EH-DETR: enhanced two-wheeler helmet detection transformer for small and complex scenes
Liwei Liu, Xinbo Yue, Ming Lu, Pingge He · Journal of Electronic Imaging · 2025
Helmet detection remains challenging due to small target sizes, complex backgrounds, and confusion with neighboring objects. An enhanced two-wheeler helmet detection model based on real time detection transformer (RT-DETR) is proposed. Enhanced helmet detection transformer (EH-DETR) incorporates a Faster RepConv Block structure designed using model re-parameterization techniques to improve detection performance while meeting real-time requirements. In addition, it introduces a mixed local channel attention module to resolve object confusion and a cross-stage partial parallel dilated convolution module to enhance feature fusion efficiency and receptive field size. To tackle the detection of small helmet objects, EH-DETR employs a channel-gated up-sampling and down-sampling technique. Experimental results demonstrate that EH-DETR enhances the mAP50 value by 2.3% and increases the frames per second to 141.3 on the helmet dataset, significantly improving the model’s capability for detecting small helmets and dense scenes while ensuring real-time performance. EH-DETR provides an effective solution for real-time two-wheeler helmet detection.