| Deep learning technology has unique advantages in visual pattern recognition of autonomous driving scenes and is crucial to the development of vehicles.Due to the rich components and diverse targets of complex road scenes,identifying driving conditions and timely interacting with external environmental information can also contribute to the safety requirements of future autonomous driving.Semantic segmentation can detect obstacles pixel by pixel in complex road environments,so it plays an important role in the field of autonomous driving.This paper introduces the main methods of visual perception of complex urban traffic scenes,and puts forward research and innovation.It aims to use the idea of computer vision to perceive different levels of visual features,and establish segmentation methods and models to improve the perception of the surrounding environment and improve its accuracy.The specific research and contributions of this paper can be summarized as follows :(1)In this paper,the network structure of deplabv3+algorithm is improved to solve the problem that the semantic segmentation model of deplabv3+algorithm is not accurate enough when segmenting large targets.The deeplabv3+algorithm is improved by adding a global correlation network to optimize the segmentation boundary of the object and increase the receptive field of the network to improve the detection accuracy of the small target.Finally,the experimental results of the improved deeplabv3+algorithm are compared with deeplabv3+.Experiments show that the improved deeplabv3+algorithm is more excellent in the segmentation of boundary details of large objects in complex roads.Compared with the original deeplabv3+segmentation algorithm,the Intersection over Union of the improved deeplabv3+algorithm is 1.04 % higher.(2)In order to ensure that the resolution of the feature map is not lost,the dilated convolution is used for fusion to obtain a larger receptive field and meet the needs of resolution.Finally,the experimental results are compared.Compared with the original difference network,the Intersection over Union of the improved difference network is increased by 2.22 %.(3)The accuracy of the segmentation model and the control of network complexity and depth are typical problems that need to be solved urgently to realize autonomous driving technology.This paper proposes a pixel-level semantic segmentation network based on image cascade network architecture,which uses multi-scale branch and cascade feature fusion to extract rich multi-layer features.In order to extract the deep correlation information from the image,different task modules are constructed for different functional processing.The orderly and parallel functional modules further ensure the efficiency of obstacle detection in complex motion scenes.The multi-scale feature module is used to establish a feature pyramid to learn the location of abnormal target obstacles.The experimental results on the urban street dataset show that compared with other popular abnormal obstacle detection algorithms,the proposed model has obvious improvement in road image detail processing. |
