This paper introduces a new 3D landmark extraction method using the range and intensity images captured by a single range camera. Speeded up robust features (SURF) detection and matching is used to extract and match features from the intensity images. The range image information is used to transfer the selected 2D features into 3D points. The range measurement bias and uncertainty of the range camera are analysed, and their models are developed for improving the range estimation. After outliers' detection and removal using random sampling consensus (RANSAC), reliable 3D points are obtained. 3D landmarks for simultaneous localisation and mapping (SLAM) are selected from the 3D points considering several factors, such as the uncertainty and geometry of their locations. Because of the availability of the SURF descriptor, the data association in SLAM has been performed using both the geometry and the descriptor information. The proposed method is tested in unstructured indoor environments, where the range camera moves in six degrees of freedom. Experimental results demonstrate the success of the proposed 3D landmark extraction method for SLAM.