使用訊息融合降低水下影像鑲嵌誤差之研究

Abstract

使用水下無人載具進行結構物表面攝影記錄工作，常因水中能見度低而必須貼近結構物表面拍攝。局部的影像紀錄必須經由影像鑲嵌技術去拼排成大範圍影像，以做為維修及檢測之根據。影像鑲嵌過程由於亮度不均、影像變形、信號雜訊等因素，產生鑲嵌誤差而次誤差隨著時間累積會造成誤差值的發散，使得所鑲嵌之影像無法被使用。本研究是針對水下影像鑲嵌技術所遭遇的困難，利用多感測器訊號融合與重複的測量，來解決誤差累積與減少影像誤差對鑲嵌結果所造成的影響。本研究使用水下無人載具上所裝配之光學攝影機，在載具移動時拍攝影像，利用相鄰影像重疊部分估測載具之位移，來達到自動鑲嵌。在影像逐一比對的過程中，先將影像轉換成頻域影像，此方法可將影像轉變成週期函數，使影像計算量降低。再者，為了減少影像誤差之影響，本研究採用杜卜勒速度聲納估載具的位置，將此位置資訊與影像估測位移整合，並藉由偵測載具路徑交叉位置之誤差資訊，使用卡曼濾波器之逆向平滑處理，重新排列位置的誤差。本文所提出之方法經由實驗驗證，具有改善整體鑲嵌影像的效果，可以提供未來水下無人載具觀測結構物、海底表面等作業之參考。

This work describes an image registration method for underwater inspection tasks. An unmanned underwater vehicle equipped with a Doppler Velocity Log, a compass, and a depth sensor is used as the platform to carry a video camera. Image sequences are captured and each image of the underwater scene is saved along the path. The images are then combined to create a large composite picture of the underwater structure. Features corresponding to the same area of the structure are found in the overlapping region of two images, relative shifts in each image are determined to form image mosaics from the image sequence. The displacements of the video camera can be estimated from the image shifts, but the estimation error of camera’s absolute position might go unbound because small errors accumulate along the surveying path. Sometimes clear features in the underwater scene are hard to determined, a Kalman filter estimator is used to combine image shifts with vehicle motion data to estimate the absolute position of vehicle. In addition, crossover points between the current and former images are monitored, a smoothing process based on the Kalman filter technique along the reverse surveying path is employed to globally minimize the alignment errors in the image chains. Underwater scenes displayed by mosaic of underwater images prove that smooth and robust estimates of underwater image shifts can be obtained by the method proposed by this study.