道路场景目标检测是智慧交通领域的重要组成部分,直接关系到众多智慧交通应用性技术的实施.然而,现有道路场景目标检测域泛化技术普遍存在域不变特征提取不充分、检测精度不高和泛化能力弱的问题.针对此问题,提出复杂天气条件下道路场景目标检测的域泛化方法.设计了道路场景域不变特征生成模型,分别提取源域图像的域内不变特征和域间不变特征,并生成更具多样性的复杂天气条件下的道路场景域不变特征,以提高目标检测模型的泛化能力;在此基础上,设计了道路场景目标检测域泛化模型,引入自蒸馏机制,使目标检测模型提取的特征拥有丰富的域不变特征,以进一步增强泛化能力,从而提高目标检测模型的检测精度.实验结果表明,所提出的目标检测域泛化模型性能与对比模型相比有明显提升,能显著提高目标检测模型的泛化能力和检测精度,其中F1-score较基线目标检测模型提升0.042$\sim$0.051,均值平均精度(mean average precision,mAP)提升3.0%$\sim$5.9%,证明了所提出的目标检测域泛化方法的有效性和优越性.
Road scene object detection is an important part of the intelligent transport field, which is directly related to the implementation of many intelligent transport applicative technologies. However, the existing domain generalization techniques for road scene object detection generally have the problems of insufficient domain-invariant feature extraction, low detection accuracy, and weak generalization ability. To address this problem, this paper proposed a domain generalization method for road scene object detection under complex weather conditions. The paper designed a road scene domain-invariant feature generation model, extracted the intra-domain invariant features and inter-domain invariant features of the source domain images, respectively, and generated more diverse road scene domain-invariant features under complex weather conditions to improve the generalization ability of the object detection model. On this basis, a domain generalization model of road scene object detection was designed, and the self-distillation mechanism was introduced to make the features extracted by the object detection model have rich domain-invariant features to further enhance its generalization ability, so as to improve the detection accuracy of the object detection model. The experimental results show that the performance of the object detection domain generalization model in this paper is significantly improved compared with the comparison models. The model can significantly improve the generalization ability and detection accuracy of the object detection model. Among them, the F1-score is increased by 0.042-0.051 compared with the baseline object detection model, and the mean average precision (mAP) is increased by 3.0%-5.9%, which proves the effectiveness and superiority of the object detection domain generalization method proposed in this paper.
[1] 林猛, 周刚, 杨亚伟, 等. 特殊天气条件下的目标检测方法综述[J]. 计算机工程与应用, 2022, 58(13): 36-47.
[2] 董天天, 曹海啸, 阚希, 等. 复杂天气下交通场景多目标识别方法研究[J]. 信息通信, 2020(11): 72-74.
[3] Huang S C, Le T H, Jaw D W. DSNet: joint semantic learning for object detection in inclement weather conditions [J]. IEEE Transaction on Pattern Analysis and Machine Intelligence, 2020, 43(8): 2623-2633.
[4] 董惠文, 田莹. 雾天车辆目标检测域适应模型[J]. 辽宁科技大学学报, 2022, 45(2): 104-112.
[5] 张永福, 宋海林. 基于跳跃特征金字塔的域适应目标检测模型[J]. 计算机技术与发展, 2022, 32(9): 28-35.
[6] Vidit V, Engilberge M, Salzmann M, et al. Clip the gap: a single domain generalization approach for object detection [C]// 2023 IEEE International Conference on Computer Vision (ICCV). 2023: 3219-3229.
[7] Wang W, Zhang J, Zhai W, et al. Robust object detection via adversarial novel style exploration [J]. IEEE Transactions on Image Processing, 2022, 31: 1949-1962.
[8] Xu M, Qin L, Chen W, et al. Multi-view adversarial discriminator: mine the non-causal factors for object detection in unseen domains [C]// 2023 IEEE International Conference on Computer Vision (ICCV). 2023: 8103-8112.
[9] Pan X, Luo P, Shi J, et al. Two at once: enhancing learning and generalization capacities via IBN-Net [C]// 2018 European Conference on Computer Vision (ECCV). 2018: 284-500.
[10] Pan X, Zhan X, Shi J, et al. Switchable whitening for deep representation learning [C]// 2019 IEEE International Conference on Computer Vision (ICCV). 2019: 1863-1871.
[11] Huang L, Zhou Y, Zhu F, et al. Iterative normalization: beyond standardization towards e-cient whitening [C]// 2019 IEEE International Conference on Computer Vision (ICCV). 2019: 4874-4883.
[12] Choi S, Jung S, Yun H, et al. RobustNet: improving domain generalization in urban-scene segmentation via instance selective whitening [C]// 2021 IEEE International Conference on Computer Vision (ICCV). 2021: 11580-11590.
[13] Zhang L, Song J, Gao A, et al. Be your own teacher: improve the performance of convolutional neural networks via self distillation [C]// 2019 IEEE International Conference on Computer Vision (ICCV). 2019: 3713-3722.
[14] Ji M, Shin S, Hwang S, et al. Reflne myself by teaching myself: feature reflnement via selfknowledge distillation [C]// 2021 IEEE International Conference on Computer Vision (ICCV). 2021: 10664-10673.
[15] 徐海, 谢洪涛, 张勇东. 视觉域泛化技术及研究进展[J]. 广州大学学报(自然科学版), 2022, 21(2): 42-59.
[16] Lu W, Wang J, Li H, et al. Domain-invariant feature exploration for domain generalization [EB/OL]. (2022-07-25) [2023-02-08]. https://arxiv.org/abs/2207.12020.
[17] Wu A, Deng C. Single-domain generalized object detection in urban scene via cyclicdisentangled self-distillation [C]// 2022 IEEE International Conference on Computer Vision (ICCV). 2022: 847-856.
[18] Ren S, He K, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks [C]// 2015 Neural Information Processing Systems (NIPS). 2015: 91-99.
