Research on traffic sign detection model with improved YOLOv8s

Abstract With the rapid advancement of intelligent driving technologies such as autonomous driving and driver assistance, the significance of traffic sign recognition in road and traffic safety has become increasingly prominent. Addressing the challenges posed by the current mainstream traffic sign detection models, which include complex background structures and the detection of small targets, this study proposes an enhanced network for detecting traffic signs based on YOLOv8s. Firstly, we modify the convolutional component of the backbone by replacing the initial convolutions with the Alterable Kernel Convolution (AKConv) network, thereby enhancing feature extraction efficiency and network performance. Secondly, we introduce the Global Attention Mechanism (GAM) within the network to effectively aggregate features, improve reconstruction performance, and maintain low computational and storage costs. Lastly, we integrate the Bi-directional Feature Pyramid Network (BiFPN) into the YOLOv8s framework, enabling the model to capture multi-scale features efficiently and enhance recognition accuracy. Experimental results conducted on the CCTSDB-2021 show the efficacy of the study improved YOLOv8s model. Precision, recall, and mAP@50% achieve impressive values of 96.0%, 79.4%, and 88.3%, respectively. These metrics represent a 4.6% improvement in precision, a 0.4% enhancement in the recall, and a 2% enhancement in mAP@50%, compared with the original YOLOv8s model. Moreover, this improved model exhibits greater precision when compared with other existing models.