Research Of RGB-D SLAM Based On Geometric Plane Constraints

Simultaneous Localization and Mapping(SLAM) technology is a crucial research topic of autonomous robots and augmented reality.With the development of RGB-D SLAM technology which makes use of the commodity RGB-D sensors,camera pose estimation and environment 3D reconstruction are capable of achieving high accuracy.However,ordinary RGB-D SLAM system takes little of the high-level semantic understanding and perception into consideration.Besides,plenty of geometric structures,which occupy a large part of the environment,have not been used efficiently in those methods.To tackle this problem,this paper sums up the current RGB-D SLAM algorithms and presents a novel plane-based RGB-D SLAM system.We follow the frontend implementation of the state-of-art RGB-D SLAM framework and segment the plane information in the backend.Meanwhile,the plane observations are used to build a global plane model and optimize the global state estimations.The main contributions of this paper are as follows:(1)Aiming at the problem that the structural information in the environment is not used properly,we build a real-time RGB-D SLAM system that concurrently reconstructs dense map and global plane model.Our method tracks the camera pose using information jointly from depth and color image in the frontend,while densely maps the environment in a moving volumetric TSDF.The planar region of keyframe is extracted and fused into global plane model in the backend.Moreover,the local plane observations are used to optimize the global state estimations.(2)In order to solve the problem that the global optimization is hard to perform in RGB-D SLAM,the global pose graph is constructed and optimized by local plane observations and plane structure priors.We adopt the minimal representation of plane in the non-linear optimization.To decrease the time consumption of optimization,we also utilize the sparsity of Jacobian matrix by marginalization.Moreover,we take the full constraint requirement of plane-based optimization into consideration,so that the number of available plane observations is increased.(3)To evaluate the capabilities of our proposed algorithm,public RGB-D benchmarks are used to test the performance of our system.The experimental results demonstrate that our method is accurate in camera trajectory,compared to other state-of-art algorithms.By adding more plane prior constraints into the pose graph,the camera pose estimation can be further optimized.Additionally,our system can reconstruct the dense map and plane model of the environment with high accuracy in real-time.