A detection method for road network interchanges with the MeshCNN based on Delaunay triangulation

Abstract

Due to their unstructured characteristics, mature convolutional neural network (CNN) models often have difficulty performing spatial analysis with vector data. Current studies used graph neural network (GCN) models to address this problem. However, the definition of cognition factors involves uncertainties, making it challenging to accurately and comprehensively define these factors. In this paper, the road interchange detection task is taken as an example to introduce the MeshCNN, a deep learning model based on triangular mesh data, aiming to provide a new solution for spatial analysis of vector data. A triangular edge classification model is first trained with simple input features. Then, interchanges are detected based on the classification results with an adaptive method. Experiments were conducted on real-world road network data from four cities. The results reveal that the proposed method outperformed the existing methods with precision and recall rate of 89.36% and 79.25% for interchange detection on the total datasets. Furthermore, our proposed method can also detect interchanges in other regions more easily than the GCN method.

Keywords

• Road network
• vector data
• interchange detection
• Delaunay triangulation
• convolutional neural network (CNN)