{"title":"Neural LiDAR Odometry With Feature Association and Reuse for Unstructured Environments","authors":"Liangshu Qian, Wei Li, Yu Hu","doi":"10.1002/rob.22607","DOIUrl":null,"url":null,"abstract":"<div>\n \n <p>Odometry plays a crucial role in autonomous tasks of field robots, providing accurate position and orientation derived from sequential sensor observations. Odometry based on Light Detection and Ranging (LiDAR) sensors has demonstrated widespread applicability in environments with rich structured features, such as urban and indoor settings. However, for unstructured environments like scrubland and rural roads, the extraction, description, and correct matching of LiDAR features between frames become challenging. Due to the lack of flat surfaces and straight lines, the existing odometry approaches, whether using hand-crafted features such as edge and planar points or learned features through networks, will face the problem of decreased positioning accuracy and potential failure. Therefore, we propose a neural LiDAR odometry based on Trans-frame Association to extract more effective features for pose estimation in unstructured environments. The Trans-frame Association module contains a fully interactive frame Transformer and a scan-aware Swin Transformer. The former applies cross-attention to features extracted from two consecutive frames, thus enhancing the accuracy and robustness of feature correspondences by considering the contextual information. The latter restricts the attention mechanism to shift along the scan lines of LiDAR, thereby leveraging the sensor's inherent higher horizontal resolution. Our Transformer has linear complexity, which guarantees the module can meet real-time requirements. Additionally, we design a Reuse Refinement Pyramid architecture to further improve the accuracy of pose estimation by reusing multiresolution features. We conducted extensive experiments on the RELLIS-3D data set and our Matian Ridge data set collected in a representative unstructured scene. The results demonstrate that our network outperforms recent learning-based LiDAR odometry methods in terms of accuracy. The code is available at https://github.com/qlsinori/FAR-LO.</p>\n </div>","PeriodicalId":192,"journal":{"name":"Journal of Field Robotics","volume":"42 7","pages":"3968-3985"},"PeriodicalIF":5.2000,"publicationDate":"2025-06-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Field Robotics","FirstCategoryId":"94","ListUrlMain":"https://onlinelibrary.wiley.com/doi/10.1002/rob.22607","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ROBOTICS","Score":null,"Total":0}
引用次数: 0
Abstract
Odometry plays a crucial role in autonomous tasks of field robots, providing accurate position and orientation derived from sequential sensor observations. Odometry based on Light Detection and Ranging (LiDAR) sensors has demonstrated widespread applicability in environments with rich structured features, such as urban and indoor settings. However, for unstructured environments like scrubland and rural roads, the extraction, description, and correct matching of LiDAR features between frames become challenging. Due to the lack of flat surfaces and straight lines, the existing odometry approaches, whether using hand-crafted features such as edge and planar points or learned features through networks, will face the problem of decreased positioning accuracy and potential failure. Therefore, we propose a neural LiDAR odometry based on Trans-frame Association to extract more effective features for pose estimation in unstructured environments. The Trans-frame Association module contains a fully interactive frame Transformer and a scan-aware Swin Transformer. The former applies cross-attention to features extracted from two consecutive frames, thus enhancing the accuracy and robustness of feature correspondences by considering the contextual information. The latter restricts the attention mechanism to shift along the scan lines of LiDAR, thereby leveraging the sensor's inherent higher horizontal resolution. Our Transformer has linear complexity, which guarantees the module can meet real-time requirements. Additionally, we design a Reuse Refinement Pyramid architecture to further improve the accuracy of pose estimation by reusing multiresolution features. We conducted extensive experiments on the RELLIS-3D data set and our Matian Ridge data set collected in a representative unstructured scene. The results demonstrate that our network outperforms recent learning-based LiDAR odometry methods in terms of accuracy. The code is available at https://github.com/qlsinori/FAR-LO.
期刊介绍:
The Journal of Field Robotics seeks to promote scholarly publications dealing with the fundamentals of robotics in unstructured and dynamic environments.
The Journal focuses on experimental robotics and encourages publication of work that has both theoretical and practical significance.