Elevated Road Network: A Metric Learning Method for Recognizing Whether a Vehicle is on an Elevated Road

Abstract

Mobile navigation is a critical component in mobile maps. Yawing detection (does a vehicle yaw) is an important task in mobile navigation. In regions containing parallel and close elevated and surface roads, it is hard to detect yawing events using traditional methods, which mainly rely on low-accuracy positions and moving directions. Recognizing whether a vehicle is moving on an elevated road can significantly improve the performance of yawing detection. We propose Elevated Road Network (ERNet), a lightweight and real industrial neural network model for mobile navigation, to solve elevated road recognition fundamentally. For an elevated road fragment and a surface road fragment in the same group (they are parallel and close), ERNet takes four types of high-level features as input and learns two 10-dim descriptors (A and B). In inference stage, ERNet predicts a 10-dim embedding (C) for a position of a vehicle. By comparing ||A-C||2 2 and ||B-C||22, and applying a technique called confidence constraint, we recognize the road type corresponding to the position. Significant improvements on elevated road recognition and yawing detection have been achieved compared with several methods in extensive experiments. ERNet is deployed as part of AMap, the famous mobile map in China, serves drivers in three large cities: Beijing, Shanghai and Guangzhou, and will cover the whole country as soon as possible.