基于Snake模型多基元融合的点云滤波方法研究

Point cloud filtering is a necessary step in the point cloud processing, analysis and applications. To break through the bottlenecks of the existing filtering methods, including low filtering accuracy in complex terrain environments, poor robustness, and too many parameter settings, this project will focus the research on fusion of multiple primitives for point cloud filtering based on the Snake model. Multiple primitives fusion filtering strategy will be adopted to avoid the drawback of poor robustness of filtering methods based on single primitive. In so doing, the adaptability of the filtering method to the complex terrain environments can be improved by making full use of the characteristic information provided by different primitives. The main contribution of the project is to develop the Snake model. By setting the target energy function with adsorption force to the ground points, the model can take both the local scale information and the global scale information of point cloud into account, which can fully exploit the automation potential of point cloud data itself. The filtering thresholds can be expressed as a linear function of terrain slope changes. In this way, the adaptive updating of filtering thresholds can be implemented so as to avoid the drawbacks of low filtering accuracy with single threshold and too many parameter settings involved for multiple thresholds. This project is expected to develop the filtering method with higher filtering accuracy, robustness, and automation degree in complicated terrain areas, and provide fundamental technical supports for the post-processing application of point clouds, such as three-dimensional building modeling, surface coverage classification, biomass estimation, etc. The research results cover wide application prospects.

点云滤波是点云数据处理、分析和应用的必经环节，为突破滤波方法在复杂地形区域滤波精度低、鲁棒性差以及需要过多参数设置的瓶颈，本项目拟基于Snake模型进行多基元融合的滤波方法研究。拟采用多基元融合的滤波策略，可以有效避免单一基元滤波法鲁棒性差的缺陷，充分利用不同基元提供的特征信息，提高滤波方法对复杂地形区域的适应性。主要构建Snake模型，通过设置具有地面点吸附力的目标能量函数，使得模型能够兼顾点云的局部尺度信息与全局尺度信息，充分挖掘点云数据本身的自动化潜力。依据该模型将滤波阈值表示为地形坡度变化的线性函数，实现滤波阈值的自适应更新，避免单一阈值滤波精度差而多阈值需要过多参数设置的缺陷。此项研究有望获取在复杂地形区域具有较高滤波精度、鲁棒性以及自动化程度的滤波方法，为建筑物三维建模、地表覆盖分类、生物量估算等点云后处理应用提供基础支持，研究成果具有广泛的应用前景。

机载LiDAR点云滤波是点云后处理应用的关键环节。本项目针对点云滤波所存在的在复杂地形区域滤波精度低、鲁棒性差以及需要过多参数设置等技术瓶颈，进行了滤波理论方法的研究。主要研究内容和成果如下：（1）针对复杂地形区域滤波精度较低的难题，本项目提出了一种基于最佳滤波窗口自适应估测的形态学滤波方法，该方法有效突破了传统形态学方法受滤波窗口设置影响较大的瓶颈，能够在地形断裂区域或者地形凸起区域有效保护地形细节信息，有效降低拒真误差、提高滤波精度；（2）针对森林区域因树冠遮挡而造成地面点云较少、滤波精度较低的问题，本项目提出了基于Mean Shift分割的森林区域点云滤波方法。该方法首先采用Mean Shift分割方法实现树冠冠幅估测，提升森林区域点云滤波的自动化程度。为提高地形坡度变化较大区域的滤波精度，本项目进行了地形去趋势化研究，有效避免了地形坡度变化较大的影响；（3）提出了一种基于高斯混合模型分离的点云滤波方法。该方法将包含有地面点和地物点的LiDAR点云数据视为高斯混合模型的叠加，通过采用期望最大法来实现高斯混合模型参数的估测，进而实现高斯混合模型的分离，从而达到地面点和地物点分离的目的。该方法有效实现了高自动化的无参点云滤波。（4）将人工智能方法应用于点云滤波中，提出了一种基于主动学习的点云滤波方法。该方法无需进行人工样本标记，能够通过不断渐进地学习地面点云和非地面点云的特征，来自动实现地面点云和非地面点云的分类。（5）将项目研究成果在点云后处理研究中进行了应用拓展，包括建筑物提取、单木分割、枝叶分离等。本项目研究通过构建不同类型、不同特点的点云滤波方法，进一步丰富了点云滤波的理论体系，将对LiDAR点云智能处理应用的进一步发展起到积极的作用。