Scale Drift-free Visual-inertial Odometry for Ground Vehicles in Highway Scenarios

Tuopu Wen,Kun Jiang,Jinyu Miao,Benny Wijaya,Mengmeng Yang,Diange Yang

DOI: https://doi.org/10.1109/itsc57777.2023.10421913

2023-01-01

Abstract:Visual-inertial odometry (VIO) has shown its capability of accurate state estimation and has been widely applied in robotics and autonomous driving areas. However, when encountering challenging cases, such as highway scenarios that lack sufficient visual features and inertial excitation, solely relying on VIO will suffer the severe scale drift. While other complementary measurements such as GNSS and wheel odometers are introduced to solve this problem, this paper focuses on improving VIO itself for ground vehicles. Inspired by the fact that vehicles are always clinging to the road, the camera's height on the vehicle to the ground plane can serve as the absolute scale measurement, therefore helping reduce the scale drift of the trajectory. Based on this idea, we design novel homography constraints that provide the up-to-scale constraints between the camera poses and the road plane. Since it is hard to extract stable feature points in texture-less road planes, the dense optical flow estimation is introduced, providing visual clues to construct the homography constraints between the consecutive frames. Then, an initial measurement of camera height is introduced to disentangle the scale ambiguity between the poses and camera height. Moreover, the camera height can be estimated online when the GNSS signal is available. Experiments on the highway scenarios of KAIST Urban datasets demonstrate that our proposed algorithm can reasonably estimate the vehicle trajectories without scale drift, outperforming the original VIO system by a large margin.