​Abstract: This article proposes a vehicle localization method that enhances MEMS sensor-based self-localization by incorporating roadside object detections obtained from on-board camera images. Although conventional MEMS-based approaches have achieved meter-level accuracy, their performance degrades on roads with limited surface variations, leading to reduced robustness. To overcome this limitation, roadside objects such as traffic lights, traffic signs, and poles are detected using a fine-tuned YOLOv8m model and utilized as visual landmarks. The spatial configurations of the detected bounding boxes are matched with a reference dataset associated with RTK-GNSS positions, enabling correction of accumulated localization errors. This approach reduces dependency on road surface features and improves localization stability in diverse urban environments. Experimental results using real-world driving data demonstrate that the proposed method achieves improved robustness and maintains localization errors within approximately 3.13 m (mean), 0.42 m (median), 12.6 m (RMSE), and 111.8 m (maximum) under challenging conditions.

Keywords: Vehicle localization, MEMS sensors. RTK-GNSS, Object detection, YOLOv8m, Bounding box.

__________________________________________________________________________________________