- 學位論文
透過單一深度相機進行手勢辨識及手部模型重建
Hand Pose Tracking and Hand Model Reconstruction from a Single Depth Camera
摘要
近年來，手勢追蹤在不同的領域中逐漸的受到重視，包含在人機互 動、手語及手勢的了解以及擴充實境等應用。尤其是在人機互動上， 運用手勢來控制如電腦、手機等電子設備已是繼鍵盤、滑鼠及觸控之 後，一種有發展潛力且更加直觀的操作方式。此外，在市場中，越來 越多商品配備著深度攝影機，而深度攝影機所取得的資訊較一般彩色 影像更能分析出物體的空間位置，因此有關於深度攝影機的運用也漸 漸實用且更加吸引人。然而，手勢追蹤一直以來都是相當有挑戰性的 問題。其難度是來自於手本身即具有複雜的關節架構，而在有限的角 度拍攝下，遮蔽的情形相當嚴重而且也會出現大量的雜訊。 在此，運用單一的深度攝影機，我們提出了一套來解決以手部模型 為主軸的手勢辨識方法。藉由建構出完整並正確的手部模型來對應到 目前深度影像上的資訊。整個追蹤方法可以被視為一個最佳化的問題， 而我們運用包括深度資訊、二維的輪廓圖、手部模型是否互相干擾以 及時間上的一致性來判斷我們所推論的手部模型準確與否以及恰不恰 當。在核心演算法的部分，我們採用了一種隨機演算法並利用梯度下 降法加以改進以便更快的找到正確的手勢。此外，我們也運用演算法 以進化的方式改進手勢這點，靠著手部本身的特性，設法去降低每一 次計算中的複雜度。
並列摘要
Hand pose tracking is gradually popular with diverse applications including gesture or sign language understanding, augmented reality and especially in human-computer interaction. Therefore, it is also a competitive system as a next natural step for communication in desktop or mobile environments. Moreover, depth sensors are more available in market and information from depth sensors is more helpful than normal RGB images to capture 3D positions of objects. Thus, the applications of commodity depth camera become useful and attractive. Nonetheless, this problem is challenging due to highly articulated hand poses, noisy input and severe self-occlusions. We propose a model-based hand pose tracking approach using a single depth camera. The method tracks a full hand and reconstructs a hand model with an accuracy pose to fit the current frame. The tracking system can be regarded as an optimization problem and we obtain the optimal hand pose with depth, silhouette, collision and temporal information. In addition, we improve the stochastic optimization method with an additional factor based on gradient descent for articulated hand tracking. In order to solve the complicated articulation problem, we present a novel strategy combining with an evolution algorithm which is used to reduce the high-dimension pose space.