TY - JOUR
T1 - Swimming style classification based on ensemble learning and adaptive feature value by using inertial measurement unit
AU - Omae, Yuto
AU - Kon, Yoshihisa
AU - Kobayashi, Masahiro
AU - Sakai, Kazuki
AU - Shionoya, Akira
AU - Takahashi, Hirotaka
AU - Akiduki, Takuma
AU - Nakai, Kazufumi
AU - Ezaki, Nobuo
AU - Sakurai, Yoshihisa
AU - Miyaji, Chikara
PY - 2017/7
Y1 - 2017/7
N2 - We have been constructing a swimming ability improvement support system. One of the issues to be addressed is the automatic classification of swimming styles (backstroke, breaststroke, butterfly, and front crawl). The mainstream swimming style classification technique of conventional researches is based on non-ensemble learning; in their classification, breaststroke and butterfly are mixed up with each other. To improve its generalization performance, we need to use better classifiers and more adaptive feature values than previously considered. Therefore, this research has introduced (1) random forest technique, one of ensemble learning techniques, and (2) feature values specific to breaststroke and butterfly to construct a four-swimming-style classifier that has resolved this issue. From subjects with 7 to 20 years history of swimming races, we have obtained their sensor data during swimming and have divided the data into learning data and test data. We have also converted them into feature values that represent their body motions. We have selected from those body-motion-representing feature values the important data to classify four swimming styles and feature values specific to breaststroke and butterfly. We have used the learning data to construct a swimming style classifier, and the test data to evaluate its classification accuracy. The evaluation results show that (1') the introduction of ensemble learning has improved the mean value of F-measure for breaststroke and butterfly by 0.053, and (2') the introduction of feature values specific to breaststroke and butterfly has improved the mean value of F-measure for breaststroke and butterfly by 0.121 as compared with (1'). The proposed swimming style classifier has performed a mean F-measure of 0.981 for the four swimming styles as well as good classification accuracies for front crawl and backstroke. Therefore, we have concluded that the swimming style classifier we have constructed has resolved the problem of mixing up breaststroke and butterfly, as well as can properly classify all different swimming styles.
AB - We have been constructing a swimming ability improvement support system. One of the issues to be addressed is the automatic classification of swimming styles (backstroke, breaststroke, butterfly, and front crawl). The mainstream swimming style classification technique of conventional researches is based on non-ensemble learning; in their classification, breaststroke and butterfly are mixed up with each other. To improve its generalization performance, we need to use better classifiers and more adaptive feature values than previously considered. Therefore, this research has introduced (1) random forest technique, one of ensemble learning techniques, and (2) feature values specific to breaststroke and butterfly to construct a four-swimming-style classifier that has resolved this issue. From subjects with 7 to 20 years history of swimming races, we have obtained their sensor data during swimming and have divided the data into learning data and test data. We have also converted them into feature values that represent their body motions. We have selected from those body-motion-representing feature values the important data to classify four swimming styles and feature values specific to breaststroke and butterfly. We have used the learning data to construct a swimming style classifier, and the test data to evaluate its classification accuracy. The evaluation results show that (1') the introduction of ensemble learning has improved the mean value of F-measure for breaststroke and butterfly by 0.053, and (2') the introduction of feature values specific to breaststroke and butterfly has improved the mean value of F-measure for breaststroke and butterfly by 0.121 as compared with (1'). The proposed swimming style classifier has performed a mean F-measure of 0.981 for the four swimming styles as well as good classification accuracies for front crawl and backstroke. Therefore, we have concluded that the swimming style classifier we have constructed has resolved the problem of mixing up breaststroke and butterfly, as well as can properly classify all different swimming styles.
KW - Ensemble learning
KW - Inertial measurement unit
KW - Machine learning
KW - Random forest
KW - Swimming style classification
UR - http://www.scopus.com/inward/record.url?scp=85026287063&partnerID=8YFLogxK
U2 - 10.20965/jaciii.2017.p0616
DO - 10.20965/jaciii.2017.p0616
M3 - Article
AN - SCOPUS:85026287063
SN - 1343-0130
VL - 21
SP - 616
EP - 631
JO - Journal of Advanced Computational Intelligence and Intelligent Informatics
JF - Journal of Advanced Computational Intelligence and Intelligent Informatics
IS - 4
ER -