Optimization On Visual Localization Based On Road Map

The ego-localization problem of mobile robot is a key problem in robot field.As the eye of the mobile robot,cameras have been widely used with the features of small size,low cost and wide application scenario.Visual localization has been popular in recent years because of the unstable localization result with traditional localization methods such as GPS in urban or indoor areas.As a classics visual localization method,visual odometry(VO)suffers from the problem of drift,which can't be used in long distance.In this paper,a novel multi-location joint filtering algorithm is proposed to restrain the VO's drift by including the constraint from a road map.Different with the traditional map-based localization methods,we believe the "node-edge" based map representation method can't well model the road map.Considering the imprecision of road map,the corrected trajectory after filtering is not strictly constaint on the edges of the road map.A flexible multi-location joint filtering is designed under the particle filter framework,which contains the VO's result and the information of road map.The method uses the VO's result as an initial localization result to match with the road map,and can well balance the information from VO and the road map.The method features making filtering only on the turning points of the trajectory,which adds little additional computation burden to VO.Experiments in various dataset including KITTI and the data acquired in our campus are carried out,and we compare our method with other localization methods with quantitative analysis,all the results demonstrate the accuracy and robustness of our method.