Authors

Abstract

To address the challenges posed by severe canopy obstruction in orchards and complex terrain—conditions under which single-GNSS systems frequently lose signal lock, as well as the susceptibility of single-sensor systems to dynamic interference—this paper proposes a multi-source fusion positioning framework based on an adaptive extended Kalman filter (AEKF), integrating GNSS-RTK and 3D LiDAR. To overcome the issue of sparse and discontinuous point cloud features in agricultural environments, a ground segmentation method based on a concentric zone model combined with an improved RANSAC algorithm is developed. This approach enables high-frequency and accurate extraction of orchard row geometric features under complex conditions, including muddy ruts and dynamic human interference, thereby establishing reliable observational constraints for local relative pose estimation. An adaptive observation noise covariance adjustment mechanism based on signal confidence is further proposed. By continuously monitoring RTK quality indicators and accuracy metrics in real time, the system dynamically suppresses unreliable state updates during periods of GNSS signal degradation and seamlessly switches to an error compensation mode based on lateral and heading constraints derived from 3D LiDAR. This effectively mitigates cumulative drift associated with dead reckoning. Experimental results demonstrate that, under challenging conditions involving intermittent canopy gaps and dynamic occlusions, the proposed system achieves a root mean square error (RMSE) of 0.042 m for lateral positioning over the entire trajectory, while the heading RMSE is maintained within 1.85°. The proposed approach effectively addresses the problem of intermittent localization loss in complex orchard environments, providing a robust state estimation framework that enables agricultural robots to operate without reliance on prior mapping, while supporting high-precision global path planning and real-time local obstacle avoidance.

Abstract in Chinese

针对果园树冠严重遮挡与复杂地形导致单一全球导航卫星系统易失锁，以及单一感知手段易受动态干扰的难题，本文提出了一种基于自适应扩展卡尔曼滤波的GNSS-RTK与3D激光雷达多源融合定位框架。针对农业环境点云特征离散的问题，设计了基于同心区模型的地面分割及改进的RANSAC算法。该算法能够在泥泞车辙和动态人员干扰等复杂工况下，实现果树行几何特征的高频精准提取，从而构建局部相对位姿的观测约束。提出了一种基于信号置信度的自适应观测噪声协方差调整机制。系统通过实时监测RTK质量标志位与精度因子，在GNSS信号退化阶段自适应阻断不可靠状态更新，并无缝切换至基于3D激光雷达侧向与航向约束的误差补偿模式，有效抑制了航位推算中的累积漂移。实验结果表明，在存在连续缺株与动态遮挡的严苛条件下，系统全过程横向定位均方根误差RMSE仅为0.042m，航向偏差RMSE被严格控制在1.85°以内。本研究有效克服了复杂果园环境中的“定位盲区”问题，为农业机器人摆脱预建图依赖、实现高精度的全局路径规划与实时局部避障提供了高鲁棒性的状态估计基础。