DBSCAN-Based Dynamic Object Recognition and Semantic Information Entropy-Assisted Vehicle LiDAR Odometry

In vehicle LiDAR odometry, a common challenge arises from the feature misalignment caused by dynamic objects. Although current popular deep learning-based semantic segmentation methods are able to effectively identify dynamic objects, they are limited to recognizing pre-trained types of dynamic objects and require substantial computational resources. Moreover, in the process of using semantic information to assist odometry feature matching, the uncertainty following point cloud semantic segmentation is not taken into account. These drawbacks lead to a lack of robustness in the backend optimization of semantic information-assisted LiDAR odometry, especially in complex environments with various dynamic objects. Thus, we propose a semantic information-based Density-Based Spatial Clustering of Applications with Noise(SI-DBSCAN) for dynamic objects recognition and a semantic information entropy(SIE)-assisted vehicle LiDAR odometry to enhance the positioning accuracy of feature-based LiDAR odometry in dynamic environments. Specifically, we design a SI-DBSCAN model to recognize dynamic objects quickly and efficiently, and innovatively apply the concept of information entropy from probability theory to describe the uncertainty of the point cloud after semantic segmentation, as well as to use it as the weighting information of the point cloud feature matching to enhance the robustness of the LiDAR odometry. Experiments performed on the KITTI dataset show that our LiDAR odometry method, which integrates the SI-DBSCAN and SIE models, achieves a significant improvement in both dynamic objects identification and odometry positioning accuracy compared to existing state-of-the-art methods.