LCRM: an open dataset for assessing the wear of road markings

Every year regions repaint road markings — but on a schedule rather than according to actual wear. The catch is that markings do not wear out on a schedule. As a result, some stretches are repainted before they need it, while others are repainted only once the crossings and stop lines have become barely visible.

We have taken a step towards changing that.

Our team has created LCRM, a road marking dataset that makes it possible to train a model not merely to find markings in an image, but to understand their condition: fine, worth monitoring, or in urgent need of repainting.

Such a system can be installed on specialised road service vehicles or on city cars. They drive their usual routes and automatically assess the condition of every marking along the way. No manual inspections and no surprises: the city gets an up-to-date wear map and can plan repairs in advance.

We used a line-scan camera, which captures the road uniformly, without the perspective distortions that prevent ordinary dashcams from seeing thin marking lines clearly. Filming took place in Moscow, Saint Petersburg and the Moscow region at different times of day. The result: more than 20,000 images with 12 types of marking and three levels of wear.

The best of the tested models recognises markings with an IoU of 91.5%, which is already close to a level suitable for real deployment. The dataset is open and can be used by any team working on road safety or autonomous driving.