具高運算效率之單攝影機視覺型同時定位與建圖系統

Abstract

解決同時定位與建圖問題最常用的方法是FastSLAM演算法。而FastSLAM2.0的運算效率雖比EKF-SLAM來的高，但FastSLAM2.0會隨著探索時間的增加，比對先前所見地標資訊的次數變多，導致運算效能降低。故本論文對此提出一改良方法，稱之為「具高運算效率同時定位與建圖演算法(ROSLAM)」，在預測機器人位置時可視當前粒子收斂情形，決定是否使用感測器資訊更新機器人位置；並且在每一時刻檢查範圍內的地標數量是否與前一刻相同來判斷是否更新地標。模擬結果顯示，本論文所提出的演算法較FastSLAM2.0與CESLAM皆有兩倍或更高的運算速度提升，並且維持一定精準度。另外，由於SLAM演算法常常搭配雷射感測器來完成任務，但較好的雷射感測器不僅重量可觀，價格也不斐。因此本論文提出一影像特徵量測系統，使用單一攝影機搭配影像處理的方式，將地面與非地面物的邊緣特徵找出並計算該特徵點與機器人的距離，來達到取代傳統雷射感測器的目的。實驗結果證明，本系統除了可以單獨使用於距離量測外，也可與ROSLAM結合，成為具高運算效率之單攝影機視覺型同時定位與建圖系統。FastSLAM is one of the most popular algorithms for solving the simultaneous localization and mapping problem. Though the effectiveness of FastSLAM2.0 is better than that of EKF-SLAM, its performance tends to slow down as the number of landmarks increases. Therefore, this thesis proposes an improved version, called Rapid Operation SLAM (ROSLAM), which uses the convergence of the particles to decide whether the current measurements should be used to update the robot’s pose or not. ROSLAM checks the number of landmarks as compared to the previous one and uses this information to decide whether to update a landmark’s position or not. Empirical evaluation using both simulation and practical experiments shows that ROSLAM is 50% faster than FastSLAM while maintaining similar accuracy of the map. Since laser scan provides clear 3D points clouds, most SLAM research uses laser scanners in spite of the high cost, weight, and power consumption. As a result, a visual SLAM system is proposed in this thesis. First, image processing technique is adopted to find the feature points between obstacles and the floor, so as to calculate the distances between the robot and the feature points. Experimental results show that the proposed system provides sufficient information for a practical visual SLAM system.