In order to improve the position estimation accuracy of ORB-SLAM2 system and solve the problem that only generate sparse maps, proposing position estimation strategy integrating the ICP (iterative closest point) algorithm and the Manhattan world hypothesis and adding dense-map building thread on the system. Firstly, extracting point, line and surface features by ORB (oriented fast and rotated brief) feature point method, LSD (least-significantdifference) algorithm and AHC (agglomerative hierarchical clustering) method, in which matching the point and line features with previous frame and matching the surface features in the global map. Then the image was divided into non-planar and planar raegions using surfel-based dense-map building strategy, the non-plane uses the ICP algorithm to compute pose, and the plane determines Manhattan world through the orthogonal relationship between plane and plane to use different estimation strategies. The Manhattan world scene realizes the drift-free rotation estimation based on the observation of the Manhattan frame through position decoupling, the translation of the scene and the pose of the non-Manhattan world scene were estimated and optimized by the traced points, lines and surface features. Finally, according to the key frame and the corresponding pose to realize the dense-map. The TUM (technische universität münchen) dataset was used to verify the map method, comparing with ORB-SLAM2 algorithm, RMSE (root mean squared error) decreased by 0.24 cm and the average positioning accuracy increased by 7.17%, which verify the feasibility and effectiveness of the proposed method for the dense map.