Selfie Sign Language Recognition with Shape Energy Features and Mahalanobis Distance Classifier

摘要

The SSLR system is simulated and tested. Created video database of 18 Indian signs for ten different signers. The hand shape and contours are generated. The features vectors are obtained in two methods, one with the height of hand shape, centroid and area of hand shape and distance of the centroid of hand portion from origin of the frame and the other with hand contour energies optimized with PCA. The performance of SSLR system for the two methods of feature vector generation is compared with the word matching score using Mahalanobis distance classifier. The word matching score with the contour energy features is improved by 10% compared to that of the feature vector with height of hand shape, centroid and area of hand shape and distance of the centroid of hand portion from origin of the frame. Further work needs the improvement in feature set and the classifier models.