Opt Lett. 2021 Nov 15;46(22):5679-5682. doi: 10.1364/OL.441817.
A long-wave infrared polarization imaging technique recently has been applied in full-time road detection. However, the existing heuristic method has the limitation of fully using the polarization information of the road. In this Letter, we propose a polarization-guided road detection network collaborating with the distinguishable polarization characteristics of the road. A two-branch network is proposed to perform accurate road detection with infrared polarization images as inputs. A coarse road map obtained by thresholding the polarization images of the road guides the network to focus on the road regions through a polarization-guided branch. We also design a road-region-aware feature fusion module to fuse the features from two branches. This customized design of the network gives full play to the advantages of deep learning networks and polarization information. Experiments on a public infrared polarization dataset of road scenes demonstrate that the proposed road detection network outperforms state-of-the-art real-time segmentation networks with fewer parameters and faster speed.
一种长波红外偏振成像技术最近已被应用于全时道路检测。然而，现有的启发式方法在充分利用道路的偏振信息方面存在局限性。在本信中，我们提出了一种偏振引导的道路检测网络，该网络与道路的可区分偏振特性协同工作。该网络由两个分支组成，用于将红外偏振图像作为输入进行准确的道路检测。通过对道路的偏振图像进行阈值处理得到的粗略道路图引导网络通过偏振引导分支将注意力集中在道路区域上。我们还设计了一个道路区域感知的特征融合模块，用于融合来自两个分支的特征。该网络的定制设计充分发挥了深度学习网络和偏振信息的优势。在道路场景的公共红外偏振数据集上的实验表明，所提出的道路检测网络在参数更少和速度更快的情况下优于最先进的实时分割网络。
