雾天场景下基于RT-DETR的轻量化车辆检测算法
1. 四川师范大学, 计算机科学学院, 四川成都 610101
2. 四川师范大学, 物理与电子工程学院, 四川成都 610101
| 摘 要： | 针对雾天场景下车辆检测任务中，检测精度与轻量化难以平衡的问题，提出了一种基于RT-DETR-r18的改进雾天车辆检测算法，该算法克服了模型的计算瓶颈，首先，设计了一种改进的主干网络（C2F_DYNCB）模块，结合动态卷积权重（DKW）机制和参数共享（DyInConv），构建了一种新型的动态卷积单元（DAIN Mixer）。该模块能够根据输入特征自适应调节深度卷积与特征融合方式，在显著减少参数量和计算量的同时，仍保持对多尺度空间特征的有效建模能力。其次，提出PEMD模块，通过引入Pola Attention线性注意力机制与EDFFN结构，在保持线性复杂度的同时大幅提高注意力图的判别性和模型性能；然后设计一个轻量化适配块MOEN模块，用于进一步整合局部信息并稳定特征分布。最后，采用WIoU-v3作为回归损失，通过在训练过程中自适应调整正负样本权重，增加对细微框的关注。实验结果表明，在REFYG雾天数据集上，改进算法的mAP50比原算法提高了3.05个百分点，模型的计算量与参数量分别降低了39.4%和32.8%，实现了显著的轻量化。 |
| 关 键 词： | RT-DETR; 雾天检测; 轻量化模型; 注意力机制; 动态卷积 |
| DOI： | 10.57237/j.se.2026.01.001 |
View:
Downloads:
收稿日期: 2025-12-11
接受日期: 2026-01-13
在线出版日期: 2026-02-03
Lightweight Vehicle Detection Algorithm Based on RT-DETR in Foggy Weather Scenarios
1. School of Computer Science, Sichuan Normal University, Chengdu 610101, China
2. School of Physics and Electronic Engineering, Sichuan Normal University, Chengdu 610101, China
| Abstract: | Addressing the difficulty in balancing detection accuracy and model lightweighting in vehicle detection tasks under foggy conditions, this paper proposes an improved foggy vehicle detection algorithm based on RT-DETR-r18. This algorithm overcomes the computational bottleneck of the model. First, an improved backbone network (C2F_DYNCB) module is designed, combining a dynamic convolution weight (DKW) mechanism and parameter sharing (DyInConv) to construct a novel dynamic convolution unit (DAIN Mixer). This module enables adaptive depth convolution and feature fusion, significantly reducing the number of parameters and computations while still maintaining detailed processing of input feature maps. Second, a PEMD module is proposed, introducing the linear attention mechanism Pola Attention and EDFFN, which significantly improves the discriminative power of the attention map and model performance while maintaining linear complexity. Then, a lightweight adaptation block, the MOEN module, is designed to further integrate local information and stabilize feature distribution. Finally, WIoU-v3 is used as the regression loss, adaptively adjusting the weights of positive and negative samples during training to increase attention to subtle bounding boxes. Experimental results show that on the REFYG foggy dataset, the improved algorithm's mAP50 is 3.05 percentage points higher than the original algorithm, and the model's computational complexity and number of parameters are reduced by 39.4% and 32.8%, respectively, achieving significant lightweighting. |
| Keywords: | RT-DETR; Foggy Weather Detection; Lightweight Model; Attention Mechanism; Dynamic Convolution |
基金
| 1. | 国家社会科学基金一般项目 (20BMZ092) |
参考文献
| [1] | Girshick R, Donahue J, Darrell T, et al. Region-based convolutional networks for accurate object detection and segmentation [J]. IEEE transactions on pattern analysis and machine intelligence, 2015, 38(1): 142-158. |
| [2] | Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need [J]. Advances in neural information processing systems, 2017, 30. |
| [3] | Zhao Y, Lv W, Xu S, et al. Detrs beat yolos on real-time object detection [C] // Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024: 16965-16974. |
| [4] | Chen Y, Dai X, Liu M, et al. Dynamic convolution: Attention over convolution kernels [C] // Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 11030-11039. |
| [5] | Meng W, Luo Y, Li X, et al. PolaFormer: Polarity-aware linear attention for vision transformers [J]. arXiv preprint arXiv: 2501.15061, 2025. |
| [6] | Ciubotariu G, Vasluianu F A, Zhou Z, et al. AIM 2025 Challenge on High FPS Motion Deblurring: Methods and Results [C] // Proceedings of the IEEE/CVF International Conference on Computer Vision. 2025: 5541-5550. |
| [7] | Zhu J, Chen X, He K, et al. Transformers without normalization [C] // Proceedings of the Computer Vision and Pattern Recognition Conference. 2025: 14901-14911. |
| [8] | Tong Z, Chen Y, Xu Z, et al. Wise-IoU: bounding box regression loss with dynamic focusing mechanism [J]. arXiv preprint arXiv: 2301.10051, 2023. |
| [9] | Cheng L, Zhang D, Zheng Y. Road object detection in foggy complex scenes based on improved YOLOv8 [J]. IEEE Access, 2024. |
| [10] | 孟志永,吴晨曦,王鹏,等.基于改进RT-DETR的井盖病害轻量化检测算法 [J/OL]. 计算机工程与应用, 1-13. (2025-09-28). https://doi.org/10.3778/j.issn.1002-8331.2506-0079 |
| [11] | Chen Y, Wang Y, Zou Z, et al. GMS-YOLO: A Lightweight Real-Time Object Detection Algorithm for Pedestrians and Vehicles Under Foggy Conditions [J]. IEEE Internet of Things Journal, 2025. |
| [12] | Li B, Ren W, Fu D, et al. Benchmarking single-image dehazing and beyond [J]. IEEE transactions on image processing, 2018, 28(1): 492-505. |
| [13] | Sakaridis C, Dai D, Van Gool L. Semantic foggy scene understanding with synthetic data [J]. International Journal of Computer Vision, 2018, 126(9): 973-992. |
| [14] | Kenk M A, Hassaballah M. DAWN: vehicle detection in adverse weather nature dataset [J]. arXiv preprint arXiv: 2008.05402, 2020. |
| [15] | Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector [C] // European conference on computer vision. Cham: Springer International Publishing, 2016: 21-37. |
| [16] | Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks [J]. Advances in neural information processing systems, 2015, 28. |
| [17] | Reis D, Kupec J, Hong J, et al. Real-time flying object detection with YOLOv8 [J]. arXiv preprint arXiv: 2305.09972, 2023. |
| [18] | Wang A, Chen H, Liu L, et al. Yolov10: Real-time end-to-end object detection [J]. Advances in Neural Information Processing Systems, 2024, 37: 107984-108011. |
| [19] | Khanam R, Hussain M. Yolov11: An overview of the key architectural enhancements [J]. arXiv preprint arXiv: 2410.17725, 2024. |
| [20] | Tian Y, Ye Q, Doermann D. Yolov12: Attention-centric real-time object detectors [J]. arXiv preprint arXiv: 2502.12524, 2025. |
| [21] | Carion N, Massa F, Synnaeve G, et al. End-to-end object detection with transformers [C] // European conference on computer vision. Cham: Springer International Publishing, 2020: 213-229. |
