基于多线激光雷达的主干形果树树干层级检测方法

Hierarchical Trunk Detection Method of Single-stem Fruit Trees Based on Multi-layer LiDAR

摘要:

针对复杂果园环境行间导航树干检测问题，提出一种基于多线激光雷达（Light detection and ranging，LiDAR）的主干形果树树干层级检测方法，使用16线VLP-16型LiDAR采集车辆周围的果园点云数据，通过目标分割和树干检测2个步骤层次化检测树干，去除非树干目标，提高树干检测精度。首先，设置环形感兴趣区域（Region of interest，ROI），采用地面拟合算法移除地面点云，消除果园目标点云之间的连通性；其次，设置矩形ROI，采用基于密度的带噪声空间聚类（Density-based spatial clustering of applications with noise，DBSCAN）算法对非地面点云进行xOy平面聚类，根据LiDAR测量分辨率和果园目标参数设置DBSCAN算法超参数，将非地面点云分割为若干目标簇；然后，从全局和局部2个尺度提取目标簇的几何和强度特征，用这些特征描述树干与其他果园目标间的差异；最后，采用训练好的树干检测器融合特征，将目标簇划分为树干与非树干2个类别，输出树干簇。树干检测步骤采用随机森林（Random forest，RF）算法进行离线特征选择与融合，使用树干和非树干训练样本，基于基尼指数（Gini index，GI）改变量评价特征重要性，从初始特征中选择22个鉴别力较强的特征，再融合这些特征生成树干检测器。实验场景为标准化种植核桃园，共采集1317帧点云数据，从中分割12213个目标簇，其中，树冠、杂草、支撑杆、围栏、土坡、农具、行人等非树干目标占比58.04%。按照帧比例1∶4将目标簇随机划分为训练集和测试集，测试集树干检测精确率为99.16%、召回率为99.21%、F1分数为99.19%，树干层级检测平均帧耗时85.25ms。本文方法能对复杂果园场景快速、精准地检测出树干，满足果园行间导航对树干检测的准确性和实时性要求。

Abstract:

Aiming to address the trunk detection challenge in inter-row navigation within complex orchard environments, a hierarchical detection method for central-leader training fruit tree trunks was proposed based on multi-layer light detection and ranging (LiDAR). A 16-layer VLP-16 LiDAR was employed to collect orchard point cloud data around the vehicle, and a two-step hierarchical trunk detection approach, i.e., object segmentation followed by trunk identification was adopted to eliminate non-trunk objects and improve detection accuracy. Firstly, an annular region of interest (ROI) was defined, and a ground fitting algorithm was applied to remove ground point clouds, disrupting the connectivity between object point clouds in the orchard. Secondly, a rectangular ROI was set, and the density-based spatial clustering of applications with noise (DBSCAN) algorithm was used for xOy plane clustering of non-ground point clouds. Hyperparameters of the DBSCAN algorithm were configured based on LiDAR measurement resolution and orchard object parameters, splitting non-ground point clouds into multiple object clusters. Thirdly, geometric and intensity features of object clusters were extracted at both global and local scales to characterize the differences between trunks and other orchard objects. Finally, a pre-trained trunk detector fused these features to classify object clusters into trunk and non-trunk categories, outputting trunk clusters. For the trunk detection step, the random forest (RF) algorithm was utilized for offline feature selection and fusion. Using trunk and non-trunk training samples, feature importance was evaluated based on changes in the Gini index (GI), and 22 highly discriminative features were selected from the initial feature set to construct the trunk detector. Experiments were conducted in a standardized walnut orchard, where 1317 frames of point cloud data were collected, yielding 12213 segmented object clusters. Non-trunk objects, including crowns, weeds, support poles, fences, slopes, farm tools, and pedestrians, accounting for 58.04% of the total clusters. The object clusters were randomly divided into training and test sets at a ratio of 1∶4. Test results showed a trunk detection precision of 99.16%, recall of 99.21%, and F1-score of 99.19%, with an average frame processing time of 85.25ms for hierarchical trunk detection. This method enabled fast and accurate trunk detection in complex orchard scenarios, meeting the accuracy and real-time requirements of trunk detection for orchard inter-row navigation.