimal feature selection, and multi-session leave-one-out cross-validation was used to evaluate model performance. Results for the unstructured condition indicated that the insole-embedded sensors can classify six common ambulatory activities with at least 95% accuracy when used alone or in combination with any of the other sensing modalities. In free-living conditions, sensor combinations that included the insole-embedded sensors demonstrated high levels of agreement with a silver-standard activity tracker. These results provide new insights into the feasibility of using instrumented insoles in combination with phone-embedded or wrist-worn sensors to enhance the accuracy of conventional methods for ambulatory activity classification.
KW - Ambulatory Activity Recognition
KW - Instrumented Insoles
KW - Machine Learning Inference
KW - Wearable Technology
UR - http://www.scopus.com/inward/record.url?scp=85136922219&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85136922219&partnerID=8YFLogxK
U2 - 10.1109/BioRob52689.2022.9925432
DO - 10.1109/BioRob52689.2022.9925432
M3 - Conference contribution
AN - SCOPUS:85136922219
T3 - Proceedings of the IEEE RAS and EMBS International Conference on Biomedical Robotics and Biomechatronics
BT - BioRob 2022 - 9th IEEE RAS/EMBS International Conference on Biomedical Robotics and Biomechatronics
Y2 - 21 August 2022 through 24 August 2022
ER -