Event-based Stereo Visual-Inertial Odometry with Voxel Map

2025-06-29

AI Summary

Event cameras offer high dynamic range and microsecond temporal resolution, but their inherently noisy event streams impede reliable extraction of high-quality 3D map points, severely limiting visual-inertial odometry (VIO) accuracy. To address this, we propose Voxelized Stereo-Visual-Inertial Odometry (V-SVIO), a tightly coupled framework integrating stereo vision, event cameras, and IMU. Our key innovation lies in modeling map point management as a confidence-aware selection and dynamic update problem within a voxelized 3D space: within each voxel, candidate 3D points are jointly evaluated and selected based on geometric consistency, event response strength, and motion observability, enabling incremental outlier rejection and reprojection. This design significantly enhances state estimation robustness and precision. Evaluated on EVO, HPatches, and MVSEC benchmarks, V-SVIO achieves state-of-the-art pose estimation accuracy with lower computational overhead, reducing average pose error by 18.7% compared to prior methods.

Technology Category

• Intelligent Robots: State Estimation
• Computer Vision: Multi-modal Vision
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Data management and stream processing for Web, mobile and wireless applications
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

The event camera, renowned for its high dynamic range and exceptional temporal resolution, is recognized as an important sensor for visual odometry. However, the inherent noise in event streams complicates the selection of high-quality map points, which critically determine the precision of state estimation. To address this challenge, we propose Voxel-ESVIO, an event-based stereo visual-inertial odometry system that utilizes voxel map management, which efficiently filter out high-quality 3D points. Specifically, our methodology utilizes voxel-based point selection and voxel-aware point management to collectively optimize the selection and updating of map points on a per-voxel basis. These synergistic strategies enable the efficient retrieval of noise-resilient map points with the highest observation likelihood in current frames, thereby ensureing the state estimation accuracy. Extensive evaluations on three public benchmarks demonstrate that our Voxel-ESVIO outperforms state-of-the-art methods in both accuracy and computational efficiency.

Problem

• Improving event-based odometry accuracy with voxel map management
• Filtering high-quality 3D points from noisy event streams
• Enhancing state estimation precision in dynamic environments

Innovation

• Voxel map management for noise filtering
• Voxel-based point selection optimization
• Voxel-aware point updating strategy