LiDAR-camera fusion for depth enhanced unsupervised odometry

Official URL: http://dx.doi.org/10.1109/VTC2022-Spring54318.2022.9860840

This paper proposes a robust and safe perception system for an odometry framework based on the fusion of LiDAR data with an RGB image. These multi-modal sensor measurements are fused using their depth proposals and confidence measures in a Bayesian inference module. The resulting fused depth map enhances unsupervised odometry estimates. Experimental results show that the LiDAR-camera fused depth map is an accurate 3D structure representation of the environment. This method can be used in an online adaption of the learningbased odometry algorithms to increase their generalizability to different scenes. We perform experiments on the benchmark odometry datasets and obtain promising results compared to the previous approaches. Compared to the state-of-the-art methods, the average translation error shows a 44% reduction, and the average rotation error is better or comparable.