本論文以加速強健特徵(SURF)建立擴張型卡爾曼過濾器(EKF)為基礎的視覺式同時定位與建圖(SLAM)所需之稀疏與續存性地圖。SURF是尺度與方向不變的特徵，在移動攝影機所擷取的序列影像中，相較其他方法所偵測的特徵，具有重現性高的優點，適合做為SLAM的地圖特徵。本論文的研究議題包括三個部份：第一部份修改SURF特徵的偵測、描述與比對程序，提高影像處理的運算速度與特徵比對的成功率。第二部份規劃有效率的資料關聯與地圖管理策略，提高機器人狀態估測的準確率。第三部份使用兩個攝影機做為SLAM感測器，左攝影機進行影像平面上預測與估測特徵的狀態，新增特徵時則使用左右攝影機所構成的立體視覺求算影像深度，加速新增特徵的初始化。所發展的演算法最後整合成為雙眼視覺式EKF SLAM系統，也在室內環境中成功測試系統的基本功能、地面基準的誤差、路徑閉合的現象、以及長距離執行SLAM任務的能力。 In this thesis, a sparse and persistent map is established using the method of speeded-up robust features (SURF) and applied on the visual simultaneous localization and mapping (SLAM) based on the extended Kalman filter (EKF). Since SURF are scale- and orientation-invariant features, they have higher repeatability than that of the features obtained by other detection methods. Even in the cases of using moving camera, the SURF method can robustly extract image features from the image sequences. Therefore, it is suitable to be utilized as the map features in SLAM. The research topic of this thesis consists of three parts: first, the procedures of detection, description and matching of the SURF method are modified to improve the image processing speed and feature recognition rate. Second, the effective procedures of data association and map management for EKF SLAM are designed to improve the accuracy of robot state estimation. Finally, two cameras are employed as the only sensor of the SLAM system. The state prediction and estimation of features on image plane is performed using only the left camera. When new features are going to be added to the map, the image depth of these new features are calculated using a stereo vision formed by the left and right cameras. The EKF SLAM with SURF-based map is developed and implemented on a binocular vision system. The integrated system has successfully tested the basic capabilities of SLAM system, including ground truth and loop closure, as well as the ability of navigating over a long distance in indoor environments.