Enhancement of gross-motor action recognition for children by CNN with OpenPose

Development of gross motor skills during the preschool period is significant not only for physical-psychological benefits but also for social participation. Several tests to evaluate gross motor competence in child have been established; however, those are hardly utilized in Japan because of staff shortage or time deficit. In order to improve the situation, the present authors have been studying a labor-saving AI assessment system based on a gross motor action recognition (GM-AR). This paper mentions the following improvements for the GM-AR reported in authors' previous study: 1) upgrading of an assistant app to prepare data set for GM-AR, 2) automatic adjustment of a person tracking algorithm after skeleton detection, and 3) improvement of accuracy of GM-AR by reconstruction of its network model. As a result of the above second, the tracking process worked correctly with 98.8% success rate to about 600 short-cut video clips. Concerning the above third, the improved GM-AR could classify the obtained dataset (including 13 kinds of GM skills) with 82.3% accuracy by stratified 8-fold cross validation.