Naver Map Navigation, Smooth Roads Only: A Narrow Road Detection Model

Navigation quality depends on accurate label classification, and in particular, detecting narrow roads in advance—so that level 2 roads are not unnecessarily guided—was the key task.

The first approach estimated road width from digital topographic map polygons. Since a simple centroid could fall outside a curved polygon, the team approximated the medial axis using a Voronoi Diagram, found candidate points inside the polygon, and added the midpoint of the polygon's minimum boundary rectangle diagonal along with the internal centroid to calculate width. In the end, width was defined as the largest of the minimum distances between candidate points and the outline, multiplied by 2.

This method worked well overall, but when the width of a digital topographic map polygon fell below a certain threshold, there were cases where the accuracy of distinguishing the actual level dropped. In particular, false positives remained where roads that were actually a lower level were classified as higher, requiring more precise discrimination.

The second approach was a model utilizing Street View images. The team first tested image classification to distinguish narrow roads from very narrow roads, but there was a problem where wide roads were misclassified as narrow when the background had a lot of soil or grass colors, as with farm roads. They then shifted direction toward road segmentation, creating a method based on YUV color similarity that found pixels similar to the road color, but this had clear limitations due to shadows, partial paving, and similarity to background colors.

In the end, they used semantic segmentation based on SegFormer. They cropped the region of interest from Street View and used the pretrained weights as-is, effectively handling only two classes: road and car. Rather than simply using the road ratio, they then found the points where the road mask ratio increased on a window-by-window basis to estimate the pixel positions of both edges of the road, and determined narrow roads based on the pixel-unit width between the two edges.

The comparison results were also clear. The polygon-based model caught more narrow roads but had a lower level 2 detection rate, while the image-based model detected fewer overall but identified level 2 more accurately. In particular, cases involving shadows, background-similar colors, and non-uniform road colors—which had been problems with the earlier color similarity method—were resolved.

Additionally, since images also contain information about parked/stopped vehicles, more sophisticated discrimination is possible through road + car masking. However, since driving vehicles and parked/stopped vehicles currently need to be distinguished, this is being used only conservatively for now. Based on narrow roads at around level 1.5, roughly 96% of Street View image coverage has been secured, leaving significant room to expand into broader quality improvement work going forward.