Research On Road Extraction Method Of Trajectory Data Based On Machine Learning

Road data production plays a significant role in professional fields and practical applications such as digital cartography,map updating,national production activities,and military operations.Vehicle trajectory data has the advantages of easy data acquisition,low cost of use,and short data update period.With its road network correlation and spatial and temporal distribution characteristics,vehicle trajectory data has become one of the high-quality data sources for road data production.However,the traditional road extraction research based on crowd sources trajectories are faced with problems such as inaccurate data accuracy,inconsistent sampling intervals,and inconsistent spatial distribution density,which seriously hinders the application of road extraction based on vehicle trajectory data in actual production practice.The rapid development of artificial intelligence technology,especially machine learning,provides a brand new solution for many professional fields.The thesis focuses on the problems of data inconsistency,complex algorithm and poor universality of algorithm in road extraction based on vehicle trajectories.Using unsupervised learning clustering and conditional generative adversarial networks model in the field of machine learning as the technical breakthroughs,this paper focuses on how to use machine learning to enhance the universality of trajectory-based road extraction in the case of inconsistent trajectory data.The main work and innovations completed in this thesis are as follows:(1)Aiming at the clustering problem with multiple constraints,we propose a similarity measurement method of kernel distance,and improve the K-DBSCAN clustering algorithm based on kernel distance.This method raises the multi-parameter space to a high-dimensional feature space of the multi-constrained clustering algorithm,and then performs the similarity measurement of the high-dimensional feature space based on the kernel function,which reduces the parameters of the clustering algorithm with the multiple constraints.On this basis,for the roadway level road extraction problem with trajectory direction constraints,we performe the roadway level trajectory clustering under distance and orientation constraints based on the K-DBSCAN algorithm,and then extracte the roadway level roads based on the clustering results.Experimental data from several cities have verified the efficiency of K-DBSCAN algorithm and the validity of road level road extraction.(2)Aiming at the problem of how deep learning models are combined with road extraction based on trajectory data,converting road extraction problems into image generation problems is proposed.And a road extraction model of trajectory density to road based on conditional generative adversarial networks is proposed.The model takes the trajectory lines kernel density map and rasterized road map as prior knowledge,and propose a residual U-Net generator to enhance the image generator of the model in combination with the residual network.Through the continuous game of “two player game”,the real road images are gradually generated.On this basis,a centerline-level road extraction deep learning method based on trajectory density is proposed,and experiments verify the effectiveness of the proposed model.(3)Aiming at the problem of roadway level road extraction based on deep learning,a trajectory orientation-road extraction model based on conditional generative adversarial network is proposed.This model uses a multi-scale receptive field discriminator to enhance the multi-scale feature perception ability of the image to solve the problem of high frequency information coincidence of input images.In the process of vehicle trajectory rasterization,an orientation-color mapping rasterization transformation method is proposed to convert the orientation information to HSV color space to solve the problem of trajectories orientation information gridding.Based on this,a roadway level road extraction deep learning method based on trajectory orientation constraints is proposed.And experiments show that the method has strong universality for different data sources.(4)Aiming at the problem of road extraction based on multi-source data fusion and high-resolution image generation,a conditional generative adversarial network road extraction model based on the fusion of vehicle trajectory and remote sensing imagery is proposed.In this model,the two-layer depth U-Net generator is used to enhance the feature perception and fusion of the two data sources,and the multi-scale discriminator is used to enhance the high-resolution image discrimination and generation ability.On this basis,a road extraction deep learning method based on the fusion of vehicle trajectory and remote sensing image is proposed.The comparative experiments showed that the method based on the fusion is better than the single data source road extraction method.