Generative Method-Based Traffic Sign Detection and Recognition in Occluded Conditions

Abstract

Traffic Sign Detection and Recognition (TSDR) is a crucial part of Autonomous Vehicles (AVs) for safe and efficient navigation. Recent advancements in Deep Learning (DL) and 5G technology have significantly improved the perception of AV by enabling real-time data processing. However, real-world challenges such as occlusions from vehicles, roadside objects, and unfavorable traffic conditions affect TSDR performance. Detection models like Faster Region-based Convolutional Neural Network (Faster R-CNN) enhanced TSDR accuracy, but the occlusions remain a significant challenge. This paper proposes a novel feature-based Generative Adversarial Network (GAN) to regenerate occluded traffic signs to improve detection accuracy. The GAN-generated features are fused with the Faster R-CNN-generated features. The architecture of GAN is tiny and integrated at the feature level within Faster R-CNN to minimize computational overhead. Thus, it leverages real-time processing and effective model deployment for AV. An artificial occlusion insertion algorithm has been proposed to train and evaluate the proposed approach against the occluded traffic sign. The experimental results demonstrate that the proposed approach improves the accuracy by 4.2% on level one occlusion, 12% on level two occlusion, and 6.5% on the augmented (occluded and clean) data. Therefore, the proposed approach is more adaptable for 5G-powered intelligent transportation systems. © 2025 Elsevier B.V., All rights reserved.