仿生型自主式水下載具利用雙魚眼攝影機在已知環境中之導航

Abstract

本文描述仿生機器魚利用雙魚眼攝影機、電子羅盤及加速度計在一個已知的水下環境下的定位演算法。首先，本文介紹一種立體魚眼攝影機的校正方法，經由校正之影像，使用離散適應性增強演算法訓練的哈爾特徵分類器分類，其結果可被用在辨識已知水中目標物，再利用雙眼特徵對應來估測機器魚與已知目標物間的相對距離，而此估測的相對距離即可當作觀測資訊，整合來自電子羅盤及加速度計的資料(航向角度及三軸加速度)所形成之運動模型中。此模型結合運動模型及觀測資訊，構成機器魚之延伸型卡曼濾波器定位演算法，以達成機器魚在已知環境中之自我定位的目的。最後，本論文展示實驗數據，以驗證此導航法之可行性，並說明定位精度與攝影機的量化誤差、感測器雜訊及背景複雜度之關係。

This work describes a localization algorithm for a fish robot by utilization of stereo fisheye cameras, a compass, and an accelerometer in a known underwater environment. A theory of the stereo fisheye cameras calibration is introduced. Classifiers which use Haar-like features are trained by discrete AdaBoost algorithm and they are used to recognize known landmarks in the underwater environment. Relative distances between the fish robot and the landmarks are then estimated by using stereo features correspondence. Taking the relative distances with respect to the landmarks as observation information, an extended Kalman filter algorithm integrates them with heading angles and 3-axis accelerations from the compass and the accelerometer into the motion model. The extended Kalman filter localization algorithm generates position estimations for BAUV’s self-localization. Finally, the localization algorithm is verified by experimental data, and it can be demonstrated that localization accuracies are limited by the quantization errors of the cameras, the sensors noises, and the background complexity.