Research On Semantic Segmentation Of Outdoor Scene Images Based On Domain Adaptation

In outdoor scenes,the automatic unmanned driving task is usually affected by a variety of complex scenes,such as rain,fog or night,which will make the segmentation model trained on the existing data set produce poor semantic segmentation performance in the above real complex scenes.Therefore,learning a semantic segmentation model that can adapt to changes in different environmental conditions is crucial for safe autonomous driving.In this thesis,we choose the outdoor scene at night as the research background.Aiming at the problem of feature distribution difference and brightness difference between daytime and nighttime images,we study the semantic segmentation algorithm of outdoor scene images based on domain adaptation.The main research work is as follows:1)Aiming at the problems of negative offset and large distribution difference caused by ignoring the semantic consistency of the same features in the existing domain adaptive night segmentation methods,a domain adaptive night semantic segmentation network based on semantic alignment optimization is proposed.The network designs a nighttime mutual exclusion classifier.In feature classification,the feature alignment degree of daytime and nighttime images is judged by the difference of their prediction results,and the adversarial loss of features is adaptively weighted according to the alignment situation to strengthen local semantic alignment.At the same time,a dual-view pseudo-supervised strategy is designed,which uses the daytime image prediction results in the target domain to provide pseudo-labels for nighttime image training to enhance nighttime segmentation prediction.The experimental results show that the network achieves 43.2 % and 46.1 % segmentation accuracy on the Dark Zurich-test and the Nighttime Driving-test,respectively.2)Aiming at the problem that the network performance degrades due to the difference of dynamic categories in the feature matching of day and night images in the target domain,an adaptive night semantic segmentation network based on dual-branch sample fusion is proposed.The network designs a dual-branch sample fusion module.In the first branch,a nighttime style image containing a daytime dynamic category is generated,while in the second branch,a daytime style image containing a nighttime dynamic category is generated by using the fusion image in the first branch to achieve complete consistency between the two in static and dynamic categories.Then they are input into the relighting network,so that the network only focuses on the change of brightness in the process of parameter optimization to improve the relighting performance,so as to enhance the final prediction result of the segmentation network in the feature generation stage.In addition,from the perspective of entropy,the night image entropy minimization method is used to narrow the entropy distribution between the target domain and the source domain,so as to indirectly complete the semantic alignment of day and night features.The experimental results show that the network achieves 45.2% and47.2% segmentation accuracy on the Dark Zurich-test and the Nighttime Driving-test,respectively.3)In order to test the robustness and generalization of the proposed network,we apply it to a variety of complex night scenes and real night scenes,and analyze the advantages and disadvantages of the proposed algorithm.The experimental results show that the proposed algorithm performs well in segmentation prediction under various night scenes,and can recover a large amount of detailed information of the target object,which has strong robustness and generalization.