A 3D Grid Mapping System Based on Depth Prediction from a Monocular Camera

Abstract

In complex unknown 3D environments, an accurate 3D volumetric representation of the environment is important for an intelligent robot. Simultaneous Localization and Mapping (SLAM) is considered as a fundamental direction in this area. RGB-D information is very important in traditional SLAM methods. The depth information obtained by sensors like some RGB-D cameras has limits in precision and accuracy. High-precision sensors like lasers and radars are often very expensive. Efficient algorithms should be adopted into those traditional SLAM system. They can not only improve the system efficiency in poorly-equipped conditions but also reduce the resources consumption of robots. To tackle the trade-off between performance and cost, this paper proposes a system producing a 3D grid map that can be used for navigation with a monocular camera and IMU of small size. Our system uses a deep neural network to predict the depth information of a monocular image and utilize the dynamic frame hopping strategy to make a smoother prediction result. Furthermore, we complete a 3D grid map directly used for navigation. The whole grid mapping process occupies little computation and storage resource at the same time. We adopt the octree structure and a keyframe method in the process of the 3D grid mapping to reduce resource consumption. Experiments in a real-world environment show that our approach achieves good results in depth prediction and can well update the 3D grid map for navigation.