應用於仿人眼視覺系統之智慧型深度偵測技術

Abstract

從兩台攝影機同時拍攝同景色的影像可以利用三角測量算出立體座標點。但是如果要用這種方法求得立體座標需要對攝影機有一些限制並且有一些參數是需要從實驗中獲得的。本論文提出用類神經網路去訓練出不需要相機參數的深度偵測系統，訊練資料由不同變化的目標點於兩張影像中的位置及其所對應的立體座標點構成。最後從誤差百分比可以知道傳統的深度偵測演算法的誤差遠遠大於本論文所提供的類神經網路深度偵測系統而且其準確度足以在立體視覺研究中被應用。

Stereo–pair images obtained from two cameras can be utilized to compute world coordinate points by using triangulation. However, there are some restrictions from cameras and parameters need to be experimentally obtained, by applying this method. This thesis proposed that, for stereo vision applications which need to evaluate the actual depth, artificial neural networks be used to train the system such that the need for parameters of cameras are eliminated. The training set for our neural network consists of a variety of points in stereo-pair and their corresponding world coordinates. The percentage error obtained from the proposed architecture set-up is comparable with those obtained through traditional depth detection algorithm and that the system is accurate enough for most stereo vision applications.