ABSTRACT

Recognition of Indian Sign Language (ISL) could bridge the gap between deaf-mute people and society. Hand recognition is a key requirement for ISL recognition system. In this paper, the hand region is segmented from the depth image using the Microsoft Kinect Sensor in the cluttered environment. The depth image obtained is then used to implement supervised machine learning by extracting and training the features of images. Here, by comparing various methods, it is depicted that ORB (Oriented FAST and Rotated BRIEF) outruns others in terms of accuracy. ORB is invariant to scale, rotation, and lighting conditions. ORB is also fused with various classification techniques to gain the optimum result. The method is applied to images of ISL 0–9 and is also compared with some standard datasets. Tuning of the ORB with k-NN classification produces an average recognition accuracy of 93.26% with ISL dataset.

摘要

印度手语（ISL）的识别可以弥合聋哑人和社会之间的差距。手部识别是 ISL 识别系统的关键要求。在本文中，手部区域是在杂乱环境中使用 Microsoft Kinect Sensor 从深度图像中分割出来的。然后将获得的深度图像用于通过提取和训练图像的特征来实现监督机器学习。在这里，通过比较各种方法，可以看出 ORB（Oriented FAST 和 Rotated Brief）在准确性方面优于其他方法。ORB 不受缩放、旋转和光照条件的影响。ORB 还融合了各种分类技术以获得最佳结果。该方法适用于 ISL 0-9 的图像，并与一些标准数据集进行了比较。