Accuracy and Efficiency Performance of the ICP Procedure Applied to Sign Language Recognition

Abstract

This work addresses the problem of recognizing the American Sign Language (ASL) hand alphabet relying only on depth information acquired from an RGB-D sensor. To accomplish this goal, a novel Iterative Closest Point (ICP) based recognition methodology is proposed where it comprehensively analyzes the inputs and outputs of the alignment as efficiency and accuracy determinants. Next, a novel classification technique, denoted Approximated KB-fit, is proposed to efficiently handle the space complexity of the database template matching. The overall accuracy of the recognition reached a performance of 99.04% in a cross-validation workbench with 520 distinct input depth images. The achieved frame rate was 7.41 FPS performed on a 2.4 GHz single processor based machine.