BEV-Patch-PF

Particle Filtering with BEV-Aerial Feature Matching for Off-Road Geo-Localization

Project Page • Paper (arXiv) • ROS 2 Deployment

BEV-Patch-PF is a GPS-free geo-localization system for long-horizon offroad localization. It combines learned bird's-eye-view (BEV) features, aerial map features, and particle filtering to estimate robot pose over time.

Installation

conda create -y -n bev-patch-pf python=3.12
conda activate bev-patch-pf
conda install -y -c conda-forge manifpy
pip install -e .

Install development dependencies with:

pip install -e .[dev]

For multi-GPU training with accelerate:

pip install accelerate

Quick Start

Demo

The demo downloads a pinned checkpoint and a pinned sparse demo dataset from Hugging Face, then launches Rerun with the demo blueprint.

python src/run_demo.py

The demo configuration is defined in config/demo.yaml.

Particle Filter Evaluation

Run offline particle filtering with a local checkpoint:

python src/run_pf.py sequence=tartandrive ckpt_path=/path/to/model.pth

run_pf.py uses Hydra overrides from config/run_pf.yaml together with sequence configs under config/sequence.

Training

Single-GPU training:

python src/train.py

Multi-GPU training:

accelerate launch --multi_gpu --num_processes=<num_gpus> --mixed_precision=fp16 src/train_ddp.py

The default training setup is defined in config/train.yaml. Training outputs are written under output/train/....

export_onnx.py exports the model components, runs onnx.checker.check_model, and performs parity checks against PyTorch.

Real-Time Deployment

After exporting ONNX, build TensorRT engines and run the ROS 2 deployment stack from the companion repository: ut-amrl/bev-patch-pf_ROS2

Export ONNX

Export the model for deployment:

python scripts/export_onnx.py --ckpt_path=/path/to/model.pth --out_dir=/path/to/export_dir

Dataset Preparation

From ROS Bagfiles

A typical pipeline is:

1. Extract sensor streams such as RGB, depth, and IMU using rosbagkit.
2. Generate a trajectory estimate with a SLAM or odometry pipeline (e.g., FAST-LIO, Adaptive-LIO).
3. Export a GeoTIFF map in the correct UTM zone (e.g., QGIS).
4. Align the trajectory into the GeoTIFF frame:

python preprocessing/align_trajectory_geotiff.py --geotiff=/path/to/map.tiff --traj=/path/to/trajectory.csv

5. Run dataset-specific preprocessing as needed, for example:

python preprocessing/preprocess_arl_jackal.py

From TartanDrive 2.0

1. Download the raw bagfiles from https://github.com/castacks/tartan_drive_2.0.
2. Extract images and odometry with rosbagkit.
3. Run preprocessing:

python preprocessing/preprocess_tartandrive.py

Citation

@misc{lee2025bevpatchpf,
  title         = {BEV-Patch-PF: Particle Filtering with BEV-Aerial Feature Matching for Off-Road Geo-Localization},
  author        = {Lee, Dongmyeong and Quattrociocchi, Jesse and Ellis, Christian and Rana, Rwik and Adkins, Amanda and Uccello, Adam and Warnell, Garrett and Biswas, Joydeep},
  year          = {2025},
  eprint        = {2512.15111},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}