UWGesture:基於超聲波穿戴式手勢識別系統

Abstract

近年來，手勢識別已被廣泛的研究。現有的研究可以分成兩大類：穿戴式及非接觸式。穿戴式即是將裝置穿戴於使用者身上，透過感測器來識別手勢；非接觸式不需要使用者在身上穿戴設備，透過環境中架設的設備感測手勢，如相機、Kinect、Wifi裝置。在此篇論文中，我們設計並開發了UWGesture，一個基於超聲波穿戴式手勢識別系統，用以識別手指及手掌的動作。不同於基於相機技術，UWGesture不依賴可見的環境以及不具使用者隱私的問題。另外，UWGesture並不需要使用者穿戴於手指上。在此篇論文中，我們開發了UWGesture的原型，包含了四個超聲波傳感器（兩個發射、兩個接收），以及開發了一個基於TI BeagleBone Black的數據接收模塊。具體來說，我們透過超聲波傳感器來感測反射訊號的都普勒效應。透過TI BeagleBone Black收集資料，並且萃取時間域及頻率域之特徵值。接著，透過支持向量機(SVM)分類器來識別手勢。我們透過了實驗結果顯示可以準確的分辨五種手勢，其中準確率及召回率皆可達到86%。

Gesture recognition has been intensively studied recent years. Observed from candidates, existing work fall into two categories: wearable approach and contactless approach. Wearable approach, as its name suggested, consists of a couple of wearable sensing devices attached to a candidate’s body for identifying the candidate’s gesture. Contactless approach frees candidates from wearing devices, by collecting data from remoted sensing devices, e.g., Camera, Kinect, Wi-Fi devices. In this paper, we proposed UWGesture, a Ultrasound-based Wearable Gesture recognition system, for the identification of finger and palm movements. Differing from cam- era based solutions, UWGesture does not rely on visible environment and does not have privacy concern. Also, UWGesture frees users from wearing devices on each finger, which should pay attention for hand washing. In addition, a prototype of UWGesture is developed, where the UWGesture consists of four ultrasound Commercial Off-The-Shelf (CTOS) ultrasound trans- ducers (2×transmitters, 2×receivers) and a TI BeagleBone Black based data processing module we developed. Specifically, the basic idea of the ultrasound transducers is to sense the doppler effect of reflection signals. Then the TI BeagleBone Black based data processing module takes the inputs from ultrasound transducers and extract features of the time domain and frequency domain. Next, Support Vector Machines (SVM) classifier is utilized to classify finger and palm gestures. We demonstrate through extensive simulations that UWGesture can identify 5 hand gestures with at least 86% recall rate and accuracy.