![]() | Only 14 pages are availabe for public view |
Abstract This thesis proposes two different methods for Isolated sign language recognition, A Hidden Markov Model based classifier that uses trajectory information for classifying a dataset of 40 Arabic Sign Language dataset, It works on relative and scaled trajectories and extracts features from Kinect device, it achieves a real time performance and an accuracy of 99.25% in signer dependent settings and an accuracy of 92.5% in signer independent settings, We also propose a multichannel deep learning model for isolated sign language recognition, The model uses hand trajectories data and leverages hand shape sequential patterns, MobileNet was adapted as a pretrained CNN model for the hand shape features, and a one dimensional Google inception like architecture is proposed for hand trajectory feature extraction along with an LSTM based model, Experimental results shows that the proposed modelachieved second best performance on the Argentinian Sign Language Dataset with 99.54% signer dependent accuracy and 97.5% accuracy on Signer Independent settings, Experimental results also shows the importance of Pooling Layers in hand trajectory deep neural networks based models for signer independent settings with an average increase of 1.9% absolute increase over not using the pooling layers in trajectory based models. |