<p>In recent years, traffic participant detection (TPD) has become an essential but challenging task in domains, such as urban management and autonomous driving. While deep learning-based object detection algorithms have shown promising performance in this area, existing algorithms face challenges with meeting real-time requirements, excessive parameter counts, and the accurate detection of multiple traffic participants. To address these challenges, we propose a novel traffic participant detector (TPD–YOLO). Initially, we integrate a detailed feature enhancement module (DFEM) into the neck structure, which enhances the ability to extract spatially detailed feature information. Subsequently, the enhanced features generated by DFEM are fed into a high-resolution small object detection head (SHead) to enrich the feature representation. Finally, the channel-wise histogram feature distillation (CHFD) compression strategy is employed to enable the student model to approximate the performance of the teacher model while requiring fewer parameters and computational resources. We validated the proposed method on the SEU_PML data set and further evaluated the model’s generalization performance on the VisDrone2019 data set. Experimental results demonstrate that the proposed TPD–YOLO achieves 73.7% mAP on the SEU_PML data set with only 3.2M parameters. Furthermore, the model achieves inference speeds of 144.9 FPS, maintaining a balance between accuracy and efficiency. This research provides an advanced small object detection solution for real-time requirements in traffic scenarios (≥ 30 FPS), offering a more effective approach to urban traffic management.</p>