Award-Winning Visual Odometry Algorithm (EDPLVO) Presented at ICRA 2022

The International Conference on Robotics and Automation (ICRA) 2022 received 3,344 submissions, accepted 1,428 papers (43.1% acceptance rate), and selected 13 outstanding papers across sub‑tracks; Meituan’s drone team won the sole Outstanding Navigation Paper award for a visual odometry study.

SLAM (Simultaneous Localization and Mapping) enables robots, including drones and autonomous delivery vehicles, to build maps of unknown environments while moving. Visual SLAM (VSLAM) relies on camera data and is favored for its rich information, low hardware cost, and broad applicability.

The awarded paper introduces EDPLVO, an efficient direct visual odometry algorithm that extends photometric error from points to lines. By proving that any 3D point on a 2D line is determined by the inverse depths of the line’s endpoints, the authors parameterize 3D collinear points, drastically reducing the number of optimization variables while strictly enforcing collinearity, which improves accuracy.

To keep computational load low, the authors add a two‑step optimization: first, with fixed inverse depths and keyframe poses, they fit 3D lines; second, they refine inverse depths and poses using the new line parameters. Both sub‑problems are easy to solve, and the authors demonstrate guaranteed convergence.

Experiments on the TUM monoVO dataset show that EDPLVO outperforms state‑of‑the‑art direct VO methods. Compared with DPLVO, the new line‑photometric consistency error plus the two‑step optimizer yields a 44% speedup in back‑end optimization and higher pose accuracy. The authors also report superior performance over other leading VO algorithms.

These results suggest that EDPLVO is well suited for autonomous navigation in drones, delivery robots, and AR/VR applications, where real‑time pose estimation on embedded hardware is critical.