基于改进LIO_SAM算法的复杂果园环境下三维地图精准构建

Precise 3D Map Construction in Complex Orchard Environments Based on Improved LIO_SAM Algorithm

摘要:

针对复杂果园环境下SLAM因点云稀疏性和低纹理特征导致的特征提取困难，以及场景相似性造成的回环检测场景区分能力不足等问题，本文提出一种基于改进LIO_SAM算法的复杂果园环境下三维地图构建算法。在前端预处理阶段通过地面分割算法Patchwork++将原始点云划分为地面点云和非地面点云;通过BFS算法对非地面点云进行精细聚类，以提高特征提取精度。在特征提取阶段，对所提取到的边缘特征进行关键点聚类并生成LinK3D特征描述子，增强复杂果园环境下特征点表征能力;后端回环检测基于LinK3D特征描述子构建在线词袋模型，实现实时回环检测，利用检测到的回环信息，完成六自由度位姿校正。通过因子图将回环因子融入全局优化，构建全局一致性地图。通过KITTI公开数据集测试了改进算法性能，结果表明，对比改进前算法，改进LIO_SAM算法绝对位姿均方根误差减少23.51%，标准误差降低26.4%。在复杂果园环境下，与同类算法在巡检机器人平台进行验证，试验结果表明，本文算法所得点云地图质量更高，轨迹与真实轨迹贴合度更好，在冬季和夏季果园环境下绝对位姿误差最大值分别为0.261 m和0.239 m，标准差分别为0.038 m和0.034 m。相较于同类算法，本文算法可在复杂果园场景下实现更高精度的定位与建图作业。

Abstract:

Aiming at the problem that SLAM in complex orchard environments is difficult to extract features due to the sparsity and low texture features of point clouds, and the loop detection scene differentiation ability is insufficient due to scene similarity, a 3D map construction algorithm based on the improved LIO_SAM algorithm was proposed in complex orchard environments. Firstly, in the front-end preprocessing stage, the original point cloud was divided into ground point cloud and non-ground point cloud by the ground segmentation algorithm Patchwork++;the non-ground point cloud was finely clustered by the BFS algorithm to improve the accuracy of feature extraction. In the feature extraction stage, the extracted edge features were clustered into key points and the LinK3D feature descriptor was generated to enhance the representation ability of feature points in complex orchard environments;the back-end loop detection built an online bag-of-words model based on the LinK3D feature descriptor to realize real-time loop detection, and used the detected loop information to complete the correction of the six-degree-of-freedom pose. Finally, the loop factor was integrated into the global optimization through the factor graph to construct a global consistency map. The performance of the improved algorithm was verified by testing the KITTI public dataset. The results showed that compared with the algorithm before improvement, the root mean square error of the absolute pose of the improved LIO_SAM algorithm was reduced by 23.51%, and the standard error was reduced by 26.4%. In order to verify the superiority and generalization of the improved algorithm in complex orchard environments, it was verified with similar algorithms on the inspection robot platform. The positioning and mapping test results in complex orchard environments showed that the point cloud map obtained by the proposed algorithm was of higher quality, and the trajectory fit the real trajectory better. The maximum value of absolute pose error in winter and summer orchard environments was controlled within 0.261 m and 0.239 m, respectively, and the standard deviation was maintained within 0.038 m and 0.034 m, respectively. Compared with similar algorithms, the proposed algorithm can achieve higher-precision positioning and mapping operations in complex orchard scenes.