Data Augmentation for GrossMotor-ActivityRecognition Using DCGAN
Hyuga Ono, Satoshi V. Suzuki · 2020
Gross Motor Activity Recognition(GM-AR) AI has been studied to develop assessment AI that automatically evaluates child's GM skill. However, collecting and creating datasets takes a lot of effort and time. Therefore, the goal is to automatically generate a GM-AR data set. Since the data set used in this paper is a special time series image, DCGAN was used for data generation. Therefore, this paper aims to generate time-series images by GM-GAN using CNN for GMAR as Discriminator. Newly designed GM-AR could achieve 68.8 % accuracy. The generated image of GM-GAN could be converted into a skeleton with both hands and feet, and the skeleton was confirmed to work with 40 frames. Therefore, GM-GAN was able to generate PK time-series images.