Normal distribution transform (NDT) algorithm is a point cloud registration algorithm applied in simultaneous localization and mapping (SLAM). According to the characteristics of terrestrial laser scanning (TLS) technique, we propose an improved NDT algorithm based on speeded up robust feature (SURF) algorithm so that it can be applied conveniently in TLS. In this algorithm, firstly the corresponding relation between the point cloud and the image is created for the point cloud visualization; the feature points are extracted from the image by using SURF algorithm and the matching feature points are identified; according to the corresponding relation, the transformation matrix is calculated, and the initial registration of point clouds is completed. In the NDT, the initial matrix is set as a unit matrix, and the point clouds are divided into three-dimensional voxel grids and registered precisely by the probability distribution function. The experimental results show that this algorithm is not only applicable to the registration for TLS, but also exhibits higher registration accuracy and less calculating time, and it has especially a good registration effect for the point clouds with different resolutions.