A novel approach for detecting roundabouts in maps based on analysis of core map data

Approaches for detecting roundabouts in maps are heavily dependent on looking at the problem from a machine-learning powered computer vision perspective. In this paper, we propose a fresh approach, taking core map data into account, that supplements existing techniques in a phenomenal way thereby significantly reducing the machine learning effort involved. As a direct consequence, our approach filters the training set to greatly reduce the scope of the ML problem resulting in increased accuracy. At the core of the proposed approach is the fact that data fields, which are used to describe maps, encapsulate geometric details about map points. If interpreted correctly, these details can be used to identify various map features including roundabouts. The proposed approach has two parts. First, an algorithm has been proposed which interprets core map data to identify roundabouts. This algorithm correctly detects roundabouts in more than 80% of the cases. Then, the remaining less than 20% cases are run through a machine learning model having extremely high accuracy because of a very specific training set. This results in an overall roundabout detection rate of more than 97%. Using this approach, we have succeeded in saving thousands of man-hours towards manual roundabout verification and correction.