Multi-scale object detection in complex road scenarios remains an arduous challenge, primarily attributed to environmental interference and background heterogeneity. To surmount these obstacles, this paper introduces an enhanced real-time Transformer-based detection model, FECS-DETR. In the design of the backbone network, a feature enhancement architecture is constructed, integrating a lightweight FasterBlock architecture with an Exponential Moving Average (EMA) attention mechanism. The lightweight FasterBlock architecture enables efficient extraction of shallow features by optimizing inter-layer parameter sharing, while the EMA attention mechanism establishes long-range context dependencies through dynamic weight allocation. This synergy significantly enhances the model’s spatio-temporal perception in intricate scenarios. For the neck network, the Cross Stage Partial Stage (CSPStage) module is introduced. Leveraging dual mechanisms of multi-scale feature decoupling and noise suppression, CSPStage effectively mitigates redundant information while substantially improving the feature continuity of occluded objects. Experimental results demonstrate that on the SODA10M dataset, FECS-DETR series achieves a 2.3%–4.0% relative growth rate in mean average precision (mAP@50) compared to the baseline RT-DETR model. The lightweight version FECS-DETR-18 has an 18% relative growth rate in Giga Floating-Point Operations (GFLOPs) compared with RT-DETR-18, and its mAP@50 improves by 2.5%. This achievement exemplifies an optimal balance between performance and efficiency. Additionally, on the KITTI and BDD100K datasets, the model attains maximum relative growth rates of 0.9% and 2.2% in mAP@50, respectively validating its excellent generalization ability. Visual analysis vividly showcases the model’s outstanding detection performance in scenarios involving overlapping objects and low-light conditions. Through three types of interference experiments, namely Gaussian noise, motion blur, and light mutation, the robustness of FECS-DETR against complex environmental perturbations is further verified, confirming the model’s superiority in practical applications.
Zhang et al. (Mon,) studied this question.