Improved Daytime Traffic Signal Detection Method Based on YOLOv8

Addressing current issues such as low accuracy in traffic signal detection algorithms, weak detection capability for small targets, and susceptibility to lighting and vehicle speed variations, we propose an improved traffic signal detection method based on YOLOv8. Initially, an EMA attention mechanism module is incorporated into the backbone network to enhance feature extraction capabilities. Subsequently, the entire Neck module is replaced with ASF's Neck module to improve detection performance for small-scale targets. Further enhancements to the Neck include introducing the P2 small target detection layer to prevent overflow of parameters and computation, thus reducing computational costs. Finally, the original detection head is replaced with SEAMHead, which integrates the SEAM attention mechanism module to compensate for various shortcomings in the original detection head during prediction tasks. Validation on a custom dataset shows that compared to the original YOLOv8, our algorithm achieves a 3.5% increase in accuracy, a 10.3% increase in recall, and an average precision mAP of 93.4%, representing a 5.5% improvement. This method effectively achieves rapid and accurate detection of traffic signals and holds broad applicability for research related to traffic signal analysis on roads.