Classifying Object of Standard Grasping Movements Using Data Glove with LSTM Networks

This paper proposes an object classification method using a data glove and deep learning. The classification is performed based on 22 joint angle values obtained from a standard grasp movement on a target object. The relation between fingers’ joint angles and the shape of the target object are described by a superellipsoid model and a simple kinematic model of the human hand. A GraspBiLSTM model is addressed as the classification model. The standard grasping movement experiment dataset of different objects is exacted by Ninapro Dataset 1 Exercise C. According to the experimental results, the accuracy of the GraspBiLSTM classification model is above 90% for 15 daily life objects with standard grasping movements. A data glove that can perform object classification is successfully developed in this work and achieves recognition of objects in an insufficient light environment.