Multilabel Classification Model for Infant Activity Recognition Using Single Inertial Sensor

Ayaka Onodera, Riku Ishioka, Yuuki Nishiyama, Kaoru Sezaki

Abstract: Recording and sharing childcare information is crucial for accurately assessing a child’s health status and taking appropriate action in case of illness or other emergencies. Although numerous applications and systems have been proposed to assist in recording and sharing these records, the process is still performed manually, presenting a significant burden for parents. Therefore, automatic recording of infants’ daily activities is required. In this study, we implement a machine learning model to recognize multilabeled infant activities using a chest-mounted low-sampling rate accelerometer. We collected accelerometer data from 24 h infants between 6 and 24 months as a dataset. Based on the data, we extracted 25 time- and frequency-domain features calculated from the single accelerometer and user features to recognize the 14 daily activities. The performance evaluation considering multilabel classification showed that our proposed model reaches over 88% in the F1 score in the best case.