LiDAR-aided Depth Registration for Visual-inertial Odometry using Spatiotemporal Validation

The traditional visual-based odometry tends to provide depth for visual point features by triangulating it from multiview, resulting in depth deviation due to camera measurement noise. In stead of it, as direct range sensor, LiDAR tends to implement depth registration to undertake this practice, which only considers the spatial validation. To this end, this paper proposes a LiDAR-aide depth registration for visual-inertial odometry using spatiotemporal validation, balancing LiDAR-aided depth registration and triangulation. Specifically, for each point feature, both the coplanarity of rays and deviation of endpoints between consecutive image frames are considered as validation index upon normal spatial validation. Furthermore, the inertial data between front and rear frames is integrated to derive pose at the next moment given pose at current moment. We implement experiment, including ablation study and comparison with the state-of-the-art algorithm, in M2DGR dataset. The results demonstrate that our proposed method outperforms others in terms of both estimated trajectory and registered depth in real-time performance.