Finger spelling recognition using depth information and support vector machine

In Sign Language fingerspelling systems, letters of the alphabet are presented by a diverse form or/and movement of the fingers. In this study, the presented work focus on developing a real-time translation framework of static fingerspelling alphabets. At first an adaptive k-means based cluster method for depth segmentation is proposed, where a flexible cluster number n is used instead of the pre-defined definitive one. Based on the segmentation step, a recognition framework using intensity and depth information is proposed and compared with some distinctive works. Discriminative features extracted from Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), and Zernike moments are used due to their simplicity and good performance. The experiments are executed on a public fingerspelling dataset, which consisted of 120,000 images representing 24 alphabet letters over five different users. The results show that the presented framework is efficient, easy implementation, and performs better than the compared approaches.