A Posture Evaluation System for Fitness Videos based on Recurrent Neural Network

Abstract

We present a posture evaluation system especially for fitness. Given a fitness video where a user repetitively performs a movement for fitness, we first detect human posture at each video frame. The evolution of posture in consecutive frames is then characterized by a recurrent neural network (RNN). This RNN examines this movement and outputs the degree of goodness (badness). This examination is important for users because prompt inspection of bad movement avoids injury and improves effectiveness of fitness. We demonstrate that the proposed system can accurately detect bad postures when the users perform two movements called Dumbbell Lateral Raise and Biceps Curl. We believe this work is one of the very few studies of using deep neural networks for fitness evaluation.