Integrated metric-topological localization by fusing visual odometry, digital map and place recognition

Abstract

Visual odometry, map-assisted methods and place recognition are all popular approaches to localize a mobile vehicle from three different perspectives. Separate implementation of these methods may cause the localization system vulnerable due to the drift issue and local pose estimation of visual odometry, the on-road assumption and tough initialization of map-assisted methods and the discontinuous output of place recognition. In order to give full play to their advantages, an integrated localization strategy is presented in this paper, where metric data such as visual odometry measurement, a digital map and topological data of place recognition results are incorporated. Place recognition assists initialization process and provides topological place estimation at all times. Gaussian-Gaussian Distribution is used for visual odometry raw measurement representation such that the errors of odometry is appropriately modelled. By comparing similarities between the digital map and odometry trajectories, we then use map-assisted approach to correct odometry estimation. Finally, a mutual check gives a criterion for judging whether metric and topological results are sufficiently consistent. Experiment results show that the integrated system outperforms subsystems with mean localization error at 2.9 m on our self-collected dataset with off-road scenarios.