Authors - Waqas Ahmed, Muhammad Khalid, Adil Khan, Gulraiz Khan Abstract - Robust traffic sign recognition being important component in integrated Autonomous Driving Assistance Systems (ADAS), driving safety systems; suffers from many challenges including Occlusions and obstructions in road signs. Occlusions may occur due to natural or man made objects; weather conditions, lightening effects, vegetation, paint sprays, parked vehicles may result in traffic sign misclassifications adding more to endangerment of human life and infrastructure. This research contributed by addressing the occlusion problem by devising an occlusion-aware masked autoencoder Vit-OCC for reconstruction of obstructed traffic sign while utilizing German Traffic Sign Recognition Benchmark (GTSRB). Study encompassed four Vit models; a standard classifier (ViT-CLS), a region-aware Cutout-augmented model (ViT-Cutout), a model initialized via random-mask Masked Auto encoding (ViT-MAE), and a proposed occlusion-aware Masked Autoencoder (ViT-OCC). All the models are evaluated using variable and increasing occlusion levels, comprising of 9 different levels from 0% being lowest occlusion level to highly obstructed 80% threshold. Results indicate different patterns of degradation among the four models, although ViT-MAE achieved highest baseline accuracy of 87.23%, ViT-Cutout exhibited stable performance with the introduction of high occlusion levels. ViT-Cutout ViT-OCC exhibited sharp drop in performance under high occlusion thresholds. This study contributed in a systematic exploration of Vision Transformer robustness across variants under controlled and increasing occlusion thresholds while concluding unique patterns in degradation and stability.
