基於機器學習與抓握訊號的物件辨識手套

Object Classification using Grasping Signal and Machine Learning

俞順馨(Shun-Hsin Yu)

指導教授 ： 蔡佳宏

國立交通大學/工學院/工學院機器人碩士學位學程/碩士(2020年)

摘要

本研究利用彎曲感測器及機器學習開發了一可透過抓握辨識物件之感測手套。一般物件辨識常使用視覺方法作為感測輸入來源，但若是在光線不足、受物件遮擋之情況下，則難以以視覺進行物件辨識，因此本研究 開發一感測手套，透過抓握姿態對物體進行物件辨識。研究中的感測手套利用彎曲感測器偵測五指的彎曲情形，並以Python作為程式平台，對抓取訊號 的特徵進行擷取，再利用 特徵訓練 機器學習的分類器 SVM進行物件辨識。實驗中準備了兩組共十八個抓取 物件進行辨識 準確率的評估，第一組為由3D列印機印製之三種大小、三種形狀所組成的9個物件 名 為標準物件組，此組實驗目標為 觀察抓握不同形狀、不同大小物件時五指之彎曲變化。第二組抓取物則為隨機選取的 9個日常生活用品，名為日常生活物件組，選取這個物件組是為了驗證感測手套辨識生活中各種不同外形 物件的能力。此外也針對不同核函數、不同抓取位置以及不同受試者的辨識進行討論。根據實驗結果感測手套成功地進行即時的物件辨識，而五階多項式的核函數對於抓握姿態特徵有最好的辨識率，其辨識的準確度在兩組待測物上分別為95.56%和93.33%。 本研究所開發的感測手套預期可應用在盲人輔助手套以及智慧農業等領域 。

關鍵字

機器學習 ； 物件辨識 ； 抓握

並列摘要

An object-classification glove is proposed and developed using flex sensors and machine learning. The glove can classify objects by a single grasp. Generally, the visual approach is often employed for object classification, and visions are used as the source of sensing input. However, if the lighting condition is inadequate or the sight of the object is blocked, it will be difficult to visually classify the object. Therefore, a sensing glove is developed to perform object classification through a tactile approach. The sensing glove uses five flex sensors to detect the flexion of the five fingers, and use Python as a programming platform to extract features from the grasping signal. The features are trained with support vector machine (SVM), a machine learning classifier, for object classification. In the experiments, two sets of objects were prepared for evaluating the accuracy of the proposed glove. The first set is called printed object set. This set includes 9 objects of 3 different sizes and 3 different shapes, and all the objects are printed by a 3D printer. Testing with the printed object set is to observe the flexion signals of the five fingers while grasping objects of different shapes and sizes. The second set is called daily-life object set, and it includes 9 randomly selected objects from daily-life. This object set is employed to verify the capability of the sensing glove for classifying various objects from daily-life. The classification with different kernel functions, different grasping positions, and different subjects are discussed. According to the results, the sensing gloves successfully perform real-time object classification, and the polynomial kernel function performs better than the Gaussian kernel or linear kernel in terms of accuracy rate. The accuracy of the printed object set and daily-life object set is 95.56% and 93.33%, respectively. The sensing glove proposed in this research could be potentially applied to people with vision impairment, or could be used in the field of smart agriculture.

並列關鍵字

grasping ； machine learning ； object classification

參考文獻

[1] S. J.Lederman and R. L.Klatzky, “Extracting object properties through haptic exploration,” Acta Psychol. (Amst)., vol. 84, no. 1, pp. 29–40, 1993.

[2] S. J.Lederman and R. L.Klatzky, “Hand Movements: A Window into Haptic Object Recognition,” Cogn. Psychol., vol. 19, pp. 342–368, 1987.

[3] T.Iberall, “A Neural Network for Planning Hand Shapes in Human Prehension,” pp. 2288–2293.

[4] T.Feix, R.Pawlik, H.-B.Schmiedmayer, J.Romero, and D.Kragi, “A comprehensive grasp taxonomy,” Robot. Sci. Syst. Conf. Work. Underst. Hum. Hand Adv. Robot. Manip., pp. 2–3, 2009.

[5] J.Fischer, N. W.Thompson, and J. W. K.Harrison, “The prehensile movements of the human hand,” Class. Pap. Orthop., pp. 343–345, 2014.