面向智能车定位的道路环境视觉地图构建

为实现智能车视觉定位，提出了一种基于多视角、多维度道路环境表征的高精度视觉地图构建方法，该方法明确了视觉地图的表征模型，包括视觉特征、场景结构信息以及轨迹信息等。在视觉特征中，运用前视场景全局特征描述道路环境，视觉特征不局限于某一种特征描述子；在场景结构信息中，运用俯视路面的2D结构信息进行描述，该特征与前视视觉特征构成多视角；轨迹信息则通过视觉维度以及地理维度的多维度方式完成计算，在视觉维度中，通过平面单应性计算节点间的轨迹；地理维度中，通过高精度经纬度信息消除累积误差问题。试验选取武汉理工大学内长约700 m的半开放式环形路段进行试验。试验结果表明：制图的单节点平均误差为3.1 cm，标准差为2.3 cm，最大节点误差为9.3 cm，累积误差率为0.5%。运用所制地图进行定位检测，平均定位误差约为11.8 cm，因此，研究所提出的方法可应用于半开放式路段或固定场景的视觉地图构建，为实现智能车在上述场景的定位打下基础。同时，研究提出的制图方法不需使用双目摄像机，在降低数据存储量以及制图成本的前提下，实现了对道路环境的充分表征；此外，运用路面2D特征结构信息计算轨迹，解决了视觉3D重建精度不稳定的问题，为视觉地图构建提供了新的构建思路。

Construction of a Visual Map Based on Road Scenarios for Intelligent Vehicle Localization

Current methods for intelligent vehicle visual localization are limited in terms of expense and validity, therefore a method based on constructing high-accuracy visual maps using multi-scale road scenarios is proposed. In order to model a representative visual map, various important parameters including visual features, structure information and trajectories were included. In visual features, road scenarios are represented by holistic features in forward-view; however visual feature can be obtained in several ways. The structure information is represented by 2D features in downward-view. Both visual features and structural information incorporate multi-view. Trajectories are computed by multi-scale visual and geographical dimensions. At the visual scale, trajectories are computed by homography, while at the geographical scale, high-accuracy position data substantially reduce accumulative errors. To evaluate the method, a semi-open ring route with about 700 m of road with a university campus was selected. The experimental results demonstrate that the mean error of a node is 3.1 cm and the standard deviation is 2.3 cm. The maximum error is 9.3 cm and the cumulative error rate is 0.5%. The localization error of intelligent vehicles is 11.8 cm. Hence, the proposed method can be applied to semi-open sections or fixed-road scenarios for basic intelligent vehicle localization. The proposed method is beneficial as it low cost when applied to road scenarios and does not require specialized cameras. Moreover, the use of 2D data can address the limited robustness of 3D reconstruction. The proposed methodology provides a novel approach for visual map construction for intelligent vehicle localization.