Solutions
Selected perception and robotic systems developed for field operation, integration and validation.
Autonomous robotic system for eucalyptus stem selection
Challenge: Selective thinning depends on an operator identifying which eucalyptus stems should be removed. This is time-consuming and difficult to perform consistently in dense, unstructured forest stands.
Approach: Fuse RGB-Depth data with computer-vision models to detect stems, assess the local stand and select candidates for removal.
Deployment: Integrate the perception pipeline with the robotic platform so that selected stems become actionable targets for the thinning operation.
Outcome: An end-to-end perception-to-action system that reduces reliance on manual stem selection by providing consistent, repeatable candidates for selective-thinning operations.
Computer vision algorithm
System integration
Multispectral tree-trunk classification
Challenge: Distinguish eucalyptus and pine tree trunks for forestry inventory and monitoring.
Approach: YOLOv5 and YOLOv8 detection and segmentation models trained on four-channel multispectral imagery (RGB + NIR).
Deployment: Training and evaluation pipelines adapted to four-channel image data.
Outcome: Tree-trunk detection and classification across two forest genera.
YOLOv5 four-channel repository · YOLOv8 four-channel repository
Vegetation-line clearing and implement control
Challenge: Perceive vegetation and coordinate implement control during line-clearing operations.
Approach: Computer-vision perception integrated with the implement-control interface and robotic system.
Deployment: System integration validated through interface and field-operation demonstrations.
Outcome: A perception and control workflow for automated vegetation-line clearing.
Perception system interface
System integration
3D perception and semantic mapping
Challenge: Build spatially meaningful representations of forest environments for autonomous machines.
Approach: Multispectral sensing, 3D object segmentation and semantic mapping.
Deployment: Integrated perception components that produce structured environmental information for downstream robotic tasks.
Outcome: 3D and semantic representations that support mapping and scene understanding.
Multispectral perception
3D object segmentation
Semantic mapping
Visible and thermal tree-trunk detection
Challenge: Maintain tree-trunk detection performance across sensor types and changing forest conditions.
Approach: Visible and thermal cameras paired with SSD, EfficientDet, YOLO, YOLOR and DETR-based object-detection models.
Deployment: Edge inference benchmarked across Raspberry Pi 4B, NVIDIA Jetson Nano, Google Coral USB Accelerator and OAK-D hardware.
Outcome: Comparative evidence for selecting sensors, models and edge hardware for forestry-monitoring robots.