Machine Learning Based Road Markings Extraction and Recognition From Laser Scanned Data
Information about location and types of road markings are among the crucial information that drivers perceive for a safe and a comfortable driving. Hence road markings are needed to be maintained and updated on a regular basis. With the recent leaps and bounds in the development of self-driving vehicle technology, the importance of regularly maintaining road markings, and moreover automatic detection and recognition of road markings have become more crucial than ever before.
Mobile mapping systems (MMS) which is composed of laser sensors, cameras, navigation support systems, etc. are one of the technologies that could be utilized not just for road asset management but also for creating maps related to self-driving vehicles.
Laser scanning is an active remote sensing technology. Hence, an intensity, which is primarily used for road markings extraction and recognition, is generally not affected by shadows or change in lighting conditions as in the MMS borne imageries. However, even for the same material, absolute intensity of the laser scanned data varies considerably with range, incidence angle, laser pulse strength, surface characteristics and so on. Moreover, density of laser scanned point cloud also varies with the distance from sensors which adds additional challenges for automatic extraction and recognition of road markings.
In this research, we have proposed a fully automatic work flow to accurately extract and to recognize road markings from the road surface using machine learning techniques. With our proposed algorithm, road markings extraction and recognition is largely immune to the change in the absolute value of intensity. The results showed that with our proposed algorithm, almost all of the road markings could be detected irrespective of the change in absolute intensity as long as the road markings are not faded out entirely.
We have also compared the result of automatic road markings extraction and recognition against the manual one. Our evaluation confirmed that the result of automatic extraction and recognition is close to a result obtained by a human interpreter and the proposed technique is robust enough for practical application.