Direct Monocular Visual Odometry Based on Lidar Vision Fusion*

Lidar-assisted visual odometry is a widely used method for pose estimation. However, existing lidar-visual odometry methods suffer from depth association errors, and single-point feature-based methods have insufficient accuracy and are prone to tracking failures, leading to inaccurate pose estimation. In this paper, we proposed a direct monocular visual odometry method based on lidar visual fusion. Firstly, high-gradient pixels of the lidar point cloud projection are extracted, and an initial pose estimation is performed by minimizing photometric errors, avoiding inaccurate feature-depth associations. Then, the point-line features in keyframes are combined and associated with the current frame for pose refinement, and the line features are also matched using the minimization of photometric errors. The evaluations on the KITTI Odometry and Nuscene datasets, and compared to lidar odometry and similar lidar-assisted visual odometry methods, show that our method achieves better pose estimation accuracy and robustness in the majority of scenarios.