مشخصات پژوهش

Variable lighting conditions in Iranian Rahvar CCTV footage significantly degrade seat belt detection performance. This paper proposes a three-stage YOLOv8 pipeline with conditional gamma correction to enhance robustness. The system sequentially executes: (1) windshield localization (AP@0.5 = 0.994), (2) occupant classification (AP@0.5 = 0.277), and (3) seat belt verification (AP@0.5 = 0.988). A custom 3,619-image dataset from the Ashrafi Esfahani highway captures real-world Iranian scenarios, including low-light tunnels and high-glare urban environments. The core innovation is conditional gamma correction applied dynamically based on mean pixel intensity: γ = 3.0 for low light (<50), 1.8 for medium (50–200), and 0.7 for high (>200), integrated with CLAHE (clip_limit=3.0) and denoising (h=10). Mosaic augmentation (probability=0.7) is disabled in the last 10 epochs for convergence stability. The YOLOv8s model achieves mAP@0.5 = 0.755 with 7.3 ms inference (136 FPS) on NVIDIA T4. Ablation studies reveal conditional gamma correction contributes +4.2% mAP, significantly improving lighting robustness. Compared to YOLOv5 and YOLOv7, our method yields 9.4% higher accuracy. This is the first system tailored for Iranian Rahvar CCTV, supporting automated enforcement.