Monocular Dense SLAM with Consistent Deep Depth Prediction

Abstract

Monocular simultaneous localization and mapping (SLAM) that using a single moving camera for motion tracking and 3D scene structure reconstruction, is an essential task for many applications, such as vision-based robotic navigation and augmented reality (AR). However, most existing methods can only recover sparse or semi-dense point clouds, which are not adequate for many high-level tasks like obstacle avoidance. Meanwhile, the state-of-the-art methods use multi-view stereo to recover the depth, which is sensitive to the low-textured and non-Lambertian surface. In this work, we propose a novel dense mapping method for monocular SLAM by integrating deep depth prediction. More specifically, a classic feature-based SLAM framework is first used to track camera poses in real-time. Then an unsupervised deep neural network for monocular depth prediction is introduced to estimate dense depth maps for selected keyframes. By incorporating a joint optimization method, predicted depth maps are refined and used to generate local dense submaps. Finally, contiguous submaps are fused with the ego-motion constraint to construct the globally consistent dense map. Extensive experiments on the KITTI dataset demonstrate that the proposed method can remarkably improve the completeness of dense reconstruction in near real-time.

Supported by the National Key Research and Development Program of China under Grant 2018YFB2100601, and National Natural Science Foundation of China under Grant (61872023, 61702482).