Reliability and accuracy of Thai sign language recognition with Kinect sensor

Abstract

Hand detection is widely used for recognizing hand gesture which is a primary step for sign language recognition. With larger quantities of sensors, it can provide real-time depth measurements such as Kinect sensor and help researchers find the hand position in the scene without the environment interference such as skin color background. The hand segmentation is based on depth image which is particular correspond to the distance from the signer's hand to Kinect sensor. This article presents a study of the performance for Thai sign language recognition obtained by Kinect sensor in various distance to find the most suitable distance for the accuracy and reliability. After the hand sign is detected, the histograms of oriented gradients will extract the image features of hand sign and proceed to the artificial neural network for recognizing the hand gestures. The result found that the recognition accuracy of Thai finger-spelling of Kinect sensor can work effectively in the distance range of 0.8 - 1.2 meter. The accuracies of recognition for each distance are 83.33% at 0.8m, 81.25% at 1.0m and 72.92% at 1.2m respectively. This distance range can be generated the wide range of brightness in the depth image.