Virtual Fitness Assistant Using Machine Learning

Artificial Intelligence plays a major role in modern healthcare and fitness domain by improving and enhancing individual exercise habits, tracking health behaviour, and analyzing repetitive exercise pattern and subsequently use the data to guide in fitness improvement. The thesis proposes, develops and evaluate a smart and effective AI solution based virtual gym assistant in real-time streaming video using CNN.The predominant goal of the proposed task is to apprehend and explore the art of 3D Human Pose Estimation and applying them to different aspects of the task of posture correction in exercises.Over the time, Convolution neural networks (CNN) algorithms have shown significant improvement in the area of human pose estimation on real-time datasets .In this work, we have developed a production-ready application which not only helps people work out effectively at the comfort of their homes but also provides them with real-time feedback on their posture, and act as a personal virtual trainer that will help them to do their exercises in efficient manner. Firstly, we have explored different algorithms and deep learning framework for 3D human pose estimation that could help detect different postures by representing human joints using key-points. These key points help in analyzing the joint coordinates and calculating the angle between joints using mathematical formulation. Finally, we have use these joints coordinates on evaluating pose made by a person on predefined workouts such as biceps, leg-raise, squats and push-ups.