Improving YOLOv11 for Traffic Sign Detection and Classification at Night

Abstract

Automatic traffic sign detection is an important challenge in supporting intelligent transportation systems, especially under nighttime conditions with low lighting. The YOLOv11-based model demonstrated good performance, but it was still limited in extracting adaptive features from small and distorted objects, which affected the accuracy. This study proposes a modification to YOLOv11 by adding a Deformable Convolutional Network (DCN) module to the backbone before the C3k2 block to enhance adaptive spatial feature extraction capabilities. The model was tested on a nighttime traffic sign dataset and evaluated using precision, recall, mAP0.5, and mAP0.5 up to 0.95. The experimental results showed improved performance compared to the YOLOv11 baseline, with increases of 12.5 percent, 3.3 percent, 6.4 percent, and 4.2 percent in precision, recall, mAP0.5, and mAP0.5 up to 0.95, respectively. Although there was a slight decrease in the inference speed, the model maintained real-time performance. These findings demonstrate that integrating a DCN into YOLOv11n is an effective approach for improving traffic sign detection accuracy in real-world conditions without sacrificing efficiency, thereby potentially supporting the implementation of more reliable intelligent transportation systems.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint