A Novel Approach for Recognizing Real-Time American Sign Language (ASL) Using the Hand Landmark Distance and Machine Learning Algorithms

Several sign language recognition techniques and models have been prevalent to ease the communication gap between people with hearing disabilities and people who don't understand sign language. The proposed model is a vision-based approach aims to find a better system to recognize sign gestures using the distance between hand landmarks. This process doesn't rely on complex image processing techniques, making it robust towards challenging lighting conditions, noisy backgrounds, and image resolution. To accurately detect 26 American Sign Language (ASL) letters, the model only extracts 12 features. As a result, it can recognize the letters more rapidly than the traditional image-based method and can be trained with lesser computational resources. Therefore, low-configured devices can likewise utilize the method described in this paper. Despite a number of sign language recognition methods are currently in widespread use, the inability of those models to handle the orientation of some specific letter combinations (D, G, U, H, I, and J) has a severe effect on their overall effectiveness. This proposed algorithm, added with features 11 and 12, can solve the anomalies in the orientation of similar pair alphabets. Therefore, the four most popular algorithms, Naive Bayes, Support Vector Machine (SVM), Random Forest, and K-Nearest Neighbors (KNN), have been used to predict the letters of the American Sign Language alphabet. Regardless of the algorithm employed, the model was highly efficient because of the carefully curated dataset, which included attributes directly related to the ASL alphabet signs. SVM demonstrated the best accuracy among these algorithms, outperformina the others by 97 %.