在虛擬實境之中,使用者與虛擬實境的互動乃是重要的一環。系統需確切地得知使用者的一舉一動,並於虛擬世界中加以反應,才能達到如同真實世界一般的互動。為了達到此目的,需要精確地偵測使用者於空間中的動作,而經由立體視覺即可達到此功效。立體視覺於傳統上乃是以三角測量為基礎,需要建立一個複雜的數學模型以及事先預測許多參數,經過相機校正的工作後,方可用來進行影像追蹤。更甚者,由於攝影機的鏡片扭曲問題將此數學模型帶來嚴重的非線性誤差。基於此理由,本論文選擇使用類神經倒傳遞網路來進行攝影機校正的工作,經由類神經網路其可平行處理非線性問題的特性來解決攝影機鏡片扭曲所帶來的非線性誤差,並加入卡爾曼濾波器(Kalman Filter)來加強即時追蹤時的準確度。而在本實驗中所使用的攝影機乃為任天堂公司所發行的Wii remote,經由Wii remote可有效偵測紅外線單點的特性進行追蹤使用者手部的動作,來達到與虛擬環境即時且逼真地互動的目的。
Realistic interactions between users and virtual objects in a virtual reality environment are very important. In order to make users feel like acting in the real world, accurate tracking of the motions of the users in a virtual environment is necessary. Usually, tracking in a virtual environment could be achieved by stereovision. Triangular stereo vision is based on trigonometry, and it needs to build a complex mathematics model and to forecast many parameters. In addition, stereo vision method needs to calibrate cameras for different intrinsic parameters. Another serious problem in using triangulation method is that the distortion of camera lens might cause nonlinear errors in trigonometry calculation. For the reasons above, this thesis uses a neural network to calibrate cameras. Through the neural network’s parallel computing ability, it could solve nonlinear errors caused by the distortion of the lens. This thesis uses the Kalman filter to enhance the accuracy of tracking. Wii remotes from Nintendo are used as cameras for tracking infrared points. Finally, Wii remotes are used to track hand motions so that users can realistically interact with virtual objects in a virtual environment.