Research On Semantic Segmentation Method For Vehicle Viewed Fog Scene Based On Domain Adaptive Technology

Semantic segmentation technology under the vehicle perspective in foggy weather scenes can provide strong perception capabilities for automatic driving,visual navigation,and other applications.It can also save a lot of costs on installing additional sensors.With the continuous development of deep learning technology,neural network-based methods have become the mainstream technology for semantic segmentation.Developing safe and stable intelligent semantic segmentation models and tools has important application and research value in achieving high-level automatic driving.The deep learning-based semantic segmentation method from the vehicle perspective can automatically extract high-level visual features from the image and has higher recognition accuracy and automation level compared to traditional interpretation methods.However,there is less research on semantic segmentation problems in harsh environments such as foggy weather in the current academic community.On the one hand,due to visual occlusion and image quality degradation caused by harsh environments,segmentation models trained under the full supervision paradigm in clear scenes often experience severe performance degradation.On the other hand,the cost of pixel-level semantic annotation in harsh environments is one order of magnitude higher than in clear scenes,which limits the large-scale promotion and application of neural network-based segmentation models in harsh environments.This article proposes a domain adaptive technique-based vehicle perspective foggy scene semantic segmentation method to address the aforementioned issues.Starting from the simulated clear scene data(which is easily obtainable),the method gradually transfers the segmentation knowledge from the simulated clear domain to the real clear domain and then to the real foggy domain.Specifically,the article proposes and verifies an adversarial training domain adaptation method based on an improved Transformer and a reinforced spatial attention discriminator,which effectively transfers the knowledge from the simulated clear domain to the real clear domain.Subsequently,a self-training domain adaptation algorithm based on common domain feature teachers is designed to further transfer the segmentation knowledge to the real foggy domain,thereby providing a complete solution for semantic segmentation of unlabeled foggy scene data.The main innovations of this article are summarized as follows:(1)Focusing on the insufficient feature extraction ability and severe information loss of discriminator output probabilities in existing adversarial domain adaptation models,this article proposes an adversarial training domain adaptation method based on an improved Transformer and a reinforced spatial attention discriminator,achieving competitive performance on the benchmark datasets GTA5->Cityscapes and SYNTHIA->Cityscapes.(2)Based on(1),this article designs a self-training domain adaptation method based on common domain feature teachers to address the huge domain gap caused by fog,achieving better results than traditional methods on real foggy scene datasets such as ACDC and Foggy Zurich.