无人车辆近年来得到快速发展，使用场景也不断扩大。随着被应用到军事、科考、乡村及野外运输等领域，无人车辆需要能够在崎岖的动态环境中稳定导航，但目前主流的城市道路和小范围封闭环境中的无人驾驶技术方案并不适用于此类环境。低线束激光雷达是目前最实用的无人车辆感知传感器，但崎岖路面导致的车辆颠簸会放大低线束激光雷达纵向视场角小、点云密度低的不足，因此需要对算法做针对性优化。本文对在崎岖动态环境中使用低线束激光雷达的无人车辆的自主导航的定位、感知和规划相关技术进行了研究，主要研究成果如下：首先，设计了可不依赖环境先验地图而将激光雷达与GNSS、IMU和轮速计融合的定位方法。先将GNSS输出的地理坐标转换为直角坐标，再使用FAST-LIO激光惯性SLAM方法处理雷达点云，并利用GNSS数据对FAST-LIO定位结果做坐标变换，最后通过扩展卡尔曼滤波算法在三维空间中融合4项数据得到高频的定位结果，显著提高了当GNSS精度下降时的系统定位精度。其次，提出了一种通过比较连续数据流信息检测动态物体的方法。首先分析相邻帧间的最近邻点，筛选出潜在的动态点；之后剔除其中的4种误判点，再经过区域生长和聚类得到动态物体。该法提高了复杂环境中动态物体检测的检出率和准确率，且不受动态物体种类的影响。进一步地，通过匹配动态物体的历史信息与当前信息，实现了多目标的跟踪，并采用卡尔曼滤波方法提高了动态物体状态估计的准确性和稳定性。再次，设计了基于高程图的静态环境中可行驶区域的分析方法，提出了路径-速度解耦的崎岖动态环境中无人车辆的运动规划方法。路径规划中提出了一种路面崎岖程度的评价指标并将该指标引入了路径搜索和优化过程，速度规划中采用碰撞检测分析动态物体的影响以适应大曲率路径，再在s-t图中进行二次规划获得最优的行驶速度。还提出了路径-速度迭代机制以更好地处理无人车辆与动态物体对向行驶的场景。最后，搭建起了实验用的无人车辆导航测试平台并开发了完整的导航程序，针对在崎岖动态环境中导航时可能出现的情况设计了对应的实验并进行测试，实验中车辆可以准确检测并稳定导航，验证了本文研究工作的可行性与有效性。
In recent years, unmanned vehicles have been rapidly developed, and their application scenarios are constantly expanding. With applications in in military field, scientific investigation, rural and field transportation, unmanned vehicles need to be able to navigate stably in uneven dynamic environments, but the current mainstream driverless technology solutions for urban roads and small enclosed environments are not suitable for such environments. Few-channel lidar is the most practical perception sensor for unmanned vehicle at present. However, vehicle bump caused by uneven road will amplify the disadvantages of small longitudinal field of view and low point cloud density of few-channel lidar, so it is necessary to optimize the algorithms. In this paper, the localization, perception and planning technologies of autonomous navigation of unmanned vehicles using few-channel lidar in uneven dynamic environment are studied. The main research results are as follows:Firstly, a positioning method which is able to fuse lidar with GNSS, IMU and wheel encoder without relying on environment prior map is designed. The geographical coordinates outputted by GNSS are converted to cartesian coordinates at first, then the point cloud of lidar is processed by a lidar-inertial SLAM method called FAST-LIO, and the coordinates of the localization results of FAST-LIO are transformed with the GNSS data. Finally, the extended Kalman filter algorithm is used to fuse the four kinds of data in three-dimensional space to produce high frequency localization results. This method improves the system localization accuracy significantly when the GNSS accuracy decreases.Secondly, a dynamic object detection method based on comparing continuous dataflow is proposed. At first, the nearest neighbor points between adjacent frames are analyzed to screen potential dynamic points. Then, four kinds of misjudged points are eliminated, and dynamic objects are obtained through regional growth and clustering. This method improves the detection rate and accuracy of dynamic object detection in complex environment and is not affected by the type of dynamic object. Furthermore, the multi-object tracking is realized by matching the historical information and the current information of dynamic object, and the accuracy and stability of the state estimation of dynamic object are improved by Kalman filtering.Thirdly, an analysis method of drivable area in static environment based on elevation map is designed, and a path-speed decoupling motion planning method of unmanned vehicle in uneven dynamic environment is proposed. In path planning, an evaluation indicator of road roughness is proposed and introduced into path search and optimization process. In speed planning, collision detection is used to analyze the influence of dynamic object to adapt to large curvature path. Then, the optimal traveling speed is obtained by quadratic programming in s-t diagram. A path-speed iteration mechanism is also proposed to deal with the scenario of unmanned vehicle and dynamic object moving in opposite direction better.Finally, an unmanned vehicle navigation test platform for experiments is built, and a complete navigation program is developed. Corresponding experiments are designed and tested for various situations that may occur when unmanned vehicles navigate in uneven dynamic environment. In the experiments, the vehicle detects precisely and navigates stably, which verifies the feasibility and effectiveness of the research work in this paper.
