Paulo Mendoza
Benedick Labbao
Maintaining proper posture during exercise is crucial for reducing the risk of injuries and maximizing the effectiveness of workouts. This paper presents an Exercise Posture Suggestion System utilizing deep learning techniques to analyze and recommend correct exercise posture in real-time. The system incorporates pose estimation models, human activity recognition algorithms, and error detection mechanisms to provide personalized feedback to users. Pose estimation models including ResNet-50, YOLOv8, and YOLO-NAS are evaluated based on their performance metrics and execution times. Human activity recognition models such as Conv(2+1)D with ResNet (Residual Network), 3D CNN, and CNN with Bidirectional LSTM are trained and tested using the UCF-101 dataset. An error detection algorithm, focusing on key body angles for specific exercises, enhances the accuracy of posture suggestions. The software interface is developed using Unity, providing an intuitive platform for users to receive feedback and visualize correct posture examples. The system demonstrates promising results in optimizing workout routines and reducing injury risks, with further refinement and validation necessary for widespread adoption in fitness settings. Integration of advanced deep learning techniques and human-computer interaction will continue to enhance the system's capabilities, contributing to improved health outcomes and enhanced quality of life.
This is our project for our course CPE 313 - Advanced Machine Learning and Deep Learning.
D-ResNet_PushUps.gif)
D-ResNet_Lunges.gif)
D-ResNet_Squats.gif)
