Road Scenarios Modeling And Accurate Localization For Intelligent Vehicles

High-accuracy localization is one of the core problems for intelligent vehicles.This dissertation presents a “representation-to-localization” strategy for intelligent vehicle localization.Road scenarios are first represented and modeled from the collected vehicle-borne multi-sensor data.We then developed different methods for intelligent vehicle localization by referring the pre-built road models.The proposed methods have been tested in different routes with or without GPS,times,weathers,seasons,etc.In summary,the contributions of this dissertation are listed as follow:First,a road scenarios representation model is built from the data captured by monocular/binocular cameras,DGPS,IMU,GPS receiver and laser scanner,etc.,for intelligent vehicle localization.With this model,road scenarios consist of a sequence of nodes.For each node,there are visual features,3D data,and trajectories.Among them,visual features are composed of both holistic and local features for scene matching purpose.3D data are derived in three ways: monocular 3D reconstruction,binocular 3D reconstruction,and LIDAR-based 3D reconstruction.3D data and the image features are corresponded with accurate calibration between camera and laser rangefinder(LRF).Moreover,vehicle trajectories,which describe the pose of the vehicle for each node,are computed from 3D data.These three elements can represent the uniqueness of each node.In practice,all road scenarios models were built with prototype intelligent vehicles.Second,a multi-scale localization method is proposed based on road scenarios model from monocular vision and ordinary GPS.The model follows a "coarse-to-fine" strategy.In coarse localization step,GPS data are matched within the nodes in the road model to derive initial node candidates.In node-level localization step,both holistic and local features are extracted and matched with those from the initiate candidate nodes,respectively.The matching results are fused by KNN-MFS(K-nearest neighbors-multiple feature spaces)to obtain a unique candidate node.In pose-level localization step,the local features are matched with the 3D data of the candidate to further refine the localization results by solving a perspective-n-point(PnP)problem.The method allows us to derive a “closest” node in the road model such that the pose of intelligent vehicle is obtained for accurate localization.Third,a vision-only localization method without GPS data is proposed.In this method,we utilize a topological model rather than GPS to derive coarse localization results.Specifically,“feature-zone” is first defined.Then,a topological localization approach based on feature-zone is developed,with which initial localization range is obtained.Furthermore,a Bayesian topological localization method based on vision-motion is then proposed to further narrow down the localization space.This method addresses the localization problem in GPS blind areas.Fourth,a multi-view site matching method is proposed for intelligent vehicles by fusing both the front and downward views.In this method,a double topological model is built from both views for coarse localization.In node-level localization,site matching is accomplished by matching the visual features computed from both views within those in the road scenarios models.In pose-level localization,vehicle poses are finally computed by referring to the 2D pavement structure rather than the 3D road scenes to further enhance the localization results.The proposed method can improve both the localization accuracy and the robustness,especially in varying illumination environments.The researches in this dissertation suggest low-cost solutions to accurate localization of intelligent vehicles,which can be significant to promote the development of intelligent vehicles.In the future,the researches will focus on full-view localization and extend the application to early evening and heavy weather.