Multibranch Fusion Method for Similar Object Recognition Based on Smart Tactile Glove

Abstract

The recognition of similar objects faces a challenge in hand haptics due to their identical or similar physical properties within categories, such as material, shape, and elasticity, thereby rendering traditional single-branch tactile recognition methods less accuracy. In this study, we developed a flexible and scalable smart tactile glove endowed with multiple perceptual dimensions, which is capable of capturing hand pose, fingertip pressure, and palm pressure images. The captured tactile video is further passed through a multibranch neural network constructed based on tactile signals of different dimensions after K-means keyframe extraction. By leveraging the benefits of 3-D convolution for video feature extraction, the model deeply fuses multidimensional mechanical features and temporal features of dynamic grasping processes to achieve higher recognition accuracy. The accuracy of recognizing similar objects was enhanced to 95.60%, a significant improvement over the conventional single-branch recognition network used in tactile gloves. The glove design and neural network model proposed herein open new avenues in fields, such as robotic grasping and tactile sensing.