Smart Fitness Trainer System Using Computer Vision

Authors

  • Tejas Rao C, Mohammed Zainuddin, Syed Faraaz Ahmed, Shrishail M Patil, Priyadarshini R

Abstract

Computer Vision is a field of study that helps the computer see images and understand the content of digital images such as photographs and videos. It attempts to understand the 2D space of an image and uses that to learn the 3D nature of the object. Our goal is to develop a trainer system thatvisually monitors the user’s movements for a particular exercise and then provides the accuracy score along as a feedback. In order to incorporate this we use simple Higher-Resolution Net (HRNet) architecture to achieve Human Pose Estimation, also known as Keypoint Detection and an Action Recognition model. Action recognition helps us in recognizing a human action from avideo containing complete action execution. Since the action involves spatial and temporal context a 3D Convolutional Neural Network (CNN) is used.  The introduced system uses UCF101 –Action Recognition dataset and other sources such as YouTube-8M for collection of the exercise dataset. 3D CNN also makes use of these datasets for training the model and for generating the class scores. The system includes the aforementioned HRNet model and the 3D – CNN for Action Recognition. Initially the system takes the video as an input from the user performing an exercise and then uses HRNet to identify the keypoints of that user or person and simultaneously the 3D-CNN recognizes the action class. Along with it,a keypoint scoring algorithm is used to generate scores for each keypointsor joints along with the overall accuracy of the exercise.

Keywords:— 3D Convolutional Neural Network, HRNet, Keypoint Detection, Action Recognition, Computer Vision, Human Pose Estimation, Keypoint scoring algorithm

Downloads

Published

2020-05-12

Issue

Section

Articles