Unsupervised Domain Adaptive Underground Railway Segmentation

As an important basic technique in the subway autonomous driving system,the subway railway lane segmentation technique can provide guidance on the position of the track limit and the train’s drivable area for various downstream tasks,such as intrusion monitoring and obstacle detection.This paper mainly researches on the unsupervised domain adaptive segmentation technique of subway railway lane,which aims to apply domain transfer method to transfer the segmentation model is from the open-source above-ground railway dataset with ground truth labels to the unlabeled subway track line images,so as to realize the unsupervised segmentation of subway track lines,thereby saving the cost of labeling.At present,the existing unsupervised domain adaptive segmentation methods are mainly designed for urban road scenes,whose basic network models cannot adapt to the characteristics of subway track images.Furthermore,the transfer methods used in existing research cannot cope with the large domain gap between the above-ground railway scene and the subway railway scene.In view of the existing problems,the main work of this research is as follows:1)This study proposes a shallow feature aware segmentation model based on the characteristics of subway track images.This model optimizes the number of network convolution kernels in each layer of the existing segmentation model,increases the number of shallow feature map channels,and reduces the number of high-level feature map channels,so that the segmentation model pays more attention to the feature extraction ability of small-scale targets,so as to adapt to the subway track line segmentation The image scene in the task is simple but the target is small.In addition,in view of the problem that there is no relevant dataset for the task of subway track line segmentation,this study proposes the first known subway scene railway line segmentation dataset,which contains 5630 high-definition images of subway scenes,covering tunnels,platforms,depots,etc.Scenes.This study uses this dataset to test the proposed shallow feature aware segmentation model,which proves the superiority of the proposed model in the task of subway track line segmentation.2)This study proposes an unsupervised domain adaptive segmentation method for subway track lane based on pseudo-label multi-level correction.This method includes pixel-level and image-sample-level pseudo-label correction processes,which realizes pseudo-label correction from microscopic to macroscopic.The proposed method is capable to improve the quality of pseudo-labels in the self-training stage and solve the problem of large domain gap between aboveground railway scene and subway railway lane scene that most existing work cannot cope with.In addition,this study uses the proposed shallow feature aware segmentation model as the basic model and adopts the multi-level pseudo-label correction based unsupervised domain adaptive method for subway railway lane unsupervised domain adaptive segmentation experiment.The experiment result on the subway railway lane dataset,proves that this method can significantly improve the performance of subway railway lane unsupervised domain adaptive segmentation.3)This study proposes a knowledge distillation based unsupervised domain adaptive segmentation method for subway railway lane.The Vision-Transformer model is used as a teacher model to guide the CNN-based model for segmentation.We take the advantage of both the strong transferability of the Vision-Transformer model and the high calculation efficiency of the CNN model so as to improve the performance of unsupervised domain adaptive segmentation for subway railway lane without increasing the calculation cost.In addition,this study uses the proposed shallow feature aware segmentation model as the basic model and adopts the knowledge distillation based unsupervised domain adaptive segmentation method for subway railway lane unsupervised domain adaptive segmentation experiment.The experiment result on the subway railway lane dataset proves that this method can further improve the performance of subway railway lane unsupervised domain adaptive segmentation.