Research On Road Recognition And Extraction Of UAV Image Based On Semantic Segmentation

With the vigorous development of intelligent transportation,accessing the accurate road information rapidly is a current research hotspot.Unmanned Aerial Vehicle(UAV)can obtain road images flexibly and conveniently,and semantic segmentation technology based on deep learning provides an efficient solution for understanding the complex road scenes in UAV aerial images.In this paper,the road recognition and extraction technology of UAV image based on semantic segmentation is studied,and the method of fusion of context feature and detail feature is proposed.The high-quality semantic segmentation of high-resolution UAV image is realized,and the structure of deep feature propagation based on optical flow is constructed.The real-time semantic segmentation of UAV video image data is realized.On the basis of theoretical research,a UAV road recognition and extraction system for road scene understanding is designed and developed.The specific research contents are as follows:(1)The semantic segmentation theory is analyzed and studied.Starting from the basic structure of semantic segmentation network,this paper explains the necessary structures of convolution layer and pooling layer,expounds the performance evaluation indexes of semantic segmentation,and analyzes three classical semantic segmentation networks.(2)An image semantic segmentation model that fuses contextual features and detail features is constructed.In order to solve the problem of feature loss caused by slicing or downsampling of high-resolution UAV aerial images,a context branch and a detail branch are designed to obtain the global context feature and local detail feature of the image,and the two features are fused;ResNet network is used to replace the encoder of U-NET to increase the network depth;An attention channel is integrated into the U-Net’s skip connections.Experiments on UAV image datasets prove that the model can accomplish the task of semantic segmentation of highresolution images.(3)A video semantic segmentation model based on deep feature propagation is constructed.Aiming at the redundant characteristics of video data,a deep feature propagation structure based on optical flow is constructed to speed up the segmentation;A basic semantic segmentation network consists of a shallow feature extraction branch and a deep feature extraction branch is proposed;It is proved that the shallow features can be used as the discriminant basis for key frames;It is verified that the video semantic segmentation model based on deep feature propagation can achieve real-time segmentation of video data while ensuring accuracy.(4)The UAV road recognition and extraction system for road scene understanding is designed and implemented.Based on the analysis of system requirements,the overall architecture and detailed functions of the system are designed,and tools such as python+tkinter+SQL Server are used for system development;The system can complete semantic segmentation datasets management,semantic segmentation models management,road recognition extraction,road scene understanding and other functions.