Enhancing Human-Computer Interaction: An Accurate Video-Based American Sign Language Translation Model Using Machine Learning Techniques

American Sign Language (ASL) is the primary communication method used by individuals who cannot speak or hear well. Only about 2.8% of the United States population is proficient in ASL, making communication challenging and inconvenient for people who use ASL. To address this issue, we have developed a model that processes video data and translates ASL fingerspelling signs into English letters. We evaluated several approaches to achieve this goal, with the most effective method involving splitting the video into frames, followed by using Google's MediaPipe Library to extract information on the locations of hand landmarks in each image. The extracted points were then normalized and fed into a classification model. The models tested include Random Forest, XGBoost, and Logistic Regression. The training phase yielded accuracies of 98% for the Random Forest model, 97% for the XGBoost model, and 94% for the Logistic Regression model. However, the Logistic Regression model outperformed the other models for generalization. Our model demonstrates robust performance despite variations in hand color, camera position, background, and other potential distractions, noted as roadblocks in prior research. This work contributes to enhancing Human-Computer Interaction for individuals who use ASL.