A Method For Alignment Recognition Of Mountainous Roads By Fusing Model Driven And Data Driven Methods

The road alignment design in mountainous areas is greatly affected by the terrain,the road alignment has strong continuity,and the non-linear coefficient is large,resulting in low recognition accuracy of road alignment element composition and element geometric feature parameters.This paper takes mountain road alignment elements as the research object,and analyzes domestic and foreign lines Based on the shape recognition method,a method for alignment recognition of mountainous roads by fusing model driven and data driven methods is proposed to improve the accuracy of alignment recognition.Based on the idea of multi-stage modeling,firstly,identify the composition of road linear elements through mathematical models;then,integrate the neural network model to accurately divide the boundaries of linear elements;finally,fit each linear element to calculate the geometric characteristic parameters of linear elements.The main work and contributions of this paper are as follows:1.Analyzed the characteristics of mountain road alignment data,and constructed the data set used in the paper.The data set is composed of three parts,which are remote sensing image data,linear design structure data and desensitization data.Based on the gray value distribution characteristics of remote sensing images,a dynamic binarization road extraction method is proposed,using morphological methods to extract the road alignment skeleton;based on the design and generation principle of road alignment elements,the generation of transitional curves is optimized by an iterative method.Finally,the three kinds of data are sorted and marked to form the data set in this paper.2.Established a model-driven mountain road alignment recognition model.In this paper,the model is modeled according to the idea of "road alignment composition identification-single alignment identification-road alignment segmentation point optimization".The mathematical model aims to obtain the road alignment composition and rough road segmentation.Firstly,the curvature feature extraction method is improved by analyzing the geometric characteristics of road alignment elements;Error sequence;finally,combined with the error evaluation index,the curvature feature is selected to establish a road alignment recognition model,and the recognition accuracy rate is 82.66%.3.Established a data-driven mountain road alignment recognition model.Taking the single linear element as the research object,based on the multi-dimensional data structure of mountainous road alignment point group,a two-dimensional normalized data preprocessing method is proposed.Using sample point coordinates,curvature features extracted by mathematical models,positions and logical directions as model inputs,the classification capabilities of LSTM,GRU and Transformer data-driven models for single road alignment elements were compared and analyzed,and the Apollo algorithm was used to build LSTMGRU network optimization data The driving model has a model recognition accuracy rate of94.84%,which is 6.47% higher than that of the single-layer GRU classification model.4.A method for alignment recognition of mountainous roads by fusing model driven and data driven methods is proposed.Through the idea of multi-stage modeling,a growth factor identification method is proposed,which uses the neural network to optimize the identification results of the mathematical model,and improves the accuracy of the segmentation points of linear elements.The accuracy of the model is 87.23%,which is 4.57% higher than that of the single mathematical model.5.Comprehensive case analysis.Collect the data of the Han road section of the Micang Mountain system to analyze the algorithm examples in the paper,use the least square method to fit each linear element,use the variation coefficient and limit index to evaluate the road linear elements,and build a test program to visualize the linear recognition results.