Speaker
Description
Forest environments pose significant challenges for automated driving due to dense vegetation, irregular terrain, varying illumination conditions, narrow skid roads, and limited or unreliable GNSS availability. In contrast to structured road environments, skid roads often lack clear boundaries and may contain unpredictable obstacles such as branches, uneven ground, and irregularly spaced trees. Reliable automated driving in such environments therefore requires robust perception, mapping, and navigation components that remain stable under highly unstructured and dynamically changing conditions.
To address these challenges, a multi-modal perception and navigation framework has been developed and deployed on a full-scale forestry machine. The framework integrates deep learning-based visual perception with LiDAR-based mapping and localization and extends the navigation pipeline with occupancy-aware path planning.
For semantic perception, a state-of-the-art YOLO26 model is trained using a combination of newly annotated forestry datasets and previously collected data. The dataset includes skid roads, tree trunks, and typical environmental structures encountered during timber extraction. The trained model is deployed directly on the forestry machine and enables real-time detection while driving along skid roads. This visual module provides semantic understanding of the driving corridor and surrounding objects, supporting corridor recognition and obstacle awareness under varying lighting conditions and partial occlusions, thereby establishing a foundation for vision-supported driving automation functions in forestry applications.
In parallel, a LiDAR-based landmark extraction method was implemented for geometric mapping and localization. Trees with distinct cylindrical trunk characteristics are detected in real time and parameterized as geometric primitives. These tree landmarks are incorporated into a SLAM framework to build a consistent map and estimate vehicle pose during operation. While this landmark-based approach provides accurate localization when trunk geometry is clearly observable, performance can degrade in scenarios where trunk features are irregular, partially occluded, or less distinguishable.
To improve robustness, an occupancy grid mapping module was introduced as a complementary environmental representation. Rather than relying exclusively on explicit trunk landmarks, the occupancy grid captures the spatial distribution of nearby objects along skid roads and provides a denser representation of occupied space. This representation compensates for situations in which landmark detection becomes unreliable and offers improved environmental awareness in complex forest conditions.
Building upon the perception and mapping modules, a multi-modal path planner, referred to as FlowPlanner, is developed. The planner integrates information from tree landmarks, tree clusters, and reference paths to generate a combined guidance flow field. Based on this flow representation, multiple candidate trajectories are generated online and evaluated with respect to alignment, safety margins, and kinematic feasibility. The most suitable candidate is selected as the local goal path. This approach enables real-time generation of feasible driving trajectories and supports exploratory navigation in partially unknown forest environments.
Initial field experiments in real forest conditions indicate the functional feasibility of the proposed framework. The modular system architecture supports incremental refinement and integration of additional perception and planning components, providing a scalable foundation for progressive automation in demanding forestry operations.
| Keywords | Automation; Perception; SLAM; Navigation |
|---|
