本文的目標是要實現使用網路攝影機即時追蹤影像範圍內的手部區域並且辨識手勢，應用於家電控制與人機互動等領域。我們首先利用膚色檢測和形態學處理分離影像，去除不必要的訊息，再利用背景相減法抓取手部的位置的區域ROI(Region Of Interest)。接著，為了避免雜訊影響到手部區塊，我們使用KCF(Kernelized Correlation Filters)演算法追蹤偵測到的手部區域ROI。最後將ROI的大小調整到100 * 120的大小，再將圖像輸入CNN (Convolutional Neural Networks)網路中進行多種手勢的辨識。接著重複上述追蹤和辨識的步驟達到即時的效果。本研究使用參考Alexnet和VGGnet網路的兩種架構進行訓練和比較，最後在訓練數據集中達到99.9%的辨識率，測試數據集有95.61%的辨識率。;The purposes of this paper are to achieve hand gesture recognition and tracking hand position in real time via web camera. First, using skin-color detect and morphological operations to remove unnecessary noise. Then use the background subtraction method to determine the ROI(Region Of Intereest) region of hand. After obtaining the hand region, Kernel Correlation Filters (KCF) algorithm is used to track the hand. Finally, the hand area is scaled to the size of 100 * 120, then the fixed size of the image input to our CNN (Convolutional Neural Networks) network for identification, in order to achieve the effect of identifying a variety of gestures. And repeat tracking and identification to achieve the real time performance.This research used two frameworks which referenced Alexnet and VGGnet for training and comparison. Finlly, a 99.9% recognition rate is achieved in the training data. The test data set has a recognition rate of 95.61%.
