Helmet detection often faces challenges in complex road scenarios such as heavy traffic, pedestrian interference, and severe occlusion of targets. These conditions can easily lead to low detection accuracy, false detections, and missed detections. This paper proposes a high-performance helmet recognition model based on the CPM-YOLO algorithm. First, a novel cross-scale feature fusion method, CS-FPN, is proposed to better integrate high-level semantic and low-level geometric feature information. Next, the PCT module is introduced to optimize feature extraction capabilities of the model. Additionally, a bounding box regression loss function based on the minimum point distance is adopted to enhance the model′s convergence speed and accuracy. Furthermore, the 20×20 downsampling layer and 20×20 detection head in the backbone network are removed, and a new 160×160 small-object detection head is introduced. Finally, ablation studies validate the effectiveness of each improved module in enhancing the model′s performance, and comparative experiments demonstrate the superiority and generalizability of the CPM-YOLO model.Experimental results show that compared to the baseline model, the proposed method achieves improvements of 5.5% in mAP@0.5. Additionally, the number of parameters and model size are reduced by 69.9% and 67.2%, respectively. The new model significantly reduces complexity while enhancing helmet detection capabilities in road environments.