Abstract: In order to realize the automatic navigation driving of field management machinery and improve the recognition accuracy of seedling belt trajectory, this paper proposed an extremum point clustering seedling zone identification algorithm based on offset compensation model. Firstly, based on the steering gear leveling unit, the seedling belt data were collected by lidar to construct a 2.5 D point cloud image, and establish an offset compensation model through mathematical modeling to obtain correction parameters. Then, the cluster feature points were obtained through the extreme value detection algorithm, and hierarchical clustering combined with offset compensation model. Finally, the least square method was used to linearly fit the obtained seedling belt cluster points to realize the reconstruction of the seedling belt trajectory. Through the Matlab simulation of point cloud information for machine-transplanted paddy rice with a forward vision range of 2 000～3 000 mm and a standard row spacing of 220 mm. When the obtained seedling belt trajectory was compared with the real seedling belt trajectory, the maximum lateral deviation was 16 mm, the maximum median deviation was 6 mm, and the calculation took 1.50 s. Compared with the slope blur clustering seedling belt recognition algorithm, the maximum axial deviation reduced by 29 mm, and the operation time reduced by 0.14 s, which improved the accuracy of trajectory coincidence and enhanced the real-time performance of the algorithm. Above results provided theoretical basis for the navigation and driving of low-injury seedlings for paddy field management machinery.

Key words: lidar, seedling belt identification, extreme value detection, offset compensation model, improved K-means mean clustering

摘要： 为实现田间管理机械自动导航驾驶，提高苗带轨迹识别精度，提出了一种基于偏移补偿模型的极值点聚类苗带识别算法。首先，以舵机调平单元为基础，利用激光雷达采集苗带数据构建2.5 D点云图像，通过建立偏移补偿模型获得矫正参数。然后，通过极值检测算法获得聚类要素点，结合偏移补偿模型进行分层聚类。最后，通过最小二乘法对获取的苗带聚类点进行线性拟合，实现苗带轨迹的重构。通过对前视范围2 000～3 000 mm且标准行距为220 mm的机插秧水稻进行点云信息Matlab仿真，获得的苗带轨迹与真实苗带轨迹相比，最大横向偏差为16 mm，最大中位偏差为6 mm，运算耗时1.50 s。相比斜率虚化聚类苗带识别算法，最大横向偏差减小了29 mm, 运算耗时减少0.14 s，提高了轨迹重合精度，增强了算法的实时性，可为水田管理机械低伤苗导航行驶提供理论依据。

关键词: