水下具平玻璃介面立體像對之類核線影像產製

Abstract

搭載平玻璃防水保護殼的成像系統(Flat-refractive imaging system)在水下拍攝時，受到多介質環境的影響，光線依照司奈爾定律(Snell’s Law)發生折射，導致原本在單介質環境中共軛光線與基線向量之共面特性，變為四次曲線的共軛對應關係。為了使密匹配程序能夠在核線影像上施作，同時兼顧前方交會三維定位品質，過去文獻提出將匹配和交會拆分為兩個步驟之想法。通過去除匹配影像中的折射畸變，使其接近單介質成像幾何，完成匹配後，再回到原始水下成像幾何執行前方交會。因此，本研究首先提出一套基於近似物距之折射畸變改正方法，使水下影像能夠結合於一般的核線影像產製流程，獲致接近列對列對應之影像資料，本文稱之為「類核線影像」，以便施行密匹配演算法；其次，記錄匹配影像與交會影像間的像點對應關係，以便在匹配完成後，可快速連結至前方交會使用的像點坐標，依循水下成像幾何實現高精度的三維定位。本研究首先以定性和定量方法分析折射改正中關鍵影響因子，接著進行模擬實驗，模擬實務操作中的相關配置，並引入各參數誤差，計算共軛像對間的y視差，以探討本方法有效性。最後，於實際實驗中，使用與模擬實驗對照之相機配置和近似物距模擬方法，通過實際影像資料驗證本方法產製的類核線影像之有效性，並以像對間視差圖及密點雲進行成果分析。實驗成果顯示，當相機具備短玻璃距離時，該系統對於參數和近似物距誤差有較大的容忍度，當近似物距誤差在20%至30%時，仍可有效改正折射並使影像校正至符合列對列對應。對於長玻璃物距，由於本身對誤差的容忍度較低，考量加入參數誤差後，校正效果會下降，建議需提供10%之近似物距誤差，並搭配 1 m以上之物距進行拍攝，方可獲致良好的校正成效。密點雲成果顯示，經本方法改正後，即使在近似物距為50%的條件下，仍可獲得與地真物距改正所獲得的點雲相近的定位品質，且場景幾何並未出現明顯的變形量。

The flat-refractive imaging system is affected by the multi-medium environment during underwater photography, where light refraction occurs according to Snell’s Law. This refraction alters the coplanarity between conjugate rays and baseline vectors in a single-medium environment to a quartic curve conjugate correspondence. To enable dense matching on epipolar images while maintaining 3D positioning quality, previous research proposed a two-step approach separating matching and intersection. By correcting refractive distortion to approximate single-medium imaging geometry, and after matching is completed, the original underwater geometry is then used for forward intersection. Accordingly, this study first proposes a refraction distortion correction method based on approximate object distance, allowing underwater images to be integrated into a general epipolar image production process, resulting in image data with near row-to-row correspondence, referred to as quasi-epipolar images. This facilitates the execution of dense matching algorithms. Additionally, the correspondence between image points in the matching and intersection images is recorded, allowing for quick linkage to the image point coordinates used in forward intersection, thereby achieving high-precision 3D positioning following underwater imaging geometry. This study first qualitatively and quantitatively analyzes the key factors influencing refraction correction, followed by simulation experiments to model relevant configurations applicable to practical operations. The study introduces various parameter errors and calculates the y-parallax between conjugate image pairs to evaluate the effectiveness of this method. Finally, real-world experiments are conducted using camera setups and approximate object distance simulation methods comparable to those in the simulations. These real image data verify the effectiveness of the quasi-epipolar images produced by this method, and the results are analyzed using disparity maps and dense point clouds. Experimental results indicate that when the camera is equipped with a short glass distance, the system tolerates larger errors in parameters and approximate object distances. Even with approximate object distance errors of 20%-30%, refraction can still be effectively corrected, aligning the images to near row-to-row correspondence. For long glass object distances, due to lower tolerance for errors, the correction effect declines when parameter errors are considered. It is recommended to maintain approximate object distance errors within 10% and to shoot at object distances above 1 meter for optimal correction. Dense point cloud results demonstrate that even with a 50% approximation in object distance, the corrected point clouds achieve similar positioning quality to those corrected using the true object distance, with no significant geometric deformation observed in the scene.