Speaker
Description
Forest harvesting operations increasingly require real-time, non-destructive tree-level information to support precision decision-making directly within machine workflows. Although harvester-mounted Mobile Laser Scanning (MLS) systems provide detailed three-dimensional representations of tree stems, most existing processing approaches depend on full point cloud reconstruction, frame alignment, and Simultaneous Localization and Mapping (SLAM). These pipelines are computationally demanding and typically require post-processing, limiting their applicability for real-time decision support in operational environments. Consequently, sensing, measurement, and bucking optimisation are often addressed as sequential and loosely connected tasks, constraining the development of fully integrated, sensor-driven harvesting systems.
This study presents an integrated real-time framework that operates directly on raw MLS data through two-dimensional projections of individual MLS frames, thereby avoiding explicit 3D reconstruction. Each frame is independently transformed into a structured 2D representation that enables stem instance segmentation using a 2D Convolutional Neural Network (CNN). From the segmented projections, geometric attributes such as stem diameters and longitudinal profile are estimated sequentially, allowing the progressive construction of a dynamic stem representation during machine operation. These measurements are directly coupled with an optimisation module implementing an optimal bucking strategy based on predefined product specifications and operational constraints. By integrating perception, measurement, and optimisation within a unified continuous processing pipeline, the proposed framework establishes a reconstruction-free approach to real-time, sensor-driven bucking decisions and contributes to advancing automation in forest harvesting systems.
| Keywords | Precision_harvesting; Optimal_bucking; LiDAR; Instance_segmentation |
|---|
