To address the problems of missed detections and false detections in electric bike helmet detection under real-world scenarios caused by complex background interference and multi-scale target variability, this study proposes an improved YOLOv10n-based detection method. First, the KernelWarehouse convolution module, based on dynamic convolution, replaces the original convolution, thus enhancing the model’s ability to recognize targets under background interference. Second, a bidirectional feature pyramid fusion network based on convolutional additive self-attention is designed. By introducing a small object detection head and a bidirectional feature fusion mechanism, the deep information interaction between the small-object layer and the conventional object layer is enhanced. Furthermore, the C2f module in the feature fusion network is deeply integrated with the convolutional additive self-attention module, effectively alleviating the problem of feature information loss and significantly improving small object detection performance. Finally, a lightweight multi-scale shared convolutional detection head is designed, which enhances multi-scale feature perception while significantly reducing the number of parameters and computational complexity, meeting the requirements of real-time detection. Experimental results show that, compared with the baseline model, the improved model increases mAP@0.5 by 3.1% while reducing the number of parameters by 22% and the model size by 26%, verifying the superiority of the proposed method for electric bike helmet detection tasks in complex real-world scenarios.