OSM-BKI: OpenStreetMap-Enhanced Bayesian Kernel Inference for Semantic Mapping

This project extends Lu Gan's Semantic BKI (S-BKI) with OpenStreetMap (OSM) geographic priors to improve 3D semantic mapping accuracy. The core BKI mapping engine is adapted for ROS2 and augmented with OSM-based prior fusion, uncertainty-aware inference, and multi-dataset support.

Building on S-BKI

The original S-BKI work by Lu Gan et al. introduced Bayesian Spatial Kernel Smoothing for scalable dense semantic mapping. It uses Bayesian Kernel Inference to propagate semantic observations through an octree-based 3D map, producing continuous semantic predictions with calibrated uncertainty.

Original paper:

L. Gan, R. Zhang, J. W. Grizzle, R. M. Eustice, and M. Ghaffari, "Bayesian Spatial Kernel Smoothing for Scalable Dense Semantic Mapping," IEEE Robotics and Automation Letters (RA-L), 2020.

Original repository: github.com/ganlumomo/BKISemanticMapping

What OSM-BKI Adds

This project extends S-BKI in several key ways:

OSM Prior Integration

• Geographic priors from OpenStreetMap: Buildings, roads, sidewalks, cycleways, parking areas, grasslands, trees, forests, fences, walls, stairs, water bodies, and poles are loaded from .osm files and projected into the map frame.
• OSM confusion matrix fusion: A learned confusion matrix maps semantic class predictions to OSM categories, biasing BKI inference toward OSM-consistent classes. The matrix can be optimized from ground-truth co-occurrence data.
• Point-level and voxel-level prior application: OSM priors can be applied directly to per-point softmax distributions (before BKI insertion) or at the voxel level (during prediction).
• Height-based filtering: OSM priors are modulated by a height confusion matrix, so ground-level priors (roads, sidewalks) don't influence elevated structures and vice versa.
• Configurable decay: Prior influence drops off smoothly with distance from OSM geometry boundaries.

Multi-Dataset Support

• KITTI-360: Velodyne LiDAR with 4x4 pose matrices, Mercator-projected OSM alignment.
• MCD (Multi-Campus Dataset): Body-frame poses with quaternion orientation and body-to-lidar calibration.
• CU North Campus: Custom campus dataset with aligned LiDAR and inferred labels.
• Common taxonomy: All datasets are mapped to a shared 13-class semantic label space for consistent evaluation.

Multiclass Confidence Scores

• Ingests per-point softmax distributions (e.g., from CENet) rather than hard labels, preserving prediction uncertainty through the BKI pipeline.
• Supports uncertainty-based filtering: discard or down-weight highly uncertain observations using per-class precision thresholds or top-N% discounting.

ROS2 and Visualization

• Full ROS2 (Humble+) port of the original ROS1 codebase.
• Real-time RViz visualization of semantic maps, OSM geometry overlays, variance maps, and semantic uncertainty maps.
• Published PointCloud2 scans with label-based RGB coloring in the map frame.
• OSM geometry visualization as MarkerArray (buildings, roads, vegetation, etc.).

Evaluation Pipeline

• Per-scan query against ground-truth labels with IoU and accuracy metrics.
• Results written under the sequence directory for organized multi-run comparisons.

Dependencies

• ROS2 (Humble or later)
• PCL (Point Cloud Library)
• Eigen3
• OpenCV
• yaml-cpp
• libosmium (sudo apt install libosmium2-dev)

Building

mkdir -p ros2_ws/src #(wherever you choose)
cd ros2_ws/src
git clone 
cd ../
colcon build --packages-select osm_bki
source install/setup.bash

Usage

KITTI-360

ros2 launch osm_bki kitti360_launch.py

MCD (Multi-Campus Dataset)

ros2 launch osm_bki mcd_with_osm_launch.py

CU North Campus

ros2 launch osm_bki cu_north_campus_launch.py

Configuration

Dataset-specific parameters are in config/methods/:

• kitti360.yaml — KITTI-360 sequence, scan count, resolution, OSM file, prior strength
• mcd.yaml — MCD dataset configuration
• cu_north_campus.yaml — North Campus configuration

OSM confusion matrices are in config/datasets/ and can be re-optimized:

python3 optimize_osm_cm_scaled_kitti360.py --output config/datasets/osm_confusion_matrix_optimized_kitti360.yaml

Acknowledgments

This work builds directly on Lu Gan's BKISemanticMapping. The core octree structure, Bayesian kernel inference engine, and semantic octree node representation originate from that project. We gratefully acknowledge the original authors for making their code publicly available.