Enhancing The Accuracy of Small Object Detection In Traffic Safety Attributes Using Yolov11 And Esrgan

Peningkatan Akurasi Deteksi Objek Kecil pada Atribut Keselamatan Berkendara Menggunakan Yolov11 dan ESRGAN

Abstract

This study aims to detect motorcycle rider attributes, specifically helmets and side mirrors, using a deep learning approach combining YOLOv11 and ESRGAN models. The proposed model addresses challenges in attribute detection under real-world conditions, such as low-resolution images, varying angles, and uneven lighting. The dataset comprises images of motorcycle riders captured by surveillance cameras (CCTV), which underwent preprocessing, augmentation, and resolution enhancement using ESRGAN to improve input quality.

The results demonstrate that ESRGAN significantly enhances the performance of YOLOv11, particularly for high-resolution images. The YOLOv11 + ESRGAN model with 300 epochs achieved the best performance, with precision of 75.8%, recall of 69.1%, and an F1-score of 0.7 during testing. During validation, the model reached a precision of 0.826 and recall of 0.797, indicating good generalization capabilities. Compared to the YOLOv11 model without ESRGAN, this combination significantly improved accuracy, especially in detecting small attributes such as side mirrors.

This study suggests further exploration with larger and more diverse datasets and fine-tuning to enhance detection accuracy. Additionally, integrating the model into real-world systems based on edge computing can accelerate real-time inference and reduce reliance on cloud-based servers. With broader implementation, this model has the potential to improve the efficiency and safety of AI-powered traffic monitoring systems.