基于融合的全向三维视觉理论及在车道检测和定位中的应用研究

发布时间：2018-12-12 05:09

【摘要】：本文以国家自然国家自然科学基金支持的重大研究计划"视听觉信息的认知计算"为课题研究背景,将具有360°视野的全向相机系统与三维激光雷达相结合,构建全向三维视觉信息采集系统为无人驾驶车辆提供丰富的全向图像信息和精确的三维位置信息。研究了基于融合的全向三维视觉数据的车道检测与定位问题。取得的主要成果和创新点如下:1、研究了逆透视变换对车道线特征提取的影响,在透视图像和逆透视图像上测试三种基于全景图像提出的车道线特征提取算法的性能。在全景图像数据库上的实验结果表明,同种车道线特征提取方法在逆透视图像上的敏感性、精度、鲁棒性、潜在最佳性能、平均正确率等性能指标均优于其在透视图像上的性能。2、提出了一种基于畸变全景图像的车道检测算法,在全景图像空间进行曲线拟合,直接获得车道线在车体坐标系下的参数。实车实验证明了在具有畸变的全景图像上进行车道检测的可行性。3、提出了一种车道模型参数优化的车道模型拟合方法,用统一的车道模型拟合直线车道和曲线车道。根据抛物线车道模型将车道线特征图变换到线性空间,直线和曲线车道边缘在线性空间均用统一的直线进行描述。将车道模型参数拟合问题转化为参数优化问题,并通过最小化目标函数得到车道模型最优的参数。在公开数据库和实车采集数据上对提出车道检测算法进行测试,长距离实验结果表明基于车道模型参数优化车道检测算法准确率达到99%。4、提出了一种基于表观地图和三维特征地图的定位方法。定位过程为两步:基于表观地图的初步定位和初步定位结果指导下的基于三维特征地图的定位。在城市道路场景中测试提出的定位算法,实验结果表明基于位置采样创建的表观地图,定位效果优于基于特征距离采样的建图方法;而且在表观地图初步定位的基础上,基于三维特征地图的方法定位可以提高初步定位的精度。
[Abstract]:In this paper, a major research project supported by the National Natural Science Foundation, "Cognitive Computation of Audio-Visual Information", is used as the research background. The omnidirectional camera system with 360 掳field of vision is combined with 3D lidar. The omnidirectional 3D visual information acquisition system is constructed to provide abundant omnidirectional image information and accurate 3D position information for driverless vehicles. The problem of lane detection and location based on fusion of omnidirectional 3D vision data is studied. The main achievements and innovations are as follows: 1. The influence of inverse perspective transformation on lane line feature extraction is studied, and the performance of three lane feature extraction algorithms based on panoramic image is tested on perspective image and inverse perspective image. Experimental results on panoramic image database show that the same lane feature extraction method is sensitive, accurate, robust and potentially optimal in inverse perspective images. The average accuracy is better than its performance in perspective image. 2. A lane detection algorithm based on distorted panoramic image is proposed to fit the curve in panoramic image space. The parameters of the lane line in the car-body coordinate system are obtained directly. Real vehicle experiments prove the feasibility of lane detection on distorted panoramic images. 3. A lane model fitting method with optimized lane model parameters is proposed, and a uniform lane model is used to fit straight lane and curve lane. According to the parabolic lane model, the characteristic map of the lane line is transformed into linear space, and the line and the curve lane edge are all described in the linear space by the unified straight line. The parameter fitting problem of lane model is transformed into parameter optimization problem, and the optimal parameters of lane model are obtained by minimizing the objective function. The proposed lane detection algorithm is tested on the open database and real vehicle data collection. The results of long distance experiments show that the accuracy of lane detection algorithm based on lane model parameters is 99.4. A localization method based on apparent map and three-dimensional feature map is proposed. The localization process consists of two steps: the initial location based on the apparent map and the localization based on the 3D feature map under the guidance of the preliminary location results. The proposed location algorithm is tested in the urban road scene. The experimental results show that the location effect is better than that based on feature distance sampling. On the basis of the preliminary location of the apparent map, the method based on the 3D feature map can improve the accuracy of the initial location.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：U463.6;TP391.41

【相似文献】