Optimization On Visual Localization Based On Multi-lane Road Map

Localization of autonomous vehicles is one of the key technologies in automatic drive,and it has been a hot research at present.Traditional localization methods such as GPS and IMU are difficult to meet high-precision requirement in urban area.Visual localization has been popular in recent years because of its small size,low cost and high accuracy.Visual odometry(VO)is a classic method of visual localization.As a method of dead reckoning,it needs to give a definite starting pose,and it inevitably has accumulated errors.Vehicle-mounted sensors usually cannot provide an accurate starting pose.In long-distance positioning,the accumulated error of VO greatly affects the positioning accuracy in urban area.This paper proposes an optimization method based on multi-lane road map and particle filter framework aiming at the existing problems of VO in urban area.Trajectory representation based on anchor points is used to describe vehicles.A pseudo-real-time anchor point detection method is proposed to solve the problems existing in real-time anchor point detection.In view of the two problems of VO,this paper proposes a vehicle starting pose estimation method and a localization method using multi-lane road map based on Multi-Position Joint Particle Filtering.Experiments were conducted on KITTI datasets,ApolloScape datasets,and self-collected datasets,and comparisons were made with other similar methods.The experiments prove that the methods in this paper can estimate the starting pose accurately,correct the results of VO and eliminate the accumulated error of VO.All the results show our methods have advantages in accuracy and robustness.