行車時偵測車輛周圍的障礙物,避免車禍的發生一直是重要的議題。光流法被使用在環場鳥瞰影像上,進行偵測障礙物。然而只使用光流法,有時光流值會被誤判或沒有足夠的特徵點描繪出障礙物的區域範圍。因此本論文提出結合光流與視差在環場鳥瞰影像上的資訊,更有效找出車輛周圍的障礙物,提供警示給駕駛者。我們使用光流法,找出各障礙物上特徵點對應相對高度的光流值。同時使用前後張環場鳥瞰影像進行視差運算,找出視差與高度的關係,偵測出障礙物區域及其相對高度。最後結合兩資訊,偵測出較準確的障礙物區域。本研究利用CUDA平行運算,加速此系統的運行。提出的演算法比只使用光流法偵測障礙物,減少平均5.48 % 的錯誤率。
It is always an important issue to detect obstacle while driving to prevent a vehicle accident. The optical flow method is employed to detect the obstacles in around view monitor (AVM). However, by using optical flow only, this method sometimes detects error optical flow points. Also, there is no enough feature points to be detected in order to find the region of obstacles. This paper proposes not only combines disparity map and optical flow of the bird’s-eye view images in AVM, but also provides drivers the region of obstacles nearby. We can find the relative heights of each obstacle’s feature point by optical flow and use stereo disparity to find the obstacle and its height in the meantime. Furthermore, in the study, using CUDA parallel computing to speed up calculation. Based on the combination of optical flow and stereo disparity information, we can get the region of the obstacles more accurately. Compared to using optical flow method, using the proposed method reduce the average error rate by 5.48 % .