Dense monocular SLAM using photometric calibration and ZPDepth

Although Monocular SLAM system has a good application prospect in the fields of unmanned driving and artificial intelligence by virtue of its advantages of low cost, simple deployment, and small computational resource requirement. However, there are urgent problems such as too sparse maps, missing depth information, and inability to adapt to complex lighting environments, etc. In this paper, we propose a monocular dense SLAM system (MDS-SLAM) based on photometric calibration and high-resolution depth estimation. MDS-SLAM proposes a photometric calibration method with continuous iterative updating, which converts the gray scale consistency assumption of feature points obtained from optical flow tracking into Lambertian reflection assumption, thus effectively eliminating the effect of automatic camera exposure in complex environments. Meanwhile, MDS-SLAM proposes a high-resolution depth estimation method (ZPDepth) that fuses ZoeDepth and PatchFusion, and generates 3D dense maps by incorporating the predicted depth map into the TSDF voxel grid. Finally, the trajectory and depth estimation experiments on two kinds of datasets, TUM and Replica, and real scenes show that the method in this paper outperforms other comparative systems in both complex environments and high-resolution input devices, and constructs good dense maps.