Positioning is essential for the safe and accurate navigation of vehicles. Global Positioning System (GPS) is the most widely used positioning technique but requires online external correction information to achieve lane-level position accuracy $( < 1.5\mathrm{m})$. Furthermore, GPS suffers from poor performance when receiver is in signal-shadowed area such as urban environments. To address this problem, we propose the deep learning-based visual map-matching algorithm for correcting position of vehicles. Visual Map Matching Network (VMMNet) is introduced as an end-to-end network to match high-definition map (HD map) and Bird-Eye View (BEV) image from monocular camera for position correction. Because VMMNet exploits predefined HD map, it operates offline and can localize vehicles even in GPS-denied situation. Moreover, VMMNet can achieve lane-level accuracy regardless of signal environment while it only requires monocular camera which is already equipped in most of modern vehicles. We demonstrated superiority of VMMNet by quantitatively comparing to conventional position correcting techniques. The results showed VMMNet improves position accuracy of low-cost GPS receiver to the lane level at various urban areas with large margin to conventional techniques. We also confirmed qualitatively that VMMNet can function stably as an independent positioning system replacing GPS.
Vehicle Localization Network using Simultaneous Coarse and Fine Visual Map Matching
2023-09-24
6729200 byte
Conference paper
Electronic Resource
English
Similarity invariant partial shape matching using coarse-to-fine strategy
British Library Online Contents | 2014
|Range–Visual–Inertial Odometry with Coarse-to-Fine Image Registration Fusion for UAV Localization
DOAJ | 2023
|IEEE | 2017
|Coarse-to-fine visual autonomous unmanned aerial vehicle landing on a moving platform
DOAJ | 2023
|