Georeferencing of Road Infrastructure from Photographs using Computer Vision and Deep Learning for Road Safety Applications

Georeferenced information of road infrastructure is crucial for road safety analysis. Unfortunately, for essential structures, such as fences and crash barriers, exact location information and extent is often not available hindering any kind of spatial analysis. For a GIS-based study on wildlife-vehicle collisions (WVCs) and, therein, the impact of these structures, we developed a method to derive this data from video-based road inspections. A deep learning approach was applied to identify fences and barriers in photos and to estimate the extent and location, based on the photos' metadata and perspective. We used GIS-based analysis and geometric functions to convert this data into georeferenced line segments. For a road network of 113 km, we were able to identify over 88% of all barrier lines. The main problems for the application of this method are infrastructure invisible from the road or hidden behind vegetation, and the small sections along the streets covered by photos not depicting the tops of higher dams or slopes.