基于Huber损失的点云稳健配准算法

Robust Point Cloud Registration Algorithm Based on Huber Loss

摘要:

为提高农业机械在非结构化农田环境下自主导航的可靠性，解决点云配准对初始位姿敏感且在动态植被干扰与剧烈振动下易失效的问题，本研究提出一种面向农机导航的Huber稳健点云配准算法（Huber-ICP）。该方法将Huber稳健核函数嵌入点到面ICP框架，通过Huber损失重构目标函数，并采用迭代重加权最小二乘（IRLS）进行优化，从而自适应抑制异常点（如晃动枝叶、尘土）的影响。在BotanicGarden数据集多个序列上的试验结果表明，相较于标准点到点ICP与点到面ICP，Huber-ICP绝对轨迹误差（ATE）分别降低67.97%与46.53%;相较于Pi2Pi-ICP和Pi2Pl-ICP，平均迭代次数减少46.22%和14.74%，单帧配准耗时缩短11.43%和35.42%;在1.0 m平移扰动下，配准成功率仍保持在80%以上。本研究提出的动态加权机制在无需手动调参的前提下，有效兼顾了内点精度与异常点鲁棒性，为农机在GNSS拒止、振动剧烈等复杂田间环境中的稳定实时定位提供了可靠解决方案。

Abstract:

Reliable point cloud registration is essential for agricultural machinery navigation in unstructured field environments, but conventional ICP methods are easily affected by large initial pose errors, dynamic vegetation interference and vibration-induced noise. To improve registration robustness and efficiency, a robust point-to-plane ICP method based on the Huber loss, named Huber-ICP, was proposed. The Huber robust kernel was introduced into the point-to-plane ICP objective function, and the transformation parameters were estimated through an iteratively reweighted least squares optimization process. In this way, point pairs with large residuals were assigned lower weights, whereas the contribution of inlier correspondences was retained. Experiments were conducted on sequences 00, 02, 05, 07 and 08 of the public BotanicGarden odometry dataset to evaluate the proposed method under different navigation conditions. Compared with point-to-point ICP and point-to-plane ICP, Huber-ICP reduced the absolute trajectory error by 67.97% and 46.53%, respectively. Compared with Pi2Pi-ICP and Pi2Pl-ICP, the average number of iterations was reduced by 46.22% and 14.74%, and the per-frame registration time was reduced by 11.43% and 35.42%. Under a translational perturbation of 1.0 m, the registration success rate remained above 80%. The results indicated that the dynamic reweighting strategy improved both outlier suppression and convergence efficiency. The proposed method can provide a practical registration solution for continuous localization of agricultural machinery in GNSS-denied and high-vibration field environments.