Region of attraction for a vehicle pose estimator utilizing monocular vision and lane marker maps

Currently, there is a lack of low-cost, real-time solutions for accurate autonomous vehicle localization. The fusion of a precise a priori map and a forward-facing camera can provide an alternative low-cost method for achieving centimeter-level localization. This paper analyzes the position and orientation bounds, or region of attraction, with which a real-time vehicle pose estimator can localize using monocular vision and a lane marker map. A pose estimation algorithm minimizes the residual pixellevel error between the estimated and detected lane marker features via Gauss-Newton nonlinear least-squares. Simulations of typical road scenes were used as ground truth to ensure the pose estimator will converge to the true vehicle pose. A successful convergence was defined as a pose estimate that fell within 5 cm and 0.25 degrees of the true vehicle pose. The results show that the longitudinal vehicle state is weakly observable with the smallest region of attraction. Estimating the remaining five vehicle states gives repeatable convergence within the prescribed convergence bounds over a relatively large region of attraction, even for the simple lane detection methods used herein. A main contribution of this paper is to demonstrate a repeatable and verifiable method to assess and compare lane-based vehicle localization strategies.