Enhancement of child gross-motor action recognition by motional time-series images conversion
Satoshi V. Suzuki, Yukie Amemiya, Maiko Sato · 2020
Acquisition of gross motor skills during childhood is quite significant for the physical and psychological developments; hence, early intervention and adequate care are desired if possibility of their developmental disability is discovered. Due to these reasons, several tests were presented and have been used in several countries, but they are not so popular in Japan because of staff shortage and/or time deficit. Therefore, the present authors have been studying a laborsaving AI assessment system based on the gross motor action recognition (GM-AR). In our previous studies, main scheme of the system was proposed, and several IT tools and the GMARs were developed. This paper focuses on an improvement of the GM-AR by expanding the previous result, and the following three methods are newly presented: conversion of the skeleton's time-series data into motional time-series images, the data augmentation, and new CNN-based deep learning. As a result, applying these methods to actual GM assessment including thirteen GM skills at kindergarten, total about 0.15 million data sets were obtained and the classification accuracy was extremely improved to 99.5% by cross validation.