This study enhances bicycle helmet detection by optimizing the YOLOv9 model with DualConv and CAFM attention mechanisms. These modifications improve detection accuracy while reducing resource consumption, making the model suitable for low-power platforms. A new dataset with diverse real-world scenarios was also developed to address the scarcity of bicycle helmet datasets. Experimental results demonstrate that the improved YOLOv9 surpasses the accuracy and efficiency of existing mainstream detection models. This research provides a robust and practical solution for intelligent traffic monitoring and supports future policy development.