An Enhanced FingerVision for Contact Spatial Surface Sensing
An Enhanced FingerVision for Contact Spatial Surface Sensing
复制标题
DOI：
10.1109/jsen.2021.3076815
复制
发表时间：
2021-08-01
影响因子：
4.3
通讯作者：
Liu, Honghai
中科院分区：
文献类型：
--
作者：
Yang, Yicheng;Wang, Xiaoxin;Liu, Honghai
基于视觉的触觉传感器是一种很有前途的触觉传感解决方案。本文定制了具有多层结构和仿生特性的FingerVision传感器，并利用其验证了一种两步空间表面感知方法。对于定制传感器，多层结构模仿皮肤和组织，并且传感区域和频率范围与迈斯纳小体和默克尔盘一致。通过空间重构和表面识别验证了两步传感方法。首先，我们采用的集成方法的基础上的梯度估计标记位移的表面重建。根据高度评价，重建的接触面与实际接触面的相似度平均为85.33%。然后在曲面重构过程中根据边界条件对标记点位移进行修正，改进了特征点，实现了KNN算法。当训练集仅为15%时，对19个表面类别的准确率高达99.26%。而利用原始坐标提取的特征，准确率仅为95.58%。两步感知方法暗示了表面重构和识别是相关的，并且重构步骤确实有助于识别。此外，出色的识别能力表明了我们定制的传感器和特征处理方法在通过触摸进行表面识别中的实用价值。
Vision-based tactile sensor is a promising solution for tactile sensing. In this paper, we customized FingerVision sensor with multi-layer structure and biomimetic features, and used it to verify a two-step spatial surface sensing method. For the custom sensor, the multi-layer structure imitated the skin and tissue, and the sensing area and frequency range were consistent with the Meissner's corpuscles and Merkel discs. The two-step sensing method was verified through spatial reconstruction and surface recognition. First, we employed the integration method for surface reconstruction based on the gradients estimated from the marker displacements. The similarity between the reconstructed and the actual contact surface was 85.33% in average according to the evaluation on heights. Then we improved the features by correcting the marker displacements according to the boundary condition during surface reconstruction for surface recognition implementing K-Nearest Neighbor (KNN) algorithm. The accuracy was up to 99.26% when the training set was only 15% for 19 surface classes. While with the features extracted from the original coordinates, the accuracy was only 95.58%. The two-step sensing method implied that the surface reconstruction and recognition were related, and the reconstruction step indeed helped to the recognition. Additionally, the excellent recognition capability indicated the practical value of our custom sensor and the feature processing method in surface recognition through touch.