<p>自动驾驶作为智能车发展的最终方向，它的安全行驶主要依赖于环境感知系统对周围环境的理解。在现实环境中，车辆所处的环境除了结构特征显著的结构化道路，还存在许多结构特征不明显的非结构化道路区域。目前，对于结构化道路提取的研究已日趋成熟，非结构化道路提取已逐渐成为智能车环境感知技术的研究热点。随着激光雷达技术的快速发展，以其测距精度高，响应速度快等诸多优点受到科研人员的青睐，不仅可以快速获取地物表面的高精度三维坐标，还可以获取反射强度与纹理等信息，在自动驾驶研究中得到广泛的应用。鉴于以上这些优点，本文将使用激光雷达作为主要传感器，研究基于三维激光点云的非结构化道路提取方法。</p> <p>本文的主要研究内容和成果如下：</p> <p>（1）多特征约束提取非结构化道路方法研究。相关研究表明，基于高程特征提取地面点时，直通滤波算法对于地面较为平坦的道路点云提取效果较好，对于高低起伏的地面提取效果并不理想，而布料模拟滤波算法适用于多种类型道路且提取效果较佳，但提取的地面点中仍存在许多低矮地物点云。针对这一问题，本文通过优化道路提取流程，基于点云强度信息从地面点中提取出初始非结构化道路，该方法可以去除大部分地物点，使道路边界清晰，提高了提取道路点云的准确性。</p> <p>（2）基于双边滤波与倾斜度信息的滤波方法研究。实验研究表明，多特征约束提取的非结构化道路，存在与道路连通特征相似的非道路点云问题。由于双边滤波后道路点云形成一种道路边缘倾斜度值较大，道路内部倾斜度值较小的特征，本文提出了倾斜度滤波算法。该方法以邻近点向量之间所成的水平夹角作为倾斜度滤波阈值剔除非道路点，弥补了双边滤波算法的不足，提高了最终道路点云的准确性。</p> <p>（3）探讨倾斜度滤波算法可行性研究。通过利用本文提出的倾斜度滤波算法对不同类型的非结构化道路提取进行实验分析，倾斜度滤波后提取的最终直路点云准确性为97.91%，在双边滤波的基础上提升了6.34%；检测质量为96.61%，较双边滤波提升了5.98%；倾斜度滤波后提取的最终弯路点云准确性为98.88%，在双边滤波的基础上提升了4.15%；检测质量为95.45%，较双边滤波算法提升了3.53%，从而验证了本文算法的准确性与可行性。</p>

<p>As the ultimate direction of the development of intelligent vehicles, autonomous driving mainly relies on the understanding of the surrounding environment by the environment awareness system. In the real environment, in addition to the structurally significant road, there are also many unstructured road areas where the structure characteristics are not obvious. At present, the research on structured road extraction has become increasingly mature, and unstructured road extraction has gradually become a research hotspot of intelligent vehicle environment awareness technology. With the rapid development of lidar technology, it is favored by researchers for its high ranging accuracy, fast response speed and many other advantages. It can not only quickly obtain the high-precision Three dimensional coordinates of the surface of the ground object, but also obtain the reflection intensity and texture information, which has been widely used in the research of automatic driving. In view of these advantages, this paper will use lidar as the main sensor to study the unstructured road extraction method based on 3d laser point cloud.</p> <p>The main research contents and achievements of this paper are as follows:</p> <p>(1) Research on unstructured road extraction with multi-feature constraints. Studies show that the feature extraction based on elevation point on the ground, through filtering algorithm for the ground is relatively flat road point cloud extracting effect is better, for the extracting effect is not ideal, the ground of the ups and downs and cloth simulation filtering algorithm is suitable for a variety of types of roads and extracting effect is better, yet there still exists many low ground points of the extract feature point cloud. To solve this problem, this paper extracted initial unstructured roads from ground points based on point cloud intensity information by optimizing the road extraction process. This method can remove most ground points, make the road boundary clear, and improve the accuracy of road point cloud extraction.</p> <p>(2) Research on filtering method based on bilateral filtering and inclination information. The experimental results show that the unstructured road extracted by multi-feature constraints has the problem of non-road point cloud similar to the road connected feature. Since the road point cloud after bilateral filtering forms a characteristic of large slope value at the edge of the road and small slope value at the inside of the road, this paper proposes a tilt filtering algorithm. In this method, the horizontal Angle between adjacent point vectors is used as tilt filtering threshold to eliminate non-road points, which makes up for the deficiency of bilateral filtering algorithm and improves the accuracy of the final road point cloud.</p> <p>(3) The feasibility study of tilt filtering algorithm is discussed. Through the experimental analysis of different types of unstructured road extraction by using the tilt filtering algorithm proposed in this paper, the accuracy of the final straight road point cloud extracted after the tilt filtering is 97.91%, which is 6.34% higher than that of bilateral filtering. The detection quality was 96.61%, 5.98% higher than bilateral filtering. The accuracy of the final curve point cloud extracted after tilt filtering is 98.88%, which is 4.15% higher than that of bilateral filtering. The detection quality is 95.45%, 3.53% higher than that of bilateral filtering algorithm, thus verifying the accuracy and feasibility of the proposed algorithm.</p>