適用於大型觸控面板之手勢辨識技術

Abstract

在本篇論文中，我們提出了一個僅利用一對可見光攝影機、且適用於大型觸控面版之手勢識別技術。該系統主要是由一個懸掛式投影機和一對安裝在面板頂部邊角的可見光攝影機組成。在大多數現有的相關系統中，為了降低偵測物體的難度，皆必須在觸控版周邊安裝特殊邊條以排除可能的複雜背景。有別於這些系統，我們的大型觸控面版不需要安裝這些特殊邊條。然而，在這樣的條件下，如何在雜亂的背景中精確地找出觸控點位置，以及如何辨識使用者的手勢，是我們本篇論文主要討論的議題。在我們的演算法裡，我們利用邊界偵測從單一的灰階影像中提取資訊並作整理。接著，我們使用隨機決策森林來訓練手的各個部位以便處理大小以及角度上的變化。次之，我們使用逐層淘汰的方式，加速我們的演算法速度。最後，將兩個攝影機偵測到的手指位置投影回三維座標後，我們即可推測使用者的觸控點位置以及使用者的手勢意義。

In this thesis, we present a novel hand posture recognition technique for a large-scale touch panel using only a pair of visible-light cameras. The system is composed mainly of an overhead image projector and a pair of cameras installed on the top corners of the panel for practicability. Due to the difficulty of object detection in cluttered images, most existing systems in related works have tried to avoid the influence of dynamic backgrounds by adding boarder bars beside the panel. Unlike these works, we remove the setup of boarder bars and use visible-light cameras to detect objects. How to determine precisely the positions of the touch points in cluttered backgrounds and how to recognize human hand postures are the main topics of the system. In our approach, we use edge detectors to extract information from a single gray-scale image in dynamic background under illumination variations. Furthermore, the random forest method is used to train different hand parts in order to conquer orientation and scale variations. After that, we use an adaboost-like rejection process to speed up detection speed. Finally, by projecting the detected hand posture from both cameras onto the 3D space, we can decide where the user is touching and what gesture the user is indicating.