Handy-Type Tactile Sensor for Object Recognition Using Convolutional Neural Networks

Abstract

Tactile sensation obtained from touch is one of the most important factors that determine the impression of an object. A method for identifying tactile textures is required because individual differences exist in feeling of tactile textures. In this study, we propose a handy-type tactile sensor for object recognition using convolutional neural networks (CNNs). The sensor consists of a three-axis pressure sensor and an optical motion sensor for a mouse and can detect time-series data. The object is identified using CNN from the time-series data, namely, pressure and speed of the sensor, when the sensor is moved by a user using a hand. Thus, the sensor system configuration is simple without needing a drive device, and it can be possibly constructed at a low cost. Fifteen types of objects were identified using the prototype sensor. The total average correct recognition rate by one specific user in this study was 77%. Further, the total average recognition rate by four separate users without considering each individual use of the sensor system was 48%. Although the problem in individual identification remained, this result demonstrated the potential for identification application. The proposed sensor system can be used as a functional and useful device.