Creating and maintaining model for sign language recognition.
This project consists of OpenCV, MediaPipe and Tensorflow tools with mainly Python Language. I used coordinates of face, body and hands landmarks from each frame and passed them to LSTM in order to process sequence of frames as one. To simplify the first version, a static duration=30 frames were used for each data sample. Since it is barely related to real world data in the future I will change model's capabiities to video duration independence.
- OpenCV was used to process images and getting livestream from Laptop camera.
- MediaPipe's solution Holistic was used for face, body and hands recognition
- Tensorflow was used for composing, training and saving LSTM-based model for sign language recongition
- HandsignLanguageRecognition.ipynd ("Battlefield"):
- This is the main file for data collection, wrangling, model selection and evaluation.
- test.py:
- Convenient way to test created model in livestream recognition. All variables are hard-coded, so any changes in the Battlefield file has to be transferred manually.
- saved_models folder:
- All saved model in keras extension
NOTE: Data was collected manually (as it can be seen in the Battlefield file) and will be replaced by more comprehensive datasets in the future.