基于三维候选区的深度学习高密度机载激光雷达森林点云单木探测的研究

Single tree detection, which can accurately obtain the number and locations of single trees, is the premise of large-scale fine forest investigation. Nowadays, airborne Lidar can generate high-density point clouds, which provides the opportunity for accurate single tree detection. However, the low performance of features, underutilized spatial information of high density point clouds, and the detection results of subdominant trees which are affected by the detection results of dominant trees lead the single tree detection methods in forest areas to have low detection accuracy and do not meet practical demands. In order to solve the above problem, this research will propose a single tree detection network for accurate single tree detection based on the establishment of three-dimensional candidate regions and extraction of candidate region features, by the ideas of region proposal networks and PointNet networks. The research consists of three main contents. A single tree proposal network which can detect and extract single tree candidate regions directly in any forest locations will be established. A single tree feature extraction network which can consider treetops and extract features from the point clouds in candidate regions automatically will be established. The single tree proposal network and the single tree feature extraction network are combined to form the single tree detection network. To further improve the performance of features, the adaptability to high-density point cloud, and the adaptability to different environment and data, optimization and training methods are studied. The research will provide a new approach and though for the single tree detection, and promote the development of automatic forest investigation.

单木探测实现精确获取森林单木数量及位置，是大规模森林精细调查的前提。当前机载激光雷达已能产生高密度点云，为精确单木探测提供了可能。但当前方法中存在特征性能低且无法充分利用高密度点云空间信息，以及下层木探测结果依赖上层木探测结果的问题，导致单木探测方法森林地区探测精度低，实用性差。为解决上述问题，本项目将借鉴区域提名网络和PointNet类网络的思路，以建立三维候选区并提取候选区特征的方式，构建单木探测网络，实现精确单木探测。本项目主要研究内容为三部分，建立能直接对森林任意位置，探测并提取单木候选区的单木提名网络；建立顾及树顶点预测的候选区内点云自动特征提取的单木特征提取网络，单木提名网络和单木特征提取网络组合形成单木探测网络；研究网络优化和训练，提高特征鲁棒性，对高密度点云的适应性，以及不同环境、数据的适应性。本项目研究成果将为精确单木探测提供新方法和新思路，也为森林调查自动化提供支持。

单木探测实现精确获取森林单木数量及位置，是大规模森林精细调查的前提。当前方法中存在特征性能低且无法充分利用高密度点云空间信息，以及下层木探测结果依赖上层木探测结果的问题，导致单木探测方法森林地区探测精度低，实用性差。为解决上述问题，本项目主要研究内容为三部分，建立能直接对森林任意位置，探测并提取单木候选区的单木提名网络；基于深度学习建立的单木特征提取网络;研究网络优化和训练，建立数据库获得性能良好的单木探测网络。基于这三个研究内容，本项目提出了Tree RCNN实现了高精度的单木提取。本项目研究成果将为精确单木探测提供新方法和新思路，也为森林调查自动化提供支持。