EdgeLoc: A Communication-Adaptive Parallel System for Real-Time Localization in Infrastructure-Assisted Autonomous Driving

TL;DR

EdgeLoc addresses the performance gap between traditional real-time visual odometry (VO) and end-to-end DNN localization by deploying DNN inference on roadside units (RSUs) and fusing results with on-vehicle VO in a ROS-based parallel architecture. It introduces an uncertainty-aware pose fusion that replaces Kalman filtering, and an online-learning strategy to auto-split DNN inference between vehicle and RSU to optimize latency without sacrificing accuracy. The approach demonstrates that the DNN-augmented, edge-assisted localization yields substantial accuracy gains over VO or Kalman-filter fusion, including reductions of 67.75% relative to VO, 30.26% relative to Kalman, and 29.95% relative to non-real-time collaborative inference in representative experiments. The work validates latency-aware fusion, online adaptation, and edge collaboration in real cellular networks and provides an open-source implementation for practical deployment.

Abstract

This paper presents EdgeLoc, an infrastructure-assisted, real-time localization system for autonomous driving that addresses the incompatibility between traditional localization methods and deep learning approaches. The system is built on top of the Robot Operating System (ROS) and combines the real-time performance of traditional methods with the high accuracy of deep learning approaches. The system leverages edge computing capabilities of roadside units (RSUs) for precise localization to enhance on-vehicle localization that is based on the real-time visual odometry. EdgeLoc is a parallel processing system, utilizing a proposed uncertainty-aware pose fusion solution. It achieves communication adaptivity through online learning and addresses fluctuations via window-based detection. Moreover, it achieves optimal latency and maximum improvement by utilizing auto-splitting vehicle-infrastructure collaborative inference, as well as online distribution learning for decision-making. Even with the most basic end-to-end deep neural network for localization estimation, EdgeLoc realizes a 67.75\% reduction in the localization error for real-time local visual odometry, a 29.95\% reduction for non-real-time collaborative inference, and a 30.26\% reduction compared to Kalman filtering. Finally, accuracy-to-latency conversion was experimentally validated, and an overall experiment was conducted on a practical cellular network. The system is open sourced at https://github.com/LoganCome/EdgeAssistedLocalization.

Figures (10)

• Figure 1: Challenges and system overview of deploying end-to-end localization deep neural networks on roadside units for infrastructure-assisted autonomous driving.
• Figure 2: System architecture of EdgeLoc: the core components, communication setup, and collaborative operations between the autonomous vehicle and roadside unit.
• Figure 3: The diagram includes five modules (end-to-end localization modules on the vehicle and RSU, sensor loop module, localization revision module, and action server), vertical lines representing the lifecycle of each module, arrows showing interactions and invocations between modules, and the parallel execution of real-time localization and RSU collaboration threads.
• Figure 4: Comparison of localization errors for different methods over time. The methods Visual Odometry, DNN-based E2E Localizationkendall2015posenet, and EdgeLoc are respectively shown from the second to fourth images. The red curves represent the ground truth track. The left first sub-figure is an example of input image, and the right first sub-figure is the map and the real trace.
• Figure 5: Localization error comparison among VO, DNN-based E2E localization, and EdgeLoc.