應用等階集合法手勢操控系統的研究與實作

Abstract

本論文建立一個可以辨識靜態手勢，並進一步控制機械手臂的系統，全部過程可分為取出手勢、學習及手勢辨識。 在取出手勢的步驟中，先利用連續影像相減找出變動的物體部分，再以此為影像分析的主要區域(ROI, Region of interest) ，進行應用等階集合法的手勢分割，以提高效率。為了手勢分割較完整，本論文提出新型的影像切割法則，其所定義的曲面隱函數為能量模型採用的參數，依循零階集合(Level set zero)移動後取得內外圈的資訊作為邊緣資訊，再利用尤拉格朗日方程式(Euler-Lagrange equation)與梯度下降法(Gradient descent method)建立隱函數演化式，使初始輪廓收斂到手掌邊緣，完成手勢影像切割的目的。為了摒除手掌角度的影響，手勢分割後再利用手掌的慣性積矩陣，利用特徵值分解找出慣性積矩陣的特徵向量，找到主軸並完成手掌轉正的程序。最後為了消除較不重要的特徵，再利用傅立葉描述子，將輪廓座的點標作為參數轉至頻率空間，並取前40項係數再反轉換回座標空間，使手勢輪廓較平滑，消除高頻雜訊部分。 在學習步驟中，本論文針對每種手勢事先蒐集300張訓練手勢影像，並以每張訓練手勢的形心為起點，向手掌輪廓劃分360度，每隔一度取距離並標準化作為特徵。將每張的手掌影像特徵蒐集成資料庫並做奇異值分解，取出前50項最重要的特徵，並保留投射矩陣(projection matrix) ，待欲辨識特徵與之相乘。 在線上(on-line)辨識時，在手勢影像分割出後，取出距離特徵並利用主成份分析降低特徵維度，從資料庫中尋找與之最相近的影像，即達到辨識的目的。 最後本論文將本系統應用在滑鼠游標控制上，並與本實驗室研究計畫的伺服馬達所發展之機械手臂做連結，使用者將可以非接觸式控制伺服馬達動作，進而方便遙控遠端設備。此系統辨識之速度與正確率足可滿足非接觸控制機械手臂的目的。

This study establishes a system that can recognize ten different gestures. The total process can be separated into three steps: getting gesture, learning and recognizing. In getting gesture step, we first use frame differencing method to find the region with pixel data changed, the area is called ROI(region of interest) for applying level set method to segment gesture in image. In order to segment more complete shape of gesture, this study establishes a new model of level set algorithm for image segmentation. This method use an implicit function as a variable of a defined energy model, and according to the movement of zero level set to get interior and exterior information and determine the edge position. This study use Euler-Lagrange equation and gradient descent method to get evolution of implicit function. We accomplish this step as soon as the zero level set converges on the edge of hand shape with specified gesture. In order to avoid the influence from the hand angle, this study get inertia matrix of hand and use eigenvalue/eigenvector decomposition to get direction of principle axis, then rotating hand on the vertical plane. Finally, for the purpose of eliminating unimportant features, we use Fourier descriptor to afford a more smooth hand shape. In learning step, this study collect about sample 300 frames for each gesture, and define the distance between centroid and each edge point as features, then use singular value decomposition and retain projective matrix to increase the recognition efficiency. When the system works on-line, it uses principle value analyze to reduce dimension of feature matrix, the most similar image is defined as the least average distance one, to reach for the result of pattern recognition. Finally, this study applies this system to control cursor, and combines this system with CIDM robot arm. Users can only use hand gestures to control the action of the arm. The result of some typical experimental cases show that the proposed method and hand gesture recognition based control system is successful in the application of remote control system.