Improving the Road Topological Relation Based on a Road Intersection Detection Method

Jun He, Jinpeng Li, Weijia Li

Abstract: Road network vector extraction has important practical significance for travel navigation, urban planning and many other applications. At present, road network vector extraction mainly involves segmenting road information from remote sensing images, converting raster road network information into vector road network data through vectorization method, and obtaining the final results by modifying the topological relationship manually. The current road extraction method is difficult to obtain road intersection information, especially when roads are stacked. Therefore, although relatively complete segmentation results can be obtained, it is difficult to show correct topological relations in vectorization. Therefore, this paper proposed a method combining object detection and node iterative search to correct the topological relationship of stacked road vectors. Among three mainstream detection frameworks, i.e., Faster R-CNN, Libra R-CNN, and YOLO v7, YOLO v7 achieved the best results in intersection object detection, which can reach 0.931 mAP 50 and 0.585 mAP50:95 for the test set. In addition, the road vector modified by node iteration search showed a more real road topology relationship in the shortest path analysis.