Vision is one of the most important aesthesia of mobile robot. Conventional vision for robot is so-called simultaneous localization and mapping (SLAM) technology, but this technology usually is used to take cognizance of macrostructure of environment. The objects in the scene are commonly treated as just landmarks and the detail 3D information are ignored. But our cognitive robot requires finding the objects from a large-scale environment and exploring them, 3D data of objects cannot be slid over any more. In this paper, we will develop a vision system for our cognitive robot based on classic batch monocular SLAM technology. The two main contributions of our work are introduce of loop closure method and relevant map storage approach, new 3D spatial cluster denoising technology for data refinement.