Abstract
We present a real-time system for per-tree canopy volume estimation using mobile LiDAR data collected during routine robotic navigation. Unlike prior approaches that rely on static scans or assume uniform orchard structures, our method adapts to varying field geometries via an integrated pipeline of LiDAR-inertial odometry, adaptive segmentation, and geometric reconstruction. We evaluate the system across two commercial orchards, one pistachio orchard with regular spacing and one almond orchard with dense, overlapping crowns. A hybrid clustering strategy combining DBSCAN and spectral clustering enables robust per-tree segmentation, achieving 93% success in pistachio and 80% in almond, with strong agreement to drone derived canopy volume estimates. This work advances scalable, non-intrusive tree monitoring for structurally diverse orchard environments.
Abstract (translated)
我们提出了一种实时系统,用于利用在机器人常规导航过程中采集的移动LiDAR数据来进行单棵树木冠层体积估算。与依赖静态扫描或假设果园结构均匀的传统方法不同,我们的方法通过集成的里程测量(LiDAR-惯性)、自适应分割和几何重建管道,能够适应变化多样的田野几何形状。我们在两个商业果园中评估了该系统的效果:一个是规则间距的开心果果园,另一个是密集、冠层重叠的杏仁果园。我们采用了一种结合DBSCAN与谱聚类的混合聚类策略,实现了鲁棒性的单树分割,在开心果果园中的成功率为93%,在杏仁果园中为80%;并且我们的估算结果与无人机获取的冠层体积高度一致。这项工作推动了结构多样化的果园环境中可扩展且非侵入式的树木监测技术的发展。
URL
https://arxiv.org/abs/2506.08061