Enhancing Safety-Integrity of Localization Systems in Lane-Keeping Scenarios Using a Self-Supervised Spatial Temporal Deep Network
Journal
IET Intelligent Transport Systems
Journal Volume
20
Journal Issue
1
Start Page
e70249
ISSN
1751956X
Date Issued
2026-01
Author(s)
Abstract
Accurate and reliable real-time localization for lane-keeping underpins safe navigation in autonomous vehicles. However, these systems are frequently exposed to sensor faults induced by environmental disturbances, which can result in critical integrity failures. To address this, we present a comprehensive fault detection and exclusion (FDE) framework that proactively identifies and removes faulty signals before sensor fusion, thereby preserving pose-estimation accuracy in lane-keeping scenarios. The framework incorporates a straightforward yet effective spatial-temporal localization diagnosis (STLD) model—combining lightweight CNNs for localized short-term feature extraction with GRUs for long-term dependency modelling—to deliver robust fault detection with minimal computational overhead. Trained self-supervised via systematic error injection and alert limit-based labelling, STLD requires no manual fault annotation. On the KITTI benchmark with injected multipath, drift, loss and LiDAR errors, our FDE framework effectively reduces localization failures compared with SPRT and lowers the average mean pose error across all error injection scenarios from 12.56 m without FDE to 1.20 m. These results validate the effectiveness of the proposed FDE framework in enhancing localization robustness and precision under varied fault scenarios.
Subjects
Automated Vehicle
Fault Detection and Execution
Localization Integrity
Spatial-Temporal Localization Diagnosis
Publisher
John Wiley and Sons Inc
Type
journal article