Research On Vehicle Semantic Communication Technology For Image Segmentation

Information sharing between vehicles can help autonomous vehicles obtain comprehensive road traffic information and make more favorable driving decisions.As one of the most important data sources of vehicles,the visual sensor will generate a large amount of image data,but the existing vehicle communication technology is difficult to meet the high speed(1Gbps)and high reliability(99.999%)of image transmission requirements.The semantic communication with the goal of ’expressing meaning’extracts information at the semantic level,reduces the actual amount of data required for transmission,and provides a feasible method for efficient transmission of large amounts of image data.Therefore,this paper focuses on image semantic information extraction technology,and designs an image segmentation oriented vehicle semantic communication system IS SC(Image Segmentation Semantic Communication)to improve transmission efficiency.The following two sections include the primary work and research contents of this paper:The key to the automatic driving task is to accurately understand the road scene,that is,to accurately understand the category,size and distance of vehicles,pedestrians and other targets,without understanding all the details of the image.Therefore,an image segmentation-oriented vehicle semantic communication system ISSC is proposed,which can directly perform image segmentation to obtain scene information at the receiving end.Secondly,aiming at the problems of low semantic extraction efficiency and insufficient semantic extraction in the existing image semantic communication system,Swin Transformer(Hierarchical Vision Transformer using Shifted Windows)is used to construct the semantic extraction module.Based on the multi-level local self-attention mechanism and the shift window strategy,the multi-scale global semantic features of the image are extracted to avoid the complex network structure and large amount of calculation caused by the cascaded convolutional layer.The simulation results show that the minimum signal-to-noise ratio required by IS SC system is 23 dB lower than that of the traditional method under the same segmentation accuracy requirement.In addition,under the condition of high signalto-noise ratio,the data transmission of ISSC system is reduced by 50%compared with the traditional method to achieve the same segmentation accuracy.According to the different effects of different types of targets on driving decisions,an enhanced semantic communication system En-ISSC for target differentiation segmentation is proposed.The system improves the segmentation effect of high-attention targets by improving the network structure and optimizing the loss function.At the same time,the online difficult sample mining strategy is introduced to solve the problem of learning difficulties caused by too few target samples.The simulation results show that the average intersection ratio of En-ISSC system on the concerned target is increased by about 3%.