Vision Based Vehicle Localization for Infrastructure Enabled Autonomy

Abstract

Primary objective of this research is to devise techniques to localize an autonomous vehicle in an Infrastructure Enabled Autonomy (IEA) setup. IEA is a new paradigm in autonomous vehicles research that aims at distributed intelligence architecture by transferring the core functionalities of sensing and localization to infrastructure. This paradigm is also promising in designing large scalable systems that enable autonomous car platooning on highways. A reliable camera calibration technique for such an experimental setup is discussed, followed by the technique for 2D image to 3D world coordinate transformation. In this research, information is received from: (1) on-board vehicle sensors like GPS and IMU, (2) localized car position data derived from deep learning on the real-time camera feeds and (3) lane detection data from infrastructure cameras. This data is fused together utilizing an Extended Kalman Filter (EKF) to obtain reliable position estimates of the vehicle at 50 Hz. This position information is then used to control the vehicle with an objective of following a prescribed path. Extensive simulation and experimental results are also presented to corroborate the performance of the proposed approach.