Visual Semantic Segmentation Model And Its Application In Intelligent Vehicles

As one of the key perception technologies of intelligent vehicles,visual perception can promote the development of intelligent vehicles,alleviate traffic congestion,reduce traffic accidents,and improve the safety and comfort of occupants.Vision-based semantic segmentation intelligent vehicle environment perception method has the advantage of being able to obtain the pixel area of the target and the contour information of irregular objects.However,the current semantic segmentation environment perception of intelligent vehicles still faces the problems of incomplete segmentation targets,difficult segmentation of small targets,and slow model inference speed.These problems prevent the application of vision-based semantic segmentation environment perception technology in practical scenarios of intelligent vehicles.This paper focuses on the design and optimization of the intelligent vehicle environment perception semantic segmentation model.Taking the recognition of road garbage of cleaning vehicles as the main research object,and other intelligent driving scene environmental perception targets as the secondary research objects,special and general semantic segmentation models are designed.The main research contents and innovation points of this paper are as follows:1.From the shape characteristics of road lanes and parking slots,the linear long-distance feature dependencies are analyzed,and combined with the existing dilated convolution method,semantic segmentation models with horizontal and vertical dilated convolutions are designed.Achieving more accurate road lane and parking slots segmentation on the TuSimple lane and SS parking slots segmentation datasets.The optimal connection position,size,and connection method of horizontal and vertical dilated convolution in the linear segmentation model are obtained through generalization experiments.2.From the color characteristics of road garbage of cleaning vehicles,combined with the attention mechanism of deep learning and quaternion color representation,semantic segmentation models with color attention module are designed.A road garbage segmentation(RGS)dataset for cleaning vehicle is created.The color attention module is used to learn the weight coefficients of different color channels of the input images,and then the convolutional network is guided to learn the use of each channel color.Model comparison experiments show that the designed color attention module improves the results of road garbage segmentation on specific color spaces and backbone models.The correlation between the information learned inside the semantic segmentation models and the input color is observed from experiments with different training and testing inputs.3.From the mechanism of deep supervised learning to enhance the segmentation results of the semantic segmentation model for small road garbage.The convergence rate and minima of deep supervised networks are analyzed according to local strong convexity.Deep supervision branches are added to the UNet++upsampling layers,and the encoding and decoding layers of the model are supervised at the same time.The results of model comparison experiments show that the designed model improves the segmentation results of small objects such as stones and sand on the RGS dataset.From the output of the supervised branch and the feature map visualization results,the impact of supervised learning on the training of the hidden layers can be observed.4.From the mechanism of feature fusion and deep supervised learning,high-precision intelligent vehicle environment-aw are semantic segmentation models are designed.The hidden layers are shared between the supervision branches and the high feature fusion layers,and the features learned by the deep supervision branches are transferred to the output layer using the high fusion method,which improves the segmentation accuracy of road garbage for cleaning vehicles.By comparing the results of road garbage detection and segmentation methods such as YoloV3 and KNN,the advantages of semantic segmentation models in road garbage perception tasks are illustrated.Generalization experiments were carried out,and the designed model also achieved stable performance improvement on the PSV parking slots datasets.5.The number of network layers are introduced as a hyperparameter into the model training process to obtain an efficient and lightweight semantic segmentation model based on the principle of deep supervision.Through the analysis of the real-time and accuracy requirements of lightweight semantic segmentation networks in practical applications,loss weight functions that have a linear or one-dimensional quadratic relationship with the number of network layers are designed to change the role of the network supervision branch in the training process.The model comparison experiment results show that the designed loss weight functions improve the evaluation accuracy of the lightweight semantic segmentation model on the RGS dataset.Generalization experiments on Cityscape,CamVid,and PSV segmentation datasets using deep,medium,and shallow semantic segmentation model.The results show that the middle and shallow models have achieved certain performance improvements,while the performance of deep semantic segmentation models has declined.In summary,based on the target apparent feature learning,supervised learning,and feature fusion methods,intelligent vehicle environment perception semantic segmentation models that focuses on the shape,color,and details of the segmentation targets are designed in this paper.The performance of the designed model is verified on different vehicle environment perception semantic segmentation datasets,and the research goals of model design,accuracy improvement and application deployment on intelligent vehicles for visual semantic segmentation environment perception are achieved.