A high-performance, real-time Human Activity Recognition system based on body pose estimation. This project leverages MediaPipe for robust landmark extraction and a Bidirectional GRU deep learning architecture to classify human actions from live video streams.
- Real-time Processing: Optimized pipeline achieving 30+ FPS for seamless inference.
- Geometric Feature Engineering: Beyond raw coordinates, the system calculates joint angles and relative distances to improve recognition for nuanced actions (e.g., drinking, clapping).
- Robust Normalization: Implements geometric scaling and centering, ensuring the model remains invariant to the user's distance or position relative to the camera.
- Modern UI/UX: A sleek dashboard built with
customtkinter, featuring real-time probability charts and performance metrics.
- Language: Python
- Computer Vision: OpenCV, MediaPipe
- Deep Learning: TensorFlow/Keras (Bidirectional GRU)
- GUI Framework: CustomTkinter
- Data Science: NumPy, Pandas, Scikit-learn, Matplotlib
GUI&inference.py: Main application integrating the UI and the inference engine.make_normalize_data.py: Data collection utility with automated landmark normalization and feature extraction.Train_GRU_Upgrade.py: Comprehensive training script featuring Early Stopping, LR reduction, and model evaluation.label_map.json: Configuration file for action class mapping.
# Clone the repository
git clone [https://github.com/Dev-TNT/HAR_realtime.git](https://github.com/Dev-TNT/HAR_realtime.git)
cd HAR_realtime
# Install required dependencies
pip install opencv-python mediapipe tensorflow customtkinter pandas numpy pillow scikit-learn seaborn matplotlib