WorkoutNet: A Deep Learning Model for the Recognition of Workout Actions from Still Images

Workouts or Exercises are some repetitive physical activities essential for maintaining good health. Workouts make us relax and happy, give us more energy, lower the risk of chronic disease, and rejuvenate our bodies and minds. Proper technique and posture are essential to get the most out of any workout; the same workout may lead to injury if performed incorrectly. This paper presents a method to identify a person’s workout action from realistic still images. In contrast to video-based action recognition, the challenge is more difficult for still photographs because the Spatio-temporal information is not present in images. This study proposes a deep learning-based model named WorkoutNet to recognize workout actions. The method achieved remarkable results on the Workout Action Image dataset, consisting of 2796 images prepared by us and on the publicly available Exercise image dataset. The suggested approach predicts different workout actions and classifies them into ten categories with high accuracy. The proposed WorkoutNet model attains 92.75% validation accuracy in classifying the workout action images, succeeded by the XceptionNet, VGG-19, DenseNet121, InceptionV3, MobileNet and VGG-16. Further, the classification performance has been evaluated using the F1-Measure and Confusion Matrix. The suggested model can be utilized in real-life to identify workout action images and assist humans doing workouts without a trainer or guide in understanding the right way.