ENHANCING TRAFFIC SIGN RECOGNITION ROBUSTNESS IN AUTONOMOUS VEHICLES VIA ADVERSARIAL WEATHER DOMAIN ADAPTATION

Abstract

Autonomous vehicles have demonstrated impressive capabilities in safely navigating from one point to another, but impressive performance has been hampered by their inability to perform in adverse weather conditions such as fog, rain and snow using state-of-the-art deep learning models. Such changes in domain recognition brought by weather changes is one of the challenging problems in the computer vision world especially in the problem of traffic sign recognition. The domain changes in traffic sign recognition due to weather conditions is a challenging problem in computer vision. The proposed method is based on a single unified adversarial learning framework, which learns feature representations across both clear and adverse weather domains, while preserving fine-grained texture information through the FPFENet module, enhancing multiscale edges with MEEM, adaptively enhancing multiscale bidirectional features with the ADBF-FPN module, and selectively suppressing noise with MCGM. Through extensive experiments on the German Traffic Sign Recognition Benchmark (GTSRB) dataset with the addition of synthetic weather conditions generated by the CycleGAN, significant performance gains are observed, with 96.4% mAP@0.5 (3.3% better than the baseline YOLOv8), and superior performance for fog (94.1% mAP), rain (92.8% mAP), and snow (87.6% mAP). The proposed model retains the ability to make real-time inferences at 28.4ms per frame, which is also a viable rate for autonomous driving applications. The synergistic contribution of each architectural component is confirmed by ablation studies, which contribute the most in severe weather degradation in the case of the MCGM. Complete solution for all weather traffic sign recognition. The development of the proposed approach in this research is a crucial step of meeting the safety, reliability of autonomous driving system in all environmental conditions.