Segmentation models are pixel-wise classifications of an image, dividing it into pre-defined classes. Road segmentation, i.e., a binary segmentation of road and background pixels, is a powerful technique, especially useful for mapping areas with scarce data sources or as a backbone for road safety modelling, providing road data for further analysis. The unique road network characteristics, such as format, color and infrastructure constraints, allow for tailored model development. This study evaluates four state-of-the-art models: LinkNet, U-Net++, GCB-Net, and DiResSeg, alongside post-processing techniques including clustering, thinning, smoothing, and grouping. As expected, LinkNet delivers high accuracy with relatively fast training, even on large datasets. In contrast, the complex architecture of U-Net++ results in significantly longer training times. Among models specifically designed for road segmentation, the convolutional kernel size appears to impact computational demand more than it does predictive performance. As for the post-process techniques, even the simplest ones are valuable to reduce noise, although the high level of confidence from the original models makes it difficult to differentiate noise from actual disconnected road segments. This work offers a practical comparison of accessible techniques to aid researchers in building efficient segmentation pipelines for road safety applications.
| ID | pc603 |
| Presentation | |
| Full Text | |
| Tags |
