An infrared pedestrian detection method based on segmentation and domain adaptation learning
An infrared pedestrian detection method based on segmentation and domain adaptation learning
复制标题
DOI：
10.1016/j.compeleceng.2022.107781
复制
发表时间：
2022-04
期刊：
影响因子：
--
通讯作者：
Jianlong Zhang;Chi-Yun Liu;Bin Wang;Chen Chen-Chen;Jianhui He;Yang Zhou;Ji Li
中科院分区：
文献类型：
--
作者：
Jianlong Zhang;Chi-Yun Liu;Bin Wang;Chen Chen-Chen;Jianhui He;Yang Zhou;Ji Li
红外图像由于具有夜视能力，作为可见光图像的有效补充，在监控系统中得到了广泛的应用。然而，红外行人检测的发展仍然受到红外图像的弱特征和有限的多样性的阻碍。针对这两个问题，我们设计了一个多任务的行人检测学习框架，结合语义分割分支和领域适应分支。语义分割由UNet网络和Swin Transformer构成，能够将空间约束应用于行人检测。域自适应分支将红外和可见光图像的特征进行匹配，以提高场景的多样性。此外，三个任务共享一个基本的特征提取网络，以减少计算成本。实验结果表明，在XDU-NIR 2020数据集和CVC-09数据集上，该方法的平均精度（AP）分别比EfficientDet网络高出上级2.0%和2.2%.
Benefiting from the capability of night viewing, infrared images has been widely applied to surveillance systems as an effective complement to visible-light images. However, the development of infrared pedestrian detection is still impeded by weak features and limited diversity of infrared images. Aiming at these two problems, we designed a multi-task learning framework for pedestrian detection by incorporating a semantic segmentation branch and a domain adaptation branch. Composed of UNet network with Swin Transformer, the semantic segmentation could apply spatial constraints to pedestrian detection. The domain adaptation branch aligns the features between infrared and visible-light images to improve the scene diversity. In addition, three tasks shared a basic feature extraction network to reduce computation cost. The experiment results show that the average precision (AP) of our method is superior to the EfficientDet network by 2.0% on the XDU-NIR2020 dataset and 2.2% on the CVC-09 dataset respectively.