ORB-SLAM2 consumes much time on descriptors calculating, which is a great challenge for real-time performance of SLAM. In this paper, we proposed a modified ORB-SLAM2 method named SD-SLAM2, which achieved better real-time performance than ORB-SLAM2, by replacing the extraction and description of ORB features with sparse direct method. Based on SD-SLAM2, we implemented centralized collaborative SLAM for multi-robot. Experimental results on popular public datasets demonstrate the performance of SD-SLAM2 in monomer SLAM. We also conducted field experiments for multi-robot, whose results demonstrate the performance of SD-SLAM2 in multi-robot SLAM.