| In recent years,artificial intelligence applications such as unmanned vehicles,autonomous robots and augmented reality are advancing rapidly.A key technique of those applications is Simultaneous Localization and Mapping(SLAM).A SLAM system relying on a monocular camera as the only input sensor,called monocular SLAM,is an attractive solution due to its low cost,light weight,easy deployment that has attracted lots of research interests.Dense 3D mapping from SLAM can be used to navigate unmanned vehicles and mobile robotics,and to facilitate interactions between virtual objects and real scenes in augmented reality.However,obtaining a dense map is quite challenging for existing methods.Conventional monocular SLAM based on multi-view geometry can't reconstruct a dense map easily due to the poor performance in low-texture regions.Depth prediction neural network can estimate an accurate and dense depth map,but it usually degenerates in environments of different types from the training data for its limited generalization ability.In this paper,we present a real-time monocular dense mapping system,which integrates a depth prediction network into an existing monocular SLAM framework.The two processes mutually benefit each other to progressively improve the overall performance.Specifically,a weakly-supervised depth prediction pre-trained using pairs of images is employed.Camera poses from the monocular SLAM are applied to select pairs of images,which are used to tune the CNN model on-the-fly in order to improve its accuracy in environments of different types from the training data.In addition,the depth predictions from CNN and monocular SLAM are fused to form a dense and accurate reconstruction result,to overcome the limitation of monocular SLAM in poorly textured regions.Our experiments on public benchmarks demonstrate that our method outperforms the state-of-the-art methods in terms of dense mapping performance.The system achieves realtime performance by utilizing both CPUs and GPUs.Besides,we run our system in real scenarios feasibility of our system,to demonstrate the feasibility of our system for practicable applications. |
