基于三维数据面向无人车导航的非结构化场景理解

发布时间：2018-07-04 20:58

本文选题：无人车 + 三维数据　；参考：《浙江大学》2014年博士论文

【摘要】：无人车对行驶环境的感知和理解是保证其后续行为规划正确的前提和基础,也是人工智能领域一直以来具有挑战性的课题之一。通过车载三维传感器,感知行驶环境,具有简便、信息量丰富、探测精度高等优点,尤其是64线激光雷达和双目立体视觉传感器,被越来越多的无人车三维感知系统所采用。三维信息与环境相关性强,具有很大的随机性。与结构化场景相比,室外非结构化场景中无特定的人工标记,物体的三维外形复杂多变,并且室外颠簸的路况,杂乱的植被也会给场景检测理解带来额外的困难。因此,研究无人车对非结构化场景三维信息特征模型的构建和处理,对扩展无人车系统的活动范围、提高它对复杂和恶劣地形环境的适应能力具有非常重要的意义。本文基于三维传感器,研究了面向无人车的典型非结构化场景理解问题,主要内容及贡献如下：提出了64线激光雷达内参标定的方法以及外参自标定的方法。通过分析激光发射系统的数学模型,放弃64线激光雷达内参中距离误差补偿常量的通常做法,实现一个与反射距离成线性关系的变量的更为精密的补偿。以不同距离的墙面为目标物,对64线激光雷达内参进行非线性优化。对于64线激光雷达外参,将传统实施复杂的一次性标定过程分解为两部分,实现了操作更为简单的自标定算法。第一步通过雷达坐标系中的地面与实际地平面比对,求解出激光雷达相对于车体的俯仰角和侧滚角；第二步让无人车在具有柱状物体(路灯杆等)的城市道路中行驶,匹配前后帧中柱状物体在雷达坐标系中航向角的以及平移的变化,与GPS记录的车体航向角及平移变化做比对,求解出激光雷达相对于车体坐标系的航向角以及平移。合并两步骤求得的参数就能获得准确的激光雷达外参。针对乡村环境的感知理解。提出了基于马尔科夫随机场的路面分割算法。相比于栅格模式的道路分割,基于图的道路分割检测距离更远、精度更高、更节省内存资源；而传统的基于图的道路分割算法,仅利用单个三维点的特征,极易受到噪声干扰,只能适用于平坦的城市道路而非颠簸起伏的乡村环境道路。算法利用激光雷达扫描线在xoy平面上的几何特性以及激光雷达原始数据结构的特点,将扫描线分割为线段,通过分析线段的多个特征,构建线段属性的概率函数,并且通过马尔可夫随机场和图割的方式优化最终分割的结果。通过实际数据序列实验结果与现有算法结果的量化比较,表明本算法在乡村环境道路情况下不仅道路分割精度高,而且稳定性好。经测试,本算法能够在无人车平台上实时运行。针对植被散布环境,提出了基于多传感器融合的的障碍物分类方法。首先将三维激光雷达数据、二维彩色相机数据融合,进行稀疏深度指导的超像素分割以及深度恢复算法,在迭代过程中,让二维图像分割的结果和三维深度数据相互指导,提高算法的性能。然后提取局部的区域特征,包括三维点云局部统计特征、激光雷达强度特征,利用近红外相机与彩色相机融合的归一化差分植被指数等特征,组成多维的特征向量,采用有监督的学习方式训练SVM分类器。在现实环境中的测试表明,多传感器融合的分类方法提高了分类的准确性,增加了远距离物体分类的鲁棒性。针对未知复杂地形的感知理解,提出了一个从三维点云输入到栅格通行性输出的运算框架。在该框架下可以完成点云数据的滤波、多帧拼接、帧间匹配、点云恢复、点云数据栅格化、通行性分析以及障碍物聚类。首先针对无人车立体视觉系统三维重建后产生的噪声,分别采用基于点云局部密度的算法去除外点,采用双边带滤波器减小点云中的随机噪声。其次针对数据帧间匹配,采用经KD-Tree的数据结构加速计算的ICP进行点云匹配。针对点云恢复,采用较为简单的反距离平方和插值算法。在通行性分析中,以无人车车体大小为窗口,计算中心点的属性：若窗口内有任意两栅格不仅阶跃值大于阈值,而且栅格间梯度也大于坡度阈值,则标记状况中心栅格属性为阶跃障碍；将窗口内所有栅格值用于拟合平面,计算坡度和粗糙度,根据通行性式计算通行代价。
[Abstract]:The perception and understanding of the driving environment by unmanned vehicles is the premise and foundation to ensure the correct follow-up behavior planning. It is also one of the challenging topics in the field of artificial intelligence. It is simple, rich in information and high detection precision through the three-dimensional sensor in vehicle, which has the advantages of simple, convenient, abundant information and high detection precision, especially 64 line laser radar and double. The stereoscopic vision sensor has been adopted by more and more unmanned vehicle 3D sensing systems. The three-dimensional information is strongly correlated with the environment and has great randomness. Compared with the structured scene, there are no specific artificial markers in the outdoor unstructured scene, the three-dimensional shape of the object is complicated and more complex, and the outdoors turbulence and the chaotic vegetation will also be found. It brings additional difficulties to the detection and understanding of the scene. Therefore, it is very important to study the construction and processing of the unstructured 3D information feature model for unstructured scenes, to extend the range of the unmanned vehicle system and to improve its adaptability to the complex and abominable terrain environment. The main contents and contributions of the typical unstructured scene understanding of human cars are as follows:
The calibration method of 64 line laser radar and the method of self calibration are proposed. By analyzing the mathematical model of the laser emission system and giving up the usual practice of compensating the constant of the distance error in the internal reference of the 64 line laser radar, a more precise compensation is realized for a variable with the linear relation of the reflection distance. The target, nonlinear optimization of the 64 line laser radar internal parameter. For the 64 line laser radar external parameter, the traditional implementation of the complicated one-time calibration process is decomposed into two parts, and a more simple self calibration algorithm is realized. The first step is to solve the laser radar relative to the vehicle by comparing the ground surface to the actual surface surface in the radar coordinate system. The pitch angle and the side roll angle of the body; the second step makes the unmanned vehicle run in the city road with a columnar object (the street lamp pole, etc.), matches the change of the navigation angle and the shift of the column shaped objects in the radar coordinate system before and after the frame, and compares the direction angle and the shift change of the car body recorded by the GPS, and solves the relative system of the laser radar relative to the car body coordinate system. The parameters obtained by merging the two steps can get accurate laser radar extrinsic parameters.
In view of the perception and understanding of the rural environment, a road segmentation algorithm based on Markov random field is proposed. Compared with the grid mode, the road segmentation based on the graph is far farther, more accurate, and more memory saving, while the traditional road segmentation algorithm based on graph only uses the characteristics of a single three-dimensional point, and is very easy to be subjected to. Noise interference can only be applied to a flat urban road rather than a bumpy country road. The algorithm uses the geometric characteristics of the laser radar scanning line on the xoy plane and the characteristics of the original data structure of the laser radar. The scanning line is divided into line segments, and the probability function of the line segment attribute is constructed by analyzing the multiple features of the line segment. The results of the final segmentation are optimized by Markov random field and graph cut. The results of the actual data sequence experiment and the existing algorithm results show that the algorithm not only has high road segmentation precision, but also has good stability. The algorithm can run on the unmanned vehicle platform in real time after testing.
An obstacle classification method based on multi-sensor fusion is proposed for the vegetation distribution environment. Firstly, the 3D laser radar data and two-dimensional color camera data are fused to carry out the super pixel segmentation and depth recovery algorithm guided by the sparse depth. In the iterative process, the results of the two dimensional image segmentation and the three-dimensional depth data are interacted with each other. To improve the performance of the algorithm, the local regional features are extracted, including the local statistical features of the three dimensional point cloud, the intensity feature of the laser radar, the feature vectors of the normalized difference vegetation index, such as the fusion of the near infrared camera and the color camera, and the multi-dimensional feature vectors, and the supervised learning method is used to train the SVM classifier. The test shows that the classification method of multi-sensor fusion improves the accuracy of classification and increases the robustness of remote object classification.
In view of the perception and understanding of unknown complex terrain, a framework of input from three dimensional point cloud to grid pass is proposed. Under this framework, the filtering of point cloud data, multi frame splicing, inter frame matching, point cloud recovery, point cloud data grid, traffic analysis and obstacle clustering can be completed. The noise generated by the three dimensional reconstruction is used to remove the external points based on the local density based algorithm of point cloud, and the random noise of Dian Yunzhong is reduced by a bilateral band filter. Secondly, the point cloud is matched by the KD-Tree data structure accelerated ICP for the matching of data frames. In the traffic analysis, the attribute of the center point is calculated with the size of the car body of unmanned vehicle as the window. If there is any two grid in the window not only the step value is greater than the threshold value, but also the gradient of the grid is greater than the gradient threshold, then the grid attribute of the mark state center is a step obstacle, and all the grid values within the window are used to fit the plane. Calculate the slope and roughness and calculate the passage cost according to the traffic pattern.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：U495;TP242

【参考文献】