Hand Gesture Recognition Based on Keypoint Vector

Abstract

Human-computer interaction (HCI) is usually associated with using popular input devices such as a mouse or keyboard. In other cases hand gestures can actually be useful for human-computer interaction when hand gestures are needed to make the game controls more interesting. There are three basic controls as input mouse: move, click, and drag. Hand gestures and hand shape are different for each person. This becomes a problem during automatic recognition. Recent research has proven the success of the Deep Neural Network (DNN) for representation and high accuracy in hand gesture recognition. DNN algorithms can study complex and nonlinear relationships between features by applying multiple layers. This paper proposes hand feature based on the normalized keypoint vector using DNN. The model was trained on 2250 hand datasets which were divided into 3 classes to identify the mouse movement. The network design uses multilayer with neuron sizes (13, 12, 15, 14) with 500 epochs and achieves the best accuracy of 98.5% for normalized features. The important work in this research is the use of keypoint vector from hand gestures as features to be fed to the DNN to achieve good accuracy.