Design And Research Of Road Scene Semantic Segmentation Method Combined With Multi-Level Features

With the continuous update and iteration of computer technology,image semantic segmentation technology,which is one of the important means of scene understanding,has also made rapid development.In the field of autonomous driving,semantic segmentation technology for road scenes can assist autonomous vehicles to avoid surrounding pedestrians,vehicles,separation belts and other targets by classifying the road scene images captured by vehicles at the pixel level.Traditional convolutional neural networks cannot make full use of different levels of features when performing semantic segmentation tasks,which will lead to the loss of some spatial detail information,and it is easy to misclassify object details in the face of complex road scenes.In order to solve the above problems,this paper designs a semantic segmentation network of road scenes combining multi-level features from the aspects of segmentation accuracy and segmentation speed.The main research contents are as follows:(1)A road scene semantic segmentation network combining multi-level features is designed.The network obtains and combines different hierarchical features by using a threebranch structure containing shallow feature branches,intermediate feature branches,and deep feature branches.Based on the idea of optical flow field and attention mechanism,a feature fusion module is designed,which can establish the position correspondence between different levels,which effectively assists the fusion of multi-level features.The network achieves 70.2%segmentation MIo U on the Cityscapes dataset,which is significantly better than networks such as FCN,Seg Net,U-Net and PSPNet.(2)A lightweight road scene semantic segmentation network combining multi-level features is designed.Based on the Mobile Net V2 feature extraction network,a multi-level feature fusion module is created,which integrates the different levels of features extracted by Mobile Net V2 through bottom-up and layer-by-layer fusion,and pays more attention to smallscale objects while maintaining the lightweight model,so as to achieve the balance between model segmentation speed and segmentation accuracy.The network was experimented on the Cityscapes and Cam Vid datasets,in which the segmentation MIo U on the Cityscapes dataset reached 66.3% and FPS reached 44.6;the segmented MIo U on the Cam Vid dataset reached63.8% and FPS reached 48.1.Compared with Seg Net,ENet,and ICNet,the network improves segmentation accuracy while maintaining lightweight,achieving a balance between segmentation speed and segmentation accuracy.