Localizability Estimation for Autonomous Driving: A Deep Learning-Based Place Recognition Approach

In recent years, research and development aimed at the societal implementation of autonomous driving have attracted increasing attention. Localization, which involves obtaining information about the surrounding environment from sensor data and estimating the vehicle's position, is necessary for realizing autonomous driving. Localization based on 3D Light detection and ranging (LiDARs) and high definition (HD) maps, called map-matching, gained more attention in recent years because of its accuracy and wide availability of HD maps. In this method, the input scan from LiDAR is matched with HD map features to obtain the accurate position of the ego vehicle. However, in this method, the localization accuracy may decrease when the surrounding area does not have distinctive features. In this study, we proposed a method based on deep learning to estimate localization accuracy for autonomous driving. In particular, the input scan point cloud is fed to the deep neural network (DNN) inspired by MinkLoc3D-SI, which is used for place recognition problems, and then proceeds with the regression network, to estimate the localization accuracy. The estimated error from the proposed method can be used in the fusion process with other sensors to increase the overall localization accuracy. For training, We created a dataset using an open-source autonomous driving simulator AWSIM. Simulation results show that in more than 94% of cases, the localization error can be estimated with less than 0.1m accuracy.