Research On Semantic Cognition Method Of Road Layout In Complex Urban Traffic Environment

Traffic scene understanding is crucial for decision planning of intelligent vehicles,as well as for scene library enriching of intelligent driving test.However,current scene understanding methods are often limited to pixel-level representation,which cannot naturally support the inference and planning tasks of intelligent vehicles,and it is difficult to directly provide structured samples for the scene library.To address these issues,a semantic-aware method for road layout cognition in complex urban traffic environments is proposed in this paper.This is achieved by reasoning the semantic of road occlusion in traffic scenes,and then extracting road layout parameters.The main contents of the paper are as follows:(1)A semantic-guided road occlusion reasoning algorithm is proposed.To address the problem of redundant pixel information in road scene images captured by cameras,as well as the problem of various types of traffic participants in complex urban road scenes obstructing the complete structure of the road,this paper first uses a semantic segmentation network based on pyramid-shaped multi-scale pooling to perform pixel-level analysis of the road scene in the form of a semantic image,which compresses the redundant information in the image.Then,based on the pixel positions and regions of the foreground targets located by semantic,a mask is generated and applied to the semantic image.Finally,a generative network based on gated convolution is used to infer and generate semantically reasonable pixels in the masked image based on the contextual information of the unknown regions,and the initial inference results are supplemented by image residuals to restore small target pixels.(2)A domain-adaptive road layout parameter extraction algorithm is proposed.To address the problem of the scarcity of road layout parameters in real-world semantic images of roads,as well as the problem of different data distributions in different domains and adaptation guided by hard labels,this paper first introduces simulated road layout data with rich labels,and uses domain blending to mix the perspective-eliminated road semantic images with the simulated road images at the pixel level to retain the potential features of simulation and real domain and provide soft label guidance.Then,based on the idea of adversarial domain adaptation,a domain-consistent feature extraction network and a domain discriminator network are set up,and the zero-sum game strategy of confusion and discrimination is used to supervise the discriminator network and guide the feature extraction network to capture domain-consistent features.Finally,the road semantic image is transformed into a structured road layout parameters defined in the road layout attribute space through the feature extraction network and the parameter extraction network.(3)The testing validated the effectiveness of the road layout semantic cognition model and evaluated the intermediate outputs of each module.First,the model was tested comprehensively using the KITTI and Pbev Dataset datasets,and the results showed that this method can accurately convert road images to road layout parameters using the simulated domain label space.The method achieved an accuracy of 84.7% and 79.0% in binary and multi-class attribute conversion,respectively,and a error of 0.124 in continuous value attribute conversion.Second,the model was subjected to modular testing,and the intermediate results of each module performed well and could retain small target pixels in road scenes to a large extent.In addition,the results of the ablation experiments demonstrated the effectiveness of the model structural design.Finally,the method was compared and analyzed against existing road layout parameterization methods and various variant methods,and the comparative results showed that this method can achieve competitive conversion performance and better performance in binary attribute conversion.This paper relies on the key project of National Natural Science Foundation of China "Intelligent Vehicle complex dynamic environment depth level perception and understanding"(U1864204).This paper focuses on structured cognition of road layout in complex urban traffic environments.Through stages such as semantic segmentation of traffic scenes,road occlusion reasoning,and road layout parameter extraction,it achieves semantic analysis and parameterized understanding of road layout based on urban traffic scene images,providing a foundation for intelligent vehicle scene understanding and the construction of intelligent driving test scenarios library.