Navigation of a mobile robot is a challenging task which requires precise localization of the robot to plan routes and follow them. Localization refinement is commonly done by matching data from the robot’s LiDAR sensors and corresponding area of the map. Conventional geometry-based methods of point cloud matching may fail with shapeless objects (e.g. trees) or with weakly overlapped scans. Addition of semantic features to geometric ones may help to address this issue. However, semantic segmentation models may produce segmentation errors, so additional techniques are required to handle these errors while matching. In this paper, we propose a robust and computationally efficient point cloud matching method based on 2D projection of semantic data and iterative filtering of extracted features. Usage of 2D projections makes our method memory- and computationally-effective, whereas clusterization and integration of RANSAC technique keeps robustness to segmentation errors. We test our approach in a simulated indoor environment and on real outdoor data, comparing it with state-of-the-art semantic-free point cloud matching methods. The experimental results show that our approach outperforms competitors on both memory consumption and matching precision, and works with proper speed for real-time operation.
DOI: 10.1007/978-3-032-34387-1_30
Скачать сборник (PDF) с сайта Springer Nature (англ.): https://link.springer.com/content/pdf/10.1007/978-3-032-34387-1.pdf
Muravyev, K., Vtorushin, D. (2027). Robust and Effective Semantic-Aided Point Cloud Matching for Mobile Robot Navigation // In: Ronzhin, A., Gribova, V., Meshcheryakov, R. (eds) Interactive Collaborative Robotics. ICR 2026. Lecture Notes in Computer Science, Vol. 16790, pp. 421–435.