森林结构特征与大模型结合的地基点云单木定位

Single Tree Localization in Ground Laser Scanning Point Clouds Based on the Combination of Forest Structure Characteristics and Large Models

摘要: 森林场景的地面激光扫描数据空间结构复杂，冠层间郁闭度高，导致传统森林资源调查方法耗费大量人力物力。随着计算机视觉技术的发展，深度学习方法在交通、农业、林业等领域已取得显著成果。监督学习作为深度学习的核心范式，依赖大量标注数据训练模型，但其对数据质量和数量的要求较高，在标注数据稀缺的场景中应用受限。相比之下，基于自监督学习的大模型目标检测方法通过设计预训练任务，能够从大量未标注数据中提取通用视觉特征，具备强大的表征能力和泛化性能。这种方法不仅降低了对标注数据的依赖，还能适应更加复杂、多样化的目标检测场景。因此，针对现有问题，本文依据森林空间特有结构提出了一种基于大模型DINO-X Pro的单木定位方法。首先针对森林点云特性设计空间特征，将其输入网络，确定森林单木候选框，再根据候选框反向计算树干点。本文采用森林点云公开数据集进行单木定位实验，基于本文提出的方法在各样地单木定位中召回率，精确度，F₁指数分别达到了84.1%, 88.9%, 86.4%，与现有方法T-Rex对比均有较大提升，分别为31.1%, 17.6%, 25.8%；实验结果表明，本文方法的单木定位性能好，可靠性高，能为森林资源调查提供有效的参考数据。

Abstract: The complex spatial structure of ground-based laser scanning data in forest scenes and the high canopy closure make traditional forest resource survey methods labor- and resource-intensive. With the advancement of computer vision technology, deep learning methods have achieved significant success in fields such as transportation, agriculture, and forestry. As a core paradigm of deep learning, supervised learning relies on large amounts of labeled data to train models. However, its high demand for data quality and quantity limits its application in scenarios where labeled data is scarce. In contrast, self-supervised learning-based large-model object detection methods can extract general visual features from vast amounts of unlabeled data by designing pre-training tasks, demonstrating strong representational power and generalization capabilities. This approach not only reduces dependence on labeled data but also enables models to adapt to more complex and diverse object detection scenarios. To address existing challenges, this paper proposes a single-tree localization method based on the large model DINO-X Pro, leveraging the unique spatial structure of forests. First, spatial features tailored to forest point cloud characteristics are designed and input into the network to determine candidate bounding boxes for individual trees. Then, trunk points are back-calculated based on the candidate boxes. This study conducts single-tree localization experiments using publicly available forest point cloud datasets. The proposed method achieves recall, precision, and F₁-score of 84.1%, 88.9%, and 86.4%, respectively, across various sample plots. Compared to the existing T-Rex method, these metrics show significant improvements of 31.1%, 17.6%, and 25.8%, respectively. The experimental results demonstrate that the proposed method offers high single-tree localization performance and reliability, providing valuable reference data for forest resource surveys.