FSD-DETR: A Target Detection Method for Complex Traffic Scenarios

Small-scale objects and partially occluded targets in complex traffic scenarios impose significant challenges on the feature representation capability of target detection models, while the increasing computational burden limits detection efficiency. To address these issues, this study proposes an improved detector, termed FSD-DETR, based on RT-DETR-R18. First, a Frequency-Aware Dynamic Convolution (FDConv) module is introduced into the backbone network to enhance the fine-grained representation capability of small-scale objects through frequency-aware feature extraction. Then, deformable attention is incorporated into the AIFI encoder stage to construct AIFI-DA, which establishes adaptive spatial sampling according to object content and effectively captures local representations with relevant contextual information, thereby improving the feature representation of partially occluded targets. Furthermore, Slim-Neck is adopted for lightweight multi-scale feature fusion, and DySample is integrated into the fusion pathway to provide adaptive upsampling reconstruction capability for aggregating high-level semantic information and low-level spatial details while reducing the computational overhead of feature transformation and fusion. The proposed method is independently trained and evaluated on the KITTI, VisDrone, and BDD100K datasets to validate its detection performance in occlusion scenarios, small-object scenarios, and diverse traffic environments. Experimental results demonstrate that, compared with RT-DETR-R18, FSD-DETR achieves higher detection accuracy while reducing model parameters and computational cost. The proposed framework provides an efficient architectural solution for target detection in complex traffic scenarios.