电子测量与仪器学报

杜运亮,王明甲.基于半监督域适应的微弱光环境下行人检测研究[J].电子测量与仪器学报,2024,38(1):106-113

基于半监督域适应的微弱光环境下行人检测研究

Research on pedestrian detection in low-light conditions basedon semi-supervised domain adaptation

DOI：

英文关键词:pedestrian detection low-light semi-supervised domain adaptation YOLOv8

基金项目:国家自然科学基金（61971253）、山东省自然科学基金（ZR2014FL026）项目资助

作者	单位
杜运亮	青岛科技大学自动化与电子工程学院青岛266100
王明甲	青岛科技大学自动化与电子工程学院青岛266100

Author	Institution
Du Yunliang	College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266100, China
Wang Mingjia	College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266100, China

摘要点击次数: 51

全文下载次数: 179

中文摘要:

为了解决可见光图像在微弱光环境下会出现检测性能下降的问题,本文提出了一种半监督域适应的行人检测算法。首先,结合均值教师模型和YOLOv8检测器搭建半监督检测网络;其次,使用图像融合算法和风格迁移算法相结合的方式生成伪图像进行伪交叉训练,减少图像之间的域差异问题;最后,将基于Transform的混合注意力机制引入主干特征提取网络,在提升图像分辨率的同时进一步提升检测精度。实验结果表明:在LLVIP数据集和KAIST数据集上,该算法的检测精度分别达到89.3%和66.8%,相比SSDA-YOLO算法分别高出7.6%和19.8%;相比Efficient Teacher算法分别高出4%和8.7%;相比全监督算法ICAFusion分别高出1.8%和17.9%。与以往的算法相比,该算法具有更高的检测精度。

英文摘要:

In order to solve the problem that visible light images suffer from performance degradation in low-light conditions, this paper proposes a semi-supervised domain adapted pedestrian detection algorithm. Firstly, the semi-supervised detection network is built by combining the mean teacher model and the YOLOv8 detector. Secondly, pseudo-images are generated for pseudo-cross-training using a combination of the image fusion algorithm and the style migration algorithm to reduce the problem of domain difference between images. Finally, the hybrid attention mechanism based on Transform is introduced into the backbone feature extraction network, which further improves the detection accuracy while enhancing the image resolution. The experimental results show that the detection accuracy of the algorithm reaches 89.3% and 66.8% on the LLVIP dataset and KAIST dataset, respectively, which is 7.6% and 19.8% higher compared to the SSDA-YOLO algorithm, 4% and 8.7% higher compared to the Efficient Teacher algorithm, and 1.8% and 17.9% compared to the fully supervised algorithm ICAFusion. Compared with previous algorithms, this algorithm has higher detection accuracy.

查看全文查看/发表评论下载PDF阅读器