基于语义辅助的强度扫描上下文回环检测方法  

A semantic-assisted intensity scan context loop closure detection method

作　　者：刘名星 徐晓苏[1] Liu Mingxing;Xu Xiaosu(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学仪器科学与工程学院,南京210096

出　　处：《仪器仪表学报》2025年第2期292-304,共13页Chinese Journal of Scientific Instrument

摘　　要：在同步定位与建图(SLAM)中,回环检测是提高定位精度的重要环节。通过识别回环并校正累计误差,能够有效增强定位的精度和稳定性。然而,目前大多数基于LiDAR的回环检测方法主要依赖坐标和反射强度等低层次特征来构建描述子,未能充分利用场景中的语义信息,从而在复杂场景中面临精度和可靠性不足的问题。针对这一不足,提出了一种基于语义辅助的强度扫描上下文方法,以弥补现有方法的局限性。首先,该研究方法利用ICP算法对两帧点云进行粗配准,减少角度和位移对回环检测的影响。在此基础上,将语义信息与点云的几何信息及强度信息相结合,构建多层级特征的全局描述子。最后,通过描述子相似度计算判断回环是否存在,从而实现更可靠的回环检测。在公开数据集KITTI上的实验结果表明,该研究方法的最大F_(1)分数较scan context算法提升了19.71%,平均均方根误差较lego-loam算法降低了36%。此外,在校园环境的实测结果显示,该研究方法的最大F_(1)分数较lio-sam算法提升了19.23%,较lego-loam算法提升了70.62%;平均均方根误差较lio-sam算法降低了56.68%,较lego-loam算法降低了20.7%。这些结果表明,该研究方法不仅能够有效提升回环检测的准确性,还能在多样化场景下展现出更强的鲁棒性。通过引入语义信息,该研究显著改善了描述子在复杂场景中的区分能力,为SLAM技术的发展提供了新的思路和方法支持。In simultaneous localization and mapping(SLAM),the loop closure detection is a critical step to improve localization accuracy.By identifying loop closures and correcting accumulated errors,the accuracy and robustness of localization can be significantly enhanced.However,most existing LiDAR-based loop closure detection methods primarily rely on low-level features such as coordinates and reflectivity to construct descriptors,failing to fully utilize semantic information within the scene.As a result,these methods often face challenges in terms of accuracy and reliability in complex scenarios.To address these limitations,this article proposes a semanticassisted intensity scan context method to overcome the insufficiencies of existing approaches.First,the proposed method employs the iterative closest point(ICP)algorithm for coarse registration of two-point clouds,reducing the impact of angular and translational errors on loop closure detection.On this basis,semantic features are integrated with the three-dimensional coordinates and reflectivity information of the point clouds to generate a global descriptor that incorporates multi-level features.Finally,loop closures are determined by calculating the similarity of the descriptors,enabling more reliable detection.Experimental results on the publicly available KITTI dataset show that the proposed method achieves a maximum F_(1) score improvement of 19.71%compared with the Scan Context algorithm,while reducing the average root mean square error(RMSE)by 36%compared with the lego-loam algorithm.Additionally,real-world experiments in a campus environment show that the proposed method improves the maximum F_(1) score by 19.23%compared with the LIOSAM algorithm and by 70.62%compared with the lego-loam algorithm.Furthermore,the average RMSE is reduced by 56.68%compared with LIO-SAM and by 20.7%compared with lego-loam.These results show that the suggested method not only greatly improves the accuracy of loop closure detection but also exhibits greater robustness in diverse scen

关 键 词：同步定位与建图 回环检测 语义分割 描述子 

分 类 号：TH701[机械工程—仪器科学与技术]