Aiming at the difficulty in feature extraction and matching in visual SLAM in a sparse texture environment, this paper proposes a sparse semi-direct SLAM that integrates feature method and direct method. Based on the reprojection error model and photometric error model, a theoretical model for sparse semi-direct SLAM method is established. Based on the ORB-SLAM2, a software of sparse semi-direct SLAM method is programmed. Experiments were conducted on multiple public datasets containing sparse texture environments, and the results showed that the sparse semi-direct SLAM proposed in this paper can achieve simultaneous localization and mapping in sparse texture environments, and its tracking robustness has been greatly improved. Compared with ORB-SLAM2, the maximum error of trajectories on fr3/structure_notexture_far dataset has been reduced by 65.06%, the root mean square error has been reduced by 36.08%, and other errors have also been significantly reduced. Overall, the trajectory accuracy is better than ORB-SLAM2.