Training, evaluating and choosing the best model for sign recognition (signs from Polish Sign Language)
Training dataset consists of 29 groups corresponding with choosen letters, numbers and additional signs from PSL.
Class labels:
Data re-processing consist of:
- determining 21 hand landmarks using mediapipe library with the minimum detection confidence = 0.7
- normalization of input data to the range (0, 1)
- obtaining 42 coordinates (x, y)
Support Vector Machine
K-nearest neighbors
Random Forest Algorithm