3D Ground Plane Estimation from a Monocular Vehicle-borne Image

An algorithm is proposed to estimate the 3D ground plane region and scene depth information from a monocular image captured by a vehicle-borne camera. Firstly, information about image defocus, image saturation and dark channel prior are fused to estimate a relative depth map of the scene. Then, the 3D ground plane can be inferred by using a bilateral median filter based on the assumption that the horizon is piecewise smooth. Finally, absolute depth map can be obtained by using the principle of imaging geometry. To verify the effectiveness of the proposed algorithm, not only numerous comparative experiments are performed on an offline computer, but also it is applied to the outdoor autonomous obstacle avoidance of a robot vehicle. Experiment results demonstrate that both the 3D ground plane and scene depth information can be well estimated by the proposed algorithm, with which the robot vehicle can successfully detect and avoid the obstacles.