This project enhances the Mask R-CNN framework with ResNet50 for human pose estimation, focusing on dynamic environments. The Real-time Pose Correction Feedback System is designed to assist users in maintaining proper form during physical exercises, starting with yoga, by providing real-time visual feedback.
- Utilizes Mask R-CNN with ResNet50 for accurate human pose estimation.
- Offers real-time feedback by comparing user's posture against ideal exercise forms.
- Targets yoga for initial analysis but is adaptable to various sports for posture improvement.
The project employs a subset of a larger yoga dataset, comprising 500 annotated images to ensure a diverse range of poses for comprehensive learning and analysis.
- Model Architecture: Integrates ResNet50 with Mask R-CNN for feature extraction and pose estimation.
- Keypoint Detection: Incorporates a dense layer in the network to predict x and y coordinates of key human joints.
- Feedback System: Analyzes captured poses against ideal ones to provide corrective feedback in real-time.
- Training and Optimization: Employs data augmentation, dropout, and fine-tuning strategies to enhance model performance and prevent overfitting.
- The model demonstrates a stable training performance with notable accuracy in pose estimation.
- Real-world application potential in fitness, health monitoring, and sports training.
The project illustrates the feasibility of using Mask R-CNN for pose estimation in dynamic settings, with potential expansion to various sports and improvements in model robustness and generalization.
- Clone the repository and install the required dependencies.
- Train the model using the provided dataset or your dataset following the annotation guidelines.
- Deploy the model for real-time pose estimation and feedback.
- Python
- TensorFlow
- Keras
- OpenCV
- Matplotlib (for plotting results)