The work aims to further improve SLAM positioning accuracy and small obstacle recognition ability. The method combined with SLAM and multi-vision was used to set up the AGV kinematics model and then construct the binocular stereo vision model. The improved subregional weighting algorithm based on SURF+RANSAC could eliminate the redundant matching pairs to the greatest extent and improve the matching accuracy significantly. Secondly, on the basis of traditional SLAM navigation, monocular vision was integrated to achieve precise positioning and berthing of key workstations and propose a solution for the occlusion defect of two-dimensional code. Real-time distance measurement was realized by binocular vision. The accuracy of binocular distance detection reached ±1.88 mm and the track accuracy could be controlled at ±2 mm. Fusion of SLAM and multi-vision can effectively improve the accuracy of navigation and positioning and the small obstacle recognition and expand the application of SLAM, so it has a certain prospect of promotion.