Simultaneous localization and mapping technology is the basis of a mobile robot, also one of the most important conditions for robot intelligence. SLAM is that a moving robot builds incremental environment maps, and simultaneously make use of the maps to realize the self-positioning in accordance with its pose estimation and visual sensors. Currently, feature-based SLAM (1) approaches are to build sparse maps, but we have no ideas to realize robot path navigation and obstacle avoidance by using a sparse map. Therefore, this paper proposes a semi-dense SLAM with monocular cameras. Our semi-dense mapping runs on key frames, optimised by local bundle adjustment and loop closing, using polar search and block matching. We evaluate our system in our laboratory, and it performs well and steadly.