基于卷积神经网络的软硬触觉感知方法研究  

Research on soft and hard tactile sensing method based on convolutional neural network

作　　者：余乐[1] 李阳光[1] 陈岩[1] 吴超[1] 李洋洋[1] 王瑶[1] 

机构地区：[1]北京工商大学计算机与信息工程学院食品安全大数据技术北京市重点实验室,北京100048

出　　处：《传感器与微系统》2017年第6期35-37,41,共4页Transducer and Microsystem Technologies

基　　金：北京市自然科学基金资助项目(4174086);国家自然科学基金资助项目(61473009)

摘　　要：智能机器手的应用已经遍布医疗、军工、农业及装配行业等领域。软硬作为物体的重要物理属性之一,对机器手的抓取控制物体有重大影响。在深度学习框架下,基于卷积神经网络提出了用于触觉感知的软硬物体的识别方法。使用薄膜压力传感器采集手指按压软硬物体的数据,建立训练和测试数据集,在Caffe中训练网络,以模拟触觉识别软硬物体。实验结果显示:对软硬物体的识别准确率达94.52%,表明,卷积神经网络对于识别软硬物体有比较好的分类效果。The application of intelligent manipulator has been widely used in medical,military,agricultural,assembly industries and other fields. As one of the important physical properties of objects,softness and hardness have a great influence on grasping control object by intelligent manipulator. Within depth learning framework,a soft and hard objects recognition method for tactile perception based on convolutional neural network（ CNN） is proposed. Thin-film pressure sensor is used to acquire data of finger pressing soft and bard objects,training and test data set are set up. Train network in caffer,so as to simulate tactile to identify soft and hard object. The experimental results show that the accuracy of recognition on soft and hard objects is 94. 52 %,which indicates that the CNN has a good classification effect for recognition of soft and hard objects.

关 键 词：卷积神经网络 薄膜压力传感器 软硬 触觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]