Gestural AI – Real-Time ASL Interpreter
Video-based ASL recognition with MediaPipe features, temporal models, and a real-time Streamlit demo.
Overview
Built a real-time American Sign Language recognition system using video features, hand-tracking signals, and lightweight inference for live interaction. The best model reached 94% accuracy across a dataset of 20,000+ videos and 166,000 images.
Technical Implementation
The system compares several video and image modeling approaches:
- I3D (Inflated 3D ConvNet) for temporal feature extraction from video sequences
- ResNet for robust spatial feature learning
- MobileNet for optimized real-time inference on edge devices
Key Features
- Real-time Processing: Optimized for low-latency inference suitable for live communication
- High Accuracy: Reached 94% recognition accuracy on diverse ASL gestures
- Scalable Architecture: Deployed via Docker on Linux servers for production use
- Multi-modal Input: Supports both video and image input formats
Technologies Used
- Python for core implementation
- OpenCV for video processing and computer vision
- MediaPipe for hand tracking and pose estimation
- TensorFlow for deep learning model training and inference
- Streamlit for user interface development
- Docker for containerized deployment
Impact
The project helped me think through accessibility ML as a systems problem: model accuracy matters, but so do latency, robustness to camera conditions, and the shape of the user interface around the model.
Future Enhancements
- Integration with mobile applications for on-the-go accessibility
- Support for additional sign languages beyond ASL
- Real-time translation to spoken language
- Integration with video conferencing platforms