BEVMAPMATCH: Multimodal BEV Neural Map Matching for Robust Re-Localization of Autonomous Vehicles

Shounak Sural; Ragunathan Rajkumar

cs.CVMar 26, 2026

BEVMAPMATCH: Multimodal BEV Neural Map Matching for Robust Re-Localization of Autonomous Vehicles

Authors: Shounak Sural, Ragunathan Rajkumar

Organizations: Carnegie Mellon University, USA

Abstract

Localization in GNSS-denied and GNSS-degraded environments is a challenge for the safe widespread deployment of autonomous vehicles. Such GNSS-challenged environments require alternative methods for robust localization. In this work, we propose BEVMapMatch, a framework for robust vehicle re-localization on a known map without the need for GNSS priors. BEVMapMatch uses a context-aware lidar+camera fusion method to generate multimodal Bird's Eye View (BEV) segmentations around the ego vehicle in both good and adverse weather conditions. Leveraging a search mechanism based on cross-attention, the generated BEV segmentation maps are then used for the retrieval of candidate map patches for map-matching purposes. Finally, BEVMapMatch uses the top retrieved candidate for finer alignment against the generated BEV segmentation, achieving accurate global localization without the need for GNSS. Multiple frames of generated BEV segmentation further improve localization accuracy. Extensive evaluations show that BEVMapMatch outperforms existing methods for re-localization in GNSS-denied and adverse environments, with a Recall@1m of 39.8%, being nearly twice as much as the best performing re-localization baseline. Our code and data will be made available at https://github.com/ssuralcmu/BEVMapMatch.git.

Figures & tables

Fig. 1: Sub-meter absolute re-localization with BEVMapMatch . Here, BEV segmentation maps are generated, followed by coarse matching that finds the most likely region, marked with a green square. This likely region is used to identify point-to-point correspondences against the segmented map. Finally, fine-grained pixel-level patch alignment yields accurate 3-DOF pose estimation (shown in pink).

Fig. 2: Overview of the BEVMapMatch Architecture

Model	Segmentation	Seg. IOU	Top 1 $\times$ 1	Top 3 $\times$ 3
BEVMapMatch	BEVFusion [ 18 ]	62.7	38.6	81.1
BEVMapMatch	CAF (Ours)	75.0	46.7	86.4
BEVMapMatch	Ground Truth	100.0	50.9	89.9

TABLE I: Coarse localization accuracy comparison on the NuScenes-based dataset (1 frame).

Number of frames	Top 1 $\times$ 1 Acc. (%)	Top 3 $\times$ 3 Acc. (%)
1	46.68	86.37
2	50.76	88.77
4	53.19	88.90
8	53.46	88.65

TABLE II: Effect of temporal fusion on coarse localization performance of BEVMapMatch .

Backbone Used	Top 1 $\times$ 1 Acc. (%)	Top 3 $\times$ 3 Acc. (%)
ResNet-34 [ 12 ]	17.68	63.18
DINO-v2-Base [ 20 ]	41.15	86.12
DINO-v2-Large [ 20 ]	46.68	86.37

TABLE III: BEVMapMatch backbone ablation with fixed basemap input resolution of

224\times 224

Prior localization methods
Method	Venue	Absolute Recall (%)		Scaled Recall (%)
Method	Venue	@1m $\uparrow$	@2m $\uparrow$	Eq.@1m $\uparrow$	Eq.@2m $\uparrow$	Eq.@5m $\uparrow$	Eq.@10m $\uparrow$
OrienterNet [ 23 ]	CVPR 2023	5.83	18.92	5.83	18.92	52.83	66.21
U-BEV [ 3 ]	IROS 2024	16.89	41.60	16.89	41.60	71.33	83.46
MapLocNet [ 28 ]	IROS 2024	20.10	45.54	20.10	45.54	77.70	91.89
Ours

TABLE IV: Localization performance analysis with absolute recall for precise lane-level localization on NuScenes (val). Scaled recall metrics are used for a fair comparison against existing methods.

Fig. 3: Comparison of BEVMapMatch performance across choice of segmentation baselines and the ground truth.

Fig. 4: Robustness of BEVMapMatch with random GNSS perturbation. BEVMapMatch -CAF stays stable over varying GNSS errors since it is retrieval-based and does not use GNSS priors.

Fig. 5: Examples of localization at day and night-time with BEVMapMatch . Camera and lidar views shown (left) along with its localization output on a NuScenes base map (right). Within pink circles, red squares mark the predictions, blues mark ground truth and greens mark full overlap.

GNSS-Denied Localization

Global Navigation Satellite System (GNSS)-denied localization estimates position when satellite signals are unavailable or unreliable.

See all 58 papers

Robust Localization for Autonomous Vehicles in Highway Scenes
Apr 23, 2026Daqian Cheng, Xuchu Ding, Yujia Wu +2Autonomous Driving Datasets Autonomous Driving
Bot Auto · Archer Aviation · Prexa AI
L2G-Map: Local-to-Global Mapping via Hierarchical Diffusion Refinement and Elliptical Bayesian Fusion
Mar 4, 2025Siyu Li, Xinying Hong, Fei Teng +5HD Map Construction Autonomous Driving
School of Automation and Electrical Engineering, Zhejiang University of Science and Technology, China · School of Artificial Intelligence and Robotics and the National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, China · College of Computer Science and Electronic Engineering, Hunan University, China +1
GLIO2: A GPU-Parallelized Tightly-Coupled LiDAR-Inertial-GNSS System for Robust and Real-Time Global Localization and Mapping
Oct 8, 2026Qi Zhang, Xikun Liu, Qijun Qin +5Simultaneous Localization and Mapping LiDAR Localization
Department of Aeronautical and Aviation Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China

Apr 23, 2026J/K move · Enter open · S save

Robust Localization for Autonomous Vehicles in Highway Scenes

Daqian Cheng, Xuchu Ding, Yujia Wu, Xiang Zhang+1

Bot Auto · Archer Aviation · Prexa AI

Autonomous Driving Datasets Autonomous Driving LiDAR Localization Map-Based Localization

Open paper AskPDF arXiv

Localization for autonomous vehicles on highways remains under-explored compared to urban roads, and state-of-the-art methods for urban scenes degrade when directly applied to highways. We identify key challenges including environment changes under information homogeneity, heavy occlusion, degraded GNSS signals, and stringent downstream requirements on accuracy and latency. We propose a robust localization system to address highway challenges, which uses a dual-likelihood LiDAR front end that decouples 3D geometric structures and 2D road-texture cues to handle environment changes; a Control-EKF further leverages steering and acceleration commands to reduce lag and improve closed-loop behavior. An automated offline mapping and ground-truth pipeline keep maps fresh at high cadence for optimal localization performance. To catalyze progress, we release a public dataset covering both urban roads and highways while focusing on representative challenging highway clips, totaling 163 km; benchmarking is standardized using product-oriented accuracy metrics and certified ground truth. Compared to Apollo and Autoware, our system performs similarly on urban roads but shows superior robustness on challenging highway scenarios. The system has been validated by more than one million kilometers of road testing.

Abstract

GNSS-Denied Localization

Explore similar work

Robust Localization for Autonomous Vehicles in Highway Scenes

L2G-Map: Local-to-Global Mapping via Hierarchical Diffusion Refinement and Elliptical Bayesian Fusion

GLIO2: A GPU-Parallelized Tightly-Coupled LiDAR-Inertial-GNSS System for Robust and Real-Time Global Localization and Mapping