A Deep Learning Based Methodology for Assessing Road Marking Wear from Laser Imaging
Résumé
As road markings are an essential guidance and safety element for drivers, it is crucial to monitor their wear level. Mobile laser mapping systems are becoming increasingly popular, particularly for measuring road deterioration, as they have the advantage of being independent of ambient lighting conditions. We have developed a deep learning method that automatically evaluates the percentage of residual marking (PRM) based on laser reflectance images, without the need for a segmentation step. To do this, we propose a variant of the EfficientDet model, that integrates an extra sub-network to predict the PRM and is trainable end-to-end. Since it is not possible to measure real road marking PRM values on a large-scale and no dataset are available, we use an indirect reference obtained by supervised image analysis. It conforms well to a physical reference resulting from markings of known PRM made with stencils. We propose to use synthetic marking generation to obtain data involving low-PRM markings, which can hardly be found on roads, and with a better balance of PRM values. Experimental results, obtained from synthetic images and real images taken on closed test tracks and open roads, show that the proposed model outperforms a two-stage approach that applies automatic segmentation (Otsu method) to the results of marking detection. They confirm that our deep learning methodology is capable of assessing PRM directly from the visual appearance of the marking.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |