A New System to Construct Dense Map with Pyramid Stereo Matching Network and ORB-SLAM2

Building large-scale dense maps with stereo cameras directly is considered to be ineffective. Since neural network has a good effect on restoring image depth, a mapping system combining neural network with SLAM is proposed in this paper. Firstly, the image depth is extracted by PSMNet from the original image, and the accurate depth information is obtained through nonlinear changes. Then, the pose information of the stereo camera is obtained by ORB-SLAM2. Finally, the original image, depth information and pose data are combined to obtain a large-scale dense point cloud map through the mapping module. Experimental results show that the constructed map can reflect the true distribution of the scene structure, and the system performs well in the standard KITTI datasets.