Abstract:

Vehicle self- localization is one of the core tasks for intelligent vehicles. Basically, localization error of GPS is about 10 m, which cannot meet the localization requirement for intelligent vehicles. Moreover, the high cost makes INS (Inertial Navigation System) not practical for intelligent vehicles. This paper proposes an accurate localization method for intelligent vehicles based on GPS and image fusion from visual map. The method aims to find the pose of current position to the nearest data collection point of visual map. Firstly, coarse localization is done by GPS data matching such that several candidate positions are selected from visual map. Furthermore, holistic feature matching is applied to compute one candidate from GPS matching results. Finally, vehicle pose is computed by matching local features and solving Perspectiven- Point (PnP) problem. Localization results are further refined with the computed vehicle poses. Experiments are made in a 5 km- route roadway, which are in different weather conditions and different intelligent vehicles. Experiment results show that the mean error of localization accuracy is about 11.6 cm and the max error is 37 cm. In addition, the proposed method has good robustness to different weather conditions. The proposed method suggests a low-cost and accurate localization solution for intelligent vehicles.

Key words: intelligent transportation, vehicle localization, data fusion, intelligent vehicle, computer vision

摘要：

车辆自定位是实现智能车辆环境感知的核心问题之一.全球定位系统(Global Positioning System, GPS)定位误差通常在10 m左右，不能满足智能车辆的定位需求；惯性导航系统成本较高，不适于智能车辆的推广.本文在视觉地图基础上，提出一种基于GPS 与图像融合的智能车辆定位算法.该算法以计算当前位置距离视觉地图中最近一个数据采集点的位姿为目标，首先运用GPS信息进行初定位，在视觉地图中选取若干采集点作为初步候选，其次运用Oriented FAST and Rotated BRIEF (ORB)全局特征进行特征匹配，得到一个候选定位结果，最后通过待检测图像中的局部特征点与候选定位结果中的三维局部特征点建立透视n 点模型(Perspective-n-Point, PnP)，得到车辆当前的位姿，并以此对候选定位结果进行修正，得到最终定位结果.实验在长为5 km的路段中进行，并在不同天气及不同智能车辆平台测试.经验证，平均定位精度为11.6 cm，最大定位误差为37 cm，同时对不同天气具有较强鲁棒性.该算法满足了智能车定位需求，且大幅降低了高精度定位成本.

关键词: 智能交通, 车辆定位, 数据融合, 智能车辆, 计算机视觉