Edge-Server Distributed Architecture for Efficient Monocular Visual Localization
Journal
2025 IEEE Workshop on Signal Processing Systems, SiPS 2025
Start Page
1
End Page
5
ISBN (of the container)
979-833159831-0
Date Issued
2025-11-27
Author(s)
Abstract
Accurate monocular 6-DoF localization (Degrees of Freedom) is crucial for autonomous driving, mobile robotics, and augmented reality. While visual odometry (VO) and SLAM methods achieve real-time performance, their susceptibility to drift limits their global accuracy. Conversely, image-based localization methods leveraging global information achieve high precision but incur substantial latency, hindering practical deployment. To optimally balance accuracy and runtime efficiency, we propose a distributed localization framework combining lightweight edgeside VO with robust server-side global feature-based localization. Our system employs an adaptive matcher selection strategy that dynamically switches between efficient nearest-neighbor matching and robust deep-learning-based methods guided by motion priors and match reliability. Extensive experiments on the Extended CMU-Seasons dataset demonstrate that our approach achieves a 9.3 times speedup in runtime compared to localizationonly solutions while closely matching their accuracy, presenting an effective and practical solution for real-world embeddedsystem applications.
Event(s)
2025 IEEE Workshop on Signal Processing Systems, SiPS 2025
Subjects
computer vision
Deep learning
distributed computing system
visual localization
Publisher
IEEE
Type
conference paper