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

2025-12-17

AI Summary

To address high-precision real-time localization in GPS-denied off-road environments, this paper proposes a particle filter framework based on cross-modal feature matching. It first learns bird’s-eye-view (BEV) features from onboard RGB-D data, then matches these features against local descriptors extracted from retrieved aerial imagery to compute observation likelihoods for each particle pose hypothesis. Crucially, this work is the first to integrate learnable BEV representations with aerial image local matching into a particle filter, significantly improving robustness under challenging conditions—such as dense canopy cover and strong shadows—and enhancing generalization to both seen and unseen trajectories. Evaluated on two real-world off-road datasets, the method reduces absolute trajectory error (ATE) by 7.5× (seen) and 7.0× (unseen) over retrieval-based baselines, while achieving real-time inference at 10 Hz on a Tesla T4 GPU.

Technology Category

• Intelligent Robots: State Estimation
• Computer Vision: Multi-modal Vision
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs

Abstract

We propose BEV-Patch-PF, a GPS-free sequential geo-localization system that integrates a particle filter with learned bird's-eye-view (BEV) and aerial feature maps. From onboard RGB and depth images, we construct a BEV feature map. For each 3-DoF particle pose hypothesis, we crop the corresponding patch from an aerial feature map computed from a local aerial image queried around the approximate location. BEV-Patch-PF computes a per-particle log-likelihood by matching the BEV feature to the aerial patch feature. On two real-world off-road datasets, our method achieves 7.5x lower absolute trajectory error (ATE) on seen routes and 7.0x lower ATE on unseen routes than a retrieval-based baseline, while maintaining accuracy under dense canopy and shadow. The system runs in real time at 10 Hz on an NVIDIA Tesla T4, enabling practical robot deployment.

Problem

• Develops GPS-free geo-localization using BEV and aerial feature matching
• Reduces trajectory error in off-road environments with particle filtering
• Enables real-time robot deployment under challenging conditions like canopy

Innovation

• Particle filter integrates BEV and aerial feature maps
• Matches BEV features to aerial patches for localization
• Achieves real-time GPS-free geo-localization in off-road environments