利用深度影像即時估計手指動作

Abstract

我們提出了一個新方法能從深度圖像有效率地估計手部關節。遵從數據驅動的策略，它結合了手勢辨識與搜尋其他時間點的過程，我們採取了一個辨識物件的方法把困難的估計手指動作問題對應到一個較簡單的累計分類問題。使用深度相機讓我們的新分類器不會隨著手的形狀、服裝、飾品等而改變。我們的框架還提供手部估計問題更多的靈活性和應用性。特別是它允許玩家在廣闊的空間中以全身和手指動作來操作系統。 該系統以足以進行互動的速率運作在消費級個人電腦上。我們的評估顯示它在實際的測試集中具有較高的兼容性和穩定性。我們展示了一個利用我們的估計系統來同時使用身體和手指動作的有效率的3D遊戲劇本。

We propose a new method to efficiently estimate hand articulation configurations from a depth image sequence. Following a data-driven strategy that combines gesture reconstruction with temporal retrieval process, we take an object recognition approach that maps difficult finger motion estimation problem into a simpler histogram classification problem. Using a depth camera allows our novel classifier to estimate finger motion invariant to hand shape, clothing, accessories, etc. Our framework further provides more flexibility and applicability in hand estimation issue. In particular, it allows players to operate the system in extensive space with both full-body and finger motion. The system runs at interactive frame rates on consumer-level PCs. Our evaluation shows its high compatibility and stability on real test sets. We demonstrate efficient 3D game scenarios using both body and finger motions by our estimation system.