Arecognition And Simplifie Dmethod For Tricky Intersections Based On GoogLeNet Model

Complex intersection is a typical micro structure of road network,which has complex structure and diverse shape changes.The identification of this structure has always been a difficult and hot topic in research.Traditional recognition methods mostly rely on low-level features of artificial design.In recent years,some scholars have made a preliminary exploration on recognition methods based on deep learning.However,the existing methods have not been able to effectively describe the detailed features of complex intersections,resulting in limited recognition types and low accuracy.How to effectively and accurately describe the detail features of complex intersections is very important for the accurate recognition of complex intersection types and the simplification of complex intersections.Based on the research and analysis of complex intersections with different morphological characteristics in the national urban road network of OSM,this paper summarizes the detailed characteristics of complex intersections.On this basis,combined with the characteristics of complex intersections and convolution neural network,a model of identifying complex intersections is established by using Goog Le Net neural network,and through this model,the complex intersections in the road network are accurately identified,and the identified complex intersections are studied and analyzed.The characteristics of the main road and the auxiliary road in the complex intersections are found,and the corresponding laws are found,and then the complex intersections are simplified.The main results and conclusions are as follows:(1)Based on the characteristics of node density at intersections,Delaunay triangulation network is constructed to cluster points,and the center position and spatial range of complex intersections are initially determined;Secondly,39 important urban road networks are selected as training samples nationwide,and make full use of the advantages of vector data structure to enrich the sample type and capacity by simplifying,rotating,mirroring,etc;Finally,according to its rich local features,Goog Le Net neural network was selected for training to learn its high-level fuzzy features.Taking Tianjin OSM urban road network as an example for experimental analysis,The results show that the recall rate and precision rate of complex intersections in the road network have reached 92.55% and91.32% respectively,which are 3.31% and 11.43% higher than those based on the Alex Net network model.The recognition results of the two neural networks are basically the same for typical and low-interference complex intersections.However,for trumpet type,alfalfa type and multi disturbance irregular complex intersections,the recognition results based onthe Goog Le Net network model are 21 and 33 higher than those based on the Alex Net network model,and the accuracy is 7.40% and 6.40% higher respectively,which shows that the former is better than the latter in identifying local details.The method in this paper can effectively identify complex intersections,and significantly improves the accuracy and accuracy of recognition,and has strong generalization and anti-interference.(2)The accurate identification of main road and auxiliary road is essential for the automatic synthesis of complex intersections.Firstly,the feature points of complex intersections are extracted,then the arc segments are interrupted according to the feature points,and the parallel arc segments are identified according to the straightness of the road segment,then the parallel clusters are identified according to the road ductility,angle,distance and other features to obtain the main roads in complex intersections,and then the ramps are identified through the compactness and distance relationship to obtain the auxiliary roads in complex intersections.Taking the urban road network of Nanjing OSM as an example,the experiment shows that this method can accurately identify the main road and auxiliary road section of the complex intersection,and the recognition accuracy is93.60% and 89.43% respectively.At the same time,considering the topological relationship between the main and auxiliary roads in the complex intersection,the auxiliary roads with ramps are effectively removed,and the main roads in the complex intersection are well preserved,so that the simplified complex intersection can effectively ensure the connectivity and topology of the road network.