Direction-Based Feature Selection for Efficient LiDAR Odometry in Urban Environment

Feature-based LiDAR odometry has received increasing attention in recent years due to its significant improvement on robustness and real-time performance. Meanwhile, there have been many efforts made to obtain a more compact and informative feature set, the focus of which is on general features for optimization. In this paper, a direction-based feature selection algorithm is proposed to deal with abundant surface features in the urban environment. By theoretically analyzing the spectral attributes of the information matrix, it is found that the orientation distribution of the feature set has a significant effect on the pose uncertainty. Therefore, the degeneracy direction of the current environment is evaluated and taken as an important reference for feature evaluation. Through evaluating the geometrical character of point clouds, an informative subset of features is obtained. The experimental results show that the approach proposed in this study could reduce the computational cost of LO system and achieve a comparable accuracy with the state-of-the-art LiDAR odometry.