Viking Hill Dataset: A Lidar-Radar-Camera Dataset for Detection and Segmentation in Forest Scenes

Technical Summary: Viking Hill Dataset

Problem Statement
Autonomous robots operating in forest environments face significant perception challenges due to harsh conditions, including uneven terrain, dense vegetation, and sensor contamination (dirt, sawdust, debris). While Unmanned Ground Vehicles (UGVs) are increasingly used for forestry tasks, existing perception systems relying on cameras and LiDAR are susceptible to visual degradation and occlusion. Although 4D imaging radar offers resilience to these conditions, its application in forestry remains under-explored. A critical gap exists in the availability of co-registered, multi-modal datasets (LiDAR, Radar, Camera) with shared 3D annotations that allow for direct, fair comparison of segmentation performance across these modalities in natural forest settings.

Methodology
The authors present the Viking Hill Dataset, a multi-sensor dataset collected by a Clearpath Husky mobile robot in a forest near Örebro University, Sweden. The data collection occurred in two sessions (May and June 2024) to capture contrasting vegetation states: "short grass" (low undergrowth) and "tall grass" (peak undergrowth).

• Sensor Suite: The robot was equipped with a high-resolution 4D FMCW imaging radar (Sensrad Hugin A3), a 32-channel LiDAR (Ouster OS0-32), an RGB camera (IDS uEye), an IMU (Xsens MTi-30), and dual RTK-GNSS receivers.
• Data Processing: A reference point-cloud map was generated using SLAM (HDL Graph SLAM) aided by high-accuracy GNSS poses where available. Due to canopy occlusion limiting GNSS accuracy, the maps exhibit some global drift but remain locally accurate.
• Annotation: The dataset features 3D cuboid annotations for nine classes (e.g., Ground, Tree Trunk, Tree Canopy, Rock, Bush). These labels were manually created on the reference LiDAR map using Segments.ai tools and projected onto the radar and camera data. This approach was chosen because manual annotation of sparse radar point clouds is impractical.
• Per-Tree Metrics: The dataset includes approximate Diameter at Breast Height (DBH) annotations for 149 (low vegetation) and 169 (high vegetation) tree trunks, derived by fitting circles to accumulated LiDAR points at 1.3m height.
• Baseline Evaluation: The authors trained a MinkowskiUNet architecture (adapted from the DigiForests dataset) for semantic segmentation on both LiDAR and radar point clouds. They also conducted a cross-modality analysis comparing LiDAR and radar segmentation against a pre-trained RGB trunk detection model (PercepTreeV1) and performed a stratified evaluation based on tree diameter.

Key Contributions

1. First Multi-Modal Forestry Dataset: The first publicly available dataset combining co-registered 4D imaging radar, LiDAR, and RGB camera data in a forest environment with shared 3D annotations.
2. ROS-Based Tooling: Development of ROS tools for labeling point clouds and projecting 3D cuboids onto RGB images, supporting formats like COCO.
3. Per-Tree Diameter Annotations: Inclusion of DBH estimates to enable stratified analysis of segmentation performance relative to tree size.
4. Baseline Results: Provision of semantic segmentation baselines using MinkowskiUNet for both LiDAR and radar, along with cross-modal performance analyses.

Results

• Semantic Segmentation: The baseline MinkowskiUNet achieved competitive Intersection over Union (IoU) scores for dominant classes. For the "Ground" class, radar achieved 90.5% IoU (5-scan accumulation) compared to 98.4% for LiDAR. For "Tree Canopy," radar reached 85.5% vs. 91.0% for LiDAR. However, radar performance lagged significantly on geometrically fine structures; for "Tree Trunk," radar achieved 55.8% IoU compared to 74.3% for LiDAR.
• Cross-Modality Analysis: When analyzing tree trunk segmentation conditioned on RGB detection success:
  • In low vegetation, LiDAR maintained consistent recall regardless of whether the RGB model succeeded or failed.
  • Radar performance dropped significantly when the RGB model failed (False Negatives), largely because many missed detections were located at the edges of the radar's azimuth range or outside the camera frame, leading to sparse point returns.
  • Radar demonstrated the ability to detect trunk points up to 50m in low vegetation, though with higher variance in recall compared to LiDAR.
• Diameter-Stratified Evaluation: Segmentation quality varied with tree size. For LiDAR, IoU was highest for thin trees (65.1%) and lowest for thick trees (55.0%) in the low vegetation session. This suggests that aggregate class metrics may mask performance variations across different object sizes.

Significance and Claims
The paper claims that the Viking Hill Dataset is a unique resource for investigating the utility of imaging radar in forestry perception. The authors assert that while radar currently lags behind LiDAR in segmenting fine geometric details like tree trunks, it provides competitive performance for dominant classes (ground and canopy) and offers robustness in conditions where visual sensors fail.

The dataset is intended to stimulate research in:

• Sensor Fusion: Combining radar, LiDAR, and camera data for redundant perception in autonomous forestry.
• Robustness: Investigating sensor performance under vegetation occlusion and potential surface contamination.
• Mapping and Localization: Utilizing the RTK-GNSS-aided reference maps for SLAM and terrain analysis under canopy.

The authors maintain a modest stance, acknowledging limitations such as the use of 3D cuboids (which may introduce label noise for complex shapes like forks or branches) and the current lack of Doppler integration in the baseline segmentation. They position the dataset as a foundation for future work to refine annotation methods and explore radar's full potential in challenging natural environments.