Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VVLoc: Prior-free 3-DoF Vehicle Visual Localization

Ze Huang, Zhongyang Xiao, Mingliang Song, Longan Yang, Hongyuan Yuan, Li Sun

TL;DR

VVLoc tackles the gap in autonomous driving localization by unifying topological and metric localization within a single, multi-camera visual framework. It learns global and local-view BEV descriptors without explicit scale priors, leveraging BEV Padding and a matching-based pose estimator to recover $3$-DoF poses while providing a confidence score. The approach achieves state-of-the-art accuracy on public datasets such as NCLT and strong results on a self-collected Park-mapping dataset, including robust loop-closure detection and multi-modal mapping, with demonstrated robustness under degraded sensing. This work delivers a scalable, confidence-aware localization solution suitable for industrial deployment in autonomous driving pipelines.

Abstract

Localization is a critical technology in autonomous driving, encompassing both topological localization, which identifies the most similar map keyframe to the current observation, and metric localization, which provides precise spatial coordinates. Conventional methods typically address these tasks independently, rely on single-camera setups, and often require additional 3D semantic or pose priors, while lacking mechanisms to quantify the confidence of localization results, making them less feasible for real industrial applications. In this paper, we propose VVLoc, a unified pipeline that employs a single neural network to concurrently achieve topological and metric vehicle localization using multi-camera system. VVLoc first evaluates the geo-proximity between visual observations, then estimates their relative metric poses using a matching strategy, while also providing a confidence measure. Additionally, the training process for VVLoc is highly efficient, requiring only pairs of visual data and corresponding ground-truth poses, eliminating the need for complex supplementary data. We evaluate VVLoc not only on the publicly available datasets, but also on a more challenging self-collected dataset, demonstrating its ability to deliver state-of-the-art localization accuracy across a wide range of localization tasks.

VVLoc: Prior-free 3-DoF Vehicle Visual Localization

TL;DR

VVLoc tackles the gap in autonomous driving localization by unifying topological and metric localization within a single, multi-camera visual framework. It learns global and local-view BEV descriptors without explicit scale priors, leveraging BEV Padding and a matching-based pose estimator to recover -DoF poses while providing a confidence score. The approach achieves state-of-the-art accuracy on public datasets such as NCLT and strong results on a self-collected Park-mapping dataset, including robust loop-closure detection and multi-modal mapping, with demonstrated robustness under degraded sensing. This work delivers a scalable, confidence-aware localization solution suitable for industrial deployment in autonomous driving pipelines.

Abstract

Localization is a critical technology in autonomous driving, encompassing both topological localization, which identifies the most similar map keyframe to the current observation, and metric localization, which provides precise spatial coordinates. Conventional methods typically address these tasks independently, rely on single-camera setups, and often require additional 3D semantic or pose priors, while lacking mechanisms to quantify the confidence of localization results, making them less feasible for real industrial applications. In this paper, we propose VVLoc, a unified pipeline that employs a single neural network to concurrently achieve topological and metric vehicle localization using multi-camera system. VVLoc first evaluates the geo-proximity between visual observations, then estimates their relative metric poses using a matching strategy, while also providing a confidence measure. Additionally, the training process for VVLoc is highly efficient, requiring only pairs of visual data and corresponding ground-truth poses, eliminating the need for complex supplementary data. We evaluate VVLoc not only on the publicly available datasets, but also on a more challenging self-collected dataset, demonstrating its ability to deliver state-of-the-art localization accuracy across a wide range of localization tasks.
Paper Structure (14 sections, 8 equations, 9 figures, 7 tables, 1 algorithm)

This paper contains 14 sections, 8 equations, 9 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Visual Topological Localizer
  5. Visual Metric Localizer
  6. Loss Functions
  7. Experiments
  8. Datasets and Model Details
  9. Global Localization Evaluation
  10. Loop Closure Detection Evaluation
  11. Multi-modal Mapping Evaluation
  12. Localization Evaluation in Degraded Scenarios
  13. Additional Ablation Studies
  14. Conclusions

Figures (9)

  • Figure 1: Pipeline of the VVLoc: For the input (A) multi-camera images, we leveraging (B) BEVformerli2022bevformer for BEV queries extraction and (C) decode global descriptor from the BEV queries for topological localization. (D) For metric localization, we decode BEV queries into local-view descriptors and combine BEV Padding for metric pose estimation.
  • Figure 2: Matching Strategy Comparison. Conventional matching-based strategies produce ambiguous supervision in visual BEV, which hinders model training. VVLoc addresses this issue by introducing a method that leverages BEV Padding and local-view descriptor matching to estimate the metric relative pose accurately.
  • Figure 3: The Proposed Novel Components.(A) Radius-aware Self-attention and (B) Theta-aware Self-attention are implemented to enhance the geometric awareness of the local-view descriptors.
  • Figure 4: Qualitative Global Localization Results on the NCLT Dataset. VVLoc can retrieve the correct map frames with a wide range of perspectives and lighting changes, and give correct metric pose estimation.
  • Figure 5: Qualitative Global Localization Results on the Oxford Radar Robotcar Dataset.
  • ...and 4 more figures