In complex traffic scenes, small vehicle targets are often affected by occlusion and background noise. Low detection accuracy is caused. Feature extraction becomes difficult. Existing models also have large parameter sizes. To address these problems, an improved PDG-YOLO algorithm based on YOLOv11n is proposed.First, a lightweight PSConv algorithm is used in the downsampling stage of the backbone. A larger receptive field is obtained. Feature extraction ability is improved. Model lightweight design is also maintained. Second, a DFMFPN network based on the DFM feature fusion method is designed. Multi-path structures are adopted. Shallow and deep feature maps are fused at multiple scales. The perception ability for small targets is enhanced. Finally, a CGLU module is introduced. Feature maps in the backbone are gated. Redundant information is filtered. Important information is preserved. Experiments are conducted on the UA-DETRAC dataset and the Street-View dataset. The number of parameters is reduced by 1.6% compared with the baseline model. Precision is increased by 4%. Recall is increased by 7.6%. The mAP@50 is improved by 3.7%. The FLOPs are reduced by 3.1%. On the Street-View dataset, the mAP@50 is increased by 1.9%. The mAP@50–95 is increased by 3.4%. In addition, this paper compares the detection performance of PDG-YOLO and the baseline model through visual results. The inference frame rate of PDG-YOLO is 26.1% higher than that of the baseline model. This shows that the proposed algorithm has better real-time performance while improving detection accuracy. Therefore, the proposed method has practical application value. It can provide effective support for intelligent traffic monitoring, road vehicle detection, assisted driving perception, and edge-end traffic object recognition.