Semantic Segmentation Method Of Vehicle Video Object Based On Graph Convolution

The vehicle video image is the basic data of auto-driving,and the semantics segmentation technology for vehicle video objects directly affects the level of auto-driving.Because the traditional semantics segmentation method requires manual segmentation on a frame-by-frame basis,although the accuracy will be high,its huge workload leads to the inability to be applied in the field of auto-driving,and the fast-response deep learning-based semantics segmentation method for video objects is often not accurate.Therefore,this paper designs a semantically segmented network model which can quickly perform high-precision semantics segmentation based on improving the accuracy of semantically segmented vehicle video objects,and lays a good foundation for the future application of this model in the field of auto-driving.The main research work and results of this paper are as follows:(1)To solve the problem of insufficient samples in the dataset,this paper uses rotation,inversion,and other means to enhance the data of two open datasets Camvid and Cityscapes.At the same time,the ground true value labeling is adjusted for the data samples after data enhancement,which enlarges the number of samples in the dataset to five times the original,effectively avoiding the problem of over-fitting that may occur during network training.(2)To solve the problem of low accuracy of semantic segmentation network based on indepth learning,this paper combines traditional convolution network and convolution network to construct SegGCN semantic segmentation network.On this basis,the SE attention mechanism is added to the network.Based on this,a semantic segmentation network SESegGCN based on convolution structure is proposed,which takes advantage of the particularity of convolution structure and improves the perception field of the network.It avoids the loss of local location information.At the same time,with the addition of SE attention mechanism,only a small amount of network computing is increased,in exchange for a great improvement of network performance.(3)In order to verify the effectiveness of the proposed model,the experimental model was designed for comparative analysis,the experiment as a whole took Python as the main programming language,using two public data sets with data enhancement as the experimental object,compared with the mainstream SegNet,FCN8,Deep Labv3 and Nested UNet,after comparison,the SE-SegGCN constructed in this paper increased the semantic segmentation MPA to 89.7%,and the semantic segmentation accuracy of each subclass was also improved.Among them,the semantic segmentation accuracy of vehicles,pedestrians and other categories that are more important for automatic driving is 17% higher than that of the original network,and the average accuracy of the overall semantic segmentation is also improved by 5%,and the FPS can reach 7.147,which can meet the rapid response of semantic segmentation of vehicle video objects,and verify the effectiveness of the semantic segmentation network constructed in this paper.