GCV-SLAM: Ground Constrained Visual SLAM Through Local Ground Planes

The classical visual Simultaneous Localization and Mapping(SLAM) algorithms assume that the camera moves in a free 3D space, while it is not valid for ground vehicles whose poses are influenced by ground surfaces. In recent years, researchers have focused on improving pose estimation accuracy by assuming planar motion which is more suitable to indoor environments. In this paper, we propose a lightweight stereo-visual SLAM framework for ground vehicles in road environments that tightly integrates constraints imposed by ground surfaces. We assume that the ground vehicle exhibits planar motion locally, and extract the observed local ground plane in each keyframe. To avoid the over-parameterization problem in the graph-factor-based optimization process, the Closest Point(CP) representation is adopted to describe the local ground plane. The roll, pitch, and position information provided by the local ground plane can be utilized to constrain the pose of the ground vehicle. Moreover, local ground planes can also be regarded as geometric features, enabling the construction of coplanarity constraints between ground map points and local ground planes, as well as between local ground planes in different keyframes. The KITTI odometry dataset was selected to validate the performance of our system, and the results demonstrated that our system could improve the accuracy and efficiency of ground vehicle localization in road environments.