基于车载LiDAR与深度学习信息融合的行道树单木参数提取方法研究

Vehicle-borne LiDAR measurement technology can quickly acquire high-resolution and high-precision tree vertical structure data, which can improve the difficult and long-term traditional resource survey method of street tree and achieve a large-scale, comprehensive and rapid acquisition of tree parameter information at single-tree-scale. This project intends to use deep learning information fusion method to study the problem of single tree parameter extraction of street tree based on vehicle-borne LiDAR, so as to improve the accuracy of street tree parameter extraction. A human-computer interactive intelligent annotation method will be proposed, which can quickly label point cloud data set of street tree with various complex scenes. A deep learning instance segmentation method will be used to fuse point cloud information such as three-dimensional coordinates, the correction values of laser reflectance intensity and the echo numbers, to mine the semantic features of street tree in multiply levels from the data set, and to establish an end-to-end point cloud segmentation model to solve the segmentation problem due to point cloud adhesion and similar-shaped objects. Multi-angle LiDAR scanning will be used to obtain the real crown volume and the data set of individual tree parameters will be established. The deep neural network will be used to mine the non-linear regression relationship between crown volume and the easily measured factors such as DBH, tree height, crown width, side-looking crown volume from the data set. The crown volume prediction model will be established, which predicts the crown volume by the easily measured factors and can solve the problem of point cloud deficiency.

车载LiDAR测量技术能够快速获取高分辨率、高精度的树木垂直结构数据，改善难度大、周期长的传统行道树资源调查作业模式，实现大范围、全面快速地获取单木尺度的行道树参数信息。本项目拟采用深度学习信息融合手段，研究车载LiDAR行道树单木参数提取方法，提高行道树单木参数提取精度。提出人机交互智能标注方法，构建包含多样复杂场景的行道树点云标注数据集。采用深度学习实例分割方法融合点云三维坐标、激光反射强度校正值、回波次数，从标注数据集中挖掘多层次行道树语义特征，建立端到端的行道树点云分割模型，解决由于点云粘连、相似形状地物导致的行道树点云错误分割。采用多角度LiDAR扫描获取树冠体积实测值，建立单木参数数据集。采用深度神经网络从数据集中挖掘树冠体积与胸径、树高、冠幅、侧视树冠体积等易测因子的多元回归关系，建立树冠体积预测模型，利用易测因子估计树冠体积，解决点云缺失时行道树参数难以准确测量的问题。

车载LiDAR测量技术能够快速获取高分辨率、高精度的树木垂直结构数据，改善难度大、周期长的传统行道树资源调查作业模式，实现大范围、全面快速地获取单木尺度的行道树参数信息。本项目主要工作为：.（1）激光强度校正。基于直线滑台2D LiDAR测量系统采集的点云数据，建立激光强度校正模型。在反射率为50%的标准漫反射板上采用分段多项式回归建立距离校正模型，采用线性回归建立入射角校正模型。实验采集一面白墙的强度数据，校正后强度变异系数≤0.0063，校正后与校正前变异系数比值≤0.280，有效消除了距离和入射角变化对激光强度的影响。.（2）行道树实例分割。基于车载2D LiDAR测量系统采集的点云数据，建立行道树实例分割模型。首先提取点云深度坐标、一次回波强度和回波次数，构建三通道街道图像；然后，对图像进行切片处理，采用图像实例分割算法YOLACT训练行道树图像实例分割模型；最后，将分割结果映射回点云并进行分割优化，实现行道树点云实例分割。最终建立的行道树实例分割模型准确率为0.9865，召回率为0.9407，F1分数为0.9576，帧平均处理时间为5.261ms。.（3）行道树参数提取。基于分割得到的行道树单木点云数据，计算行道树树高、冠幅和胸径，胸径拟合优度为0.9441，均方误差为2.6cm。建立车载2D LiDAR点云数据的格网索引，根据二次回波情况将测量点分为三个类型，分别计算格网面积，采用二次多项式拟合总格网面积与总叶面积的回归关系，拟合优度为0.9617，均方误差为1404.1652cm2，帧平均处理时间为1.0787ms，与基于总点云数的总叶面积估测方法相比，拟合优度提升了0.3858，克服了LiDAR移动速度和测量距离引起的点云密度变化对估测精度的影响。.项目成果能够提升行道树资源调查的智能化水平，为实现快速、高效、准确地行道树参数提取做出贡献。

内容获取失败，请点击重试