- 學位論文
即時手部動作辨識系統之研發
Development of Real Time Hand Motion Identification System
摘要
摘要 語言是日常生活中人與人之間溝通的主要工具，但語障者無法使用語言和一般人進行交談，在生活中或是工作上必須依賴其他方式與他人溝通，一般較常見的是手語。一般人若未受過手語訓練，則不易了解手語所要表達的內容，所以語障者在生活與工作中仍會受到層層的阻礙，因此本研究利用德州儀器出產的數位訊號處理器(TMS320VC5402)為基礎，發展一套即時手部動作辨識系統。 本研究以數位訊號處理器為基礎，設計一四層板的DSP kit，其大小約10.5cm×8.5cm，周邊元件則包括flash memory、SRAM與電源調整IC等。DSP kit則透過JTAG的方式與個人電腦相連接，進行程式的編輯、除錯與執行。在單機執行方面，則是利用燒錄器將程式燒錄於flash memory中，透過boot loading即可執行程式。 在肌電訊號辨識方面，利用本實驗室自行研發的7通道主動電極腕套擷取肌電訊號，設定閥值判斷訊號的起始點與終止點，特徵化則使用ZC與WAMP兩種特徵化方法，計算動作區間的特徵值，將特徵值輸入前向式的類神經網路辨識。本研究的類神經網路使用前向式類神經網路，14個輸入層節點、16個隱藏層節點以及11個節點的輸出層，而網路中所使用的權值則利用Matlab作權值的訓練，本研究採用此類神經網訓練的權值作為程式運算的權值。前向式網路的測試部份，DSP的訓練樣本辨識率為91％，測試樣本72％，可能因為肌電訊號分割及定點數運算，使得測試樣本的辨識率降低，若接上A/D轉換則可避免訊號分割的問題。
並列摘要
Abstract Language is the primary tool in our daily communication. However, for people who are unable to talk, it is difficult to have a conversation with others. Most of them have to depend on other communication method in day-to-day life and on their work. Sign language is the primary choose. However, it is not easy for people never being trained in sign language to understand what was expressed. Therefore, people using sign language still have problems in their life and on the work. This research propose a real-time system to discriminate the motion of arm by mean of EMG classification. It is hope that this system can be further developed into a sign language translator. In this study, a four-layer DSP kit is developed including flash memory, SRAM, power and a JTAG connector. The DSP kit can connect with personal computer using JTAG to program and debug. The DSP kit is operated in microcomputer mode and has the ability to boot load from the flash memory. A custom made active-electrodes ring is used to acquire the EMG signal, Zero-crossing and Willison amplitude are used to extract features from the EMG signal. Additionally, a forward neural network is used to perform the identification. The neural network proposed in this study including 14 input nodes, 16 hide nodes and 11 output nodes. The corresponding weight at each node were first obtained using Matlab, and used in the DSP kit. It is found that using forward neural network, and weight calculated by Matlab, the DSP can achieve 91％of correct identification in the training set, and 72％ in the test set. The low rate of identification may be caused by the fixed-point calculation in DSP kit and EMG signal truncation in the test period. In the future, with added A/D the signal truncation problem can be resolved.