Sensor fusion for autonomous robotics in unstructured environments

实现了非城市环境下自主移动的系统的地形分类、导航和地图绘制。这些方法将适用于来自3D激光扫描仪，位置和里程传感器，GPS和彩色摄像机的数据。通过融合传感器数据，将附加信息集成到激光测量中并合并以创建语义3D地图。该方法将能够处理非结构化地形中的许多不确定性，而不需要地图或路线信息。因此，必须尽可能准确地分析和解释地形，以规划最佳路线。检测到的障碍物分为静态或动态对象，如森林、道路或车辆，并在3D地图中输入语义。因此，自主车辆可以在未知地形中决定和规划安全导航的路径。所描述的方法不同于其他方法，因为实现了用于导航和创建实时语义3D地图的完整融合和传感器集成。现有项目的第一个扩展是为地形分类创建语义马尔可夫随机场。这将使我们能够将机器人环境中检测到的物体整合到分类过程中。例如，这包括在道路或可驾驶表面上驾驶车辆的知识，这些知识可以用于在发现其他动态障碍物的区域中调整分类。第二个扩展是开发用于从3D激光扫描创建地图的过程。开发了新的方法，允许机器人在更长的时间内保持现有的地图，并扩展，即使机器人已经断电并重新启动。此外，将开发和测试一个新的想法，以提高映射的鲁棒性，该映射保持了关于轨迹过程的几个根本不同的假设。尤其是现在，不意味着被认为只是一个可能的轨迹。因此，可以做出关于数据关联的更好的决策。

Terrain classification, navigation, and mapping for autonomous mobile system will be realized for the non-urban environments. The methods will work on data from 3D laser scanners, position and odometry sensors, GPS, and color cameras. By fusion of sensor data, additional information is integrated into the laser measurements and merged to create a semantic 3D map. The approach will be able to deal with a number of uncertainties in unstructured terrain, without need of map or route information. Therefore, the terrain must be analyzed and interpreted as exactly as possible in order to plan the best possible route.Detected obstacles are classified as static or dynamic objects, such as a forest, road or vehicle, and semantics as entered in the 3D map. Thus the autonomous vehicle can decide and plan a path for safe navigation in unknown terrain. The approach described is different from others as a complete fusion and sensor integration for navigation and creating a real-time semantic 3D map is implemented. The first extension to the existing projects provides for the creation of a semantic Markov random field for terrain classification. This will enable us to integrate detected objects in the environment of the robot into the classification process.This includes for example the knowledge, to drive a vehicles on roads or drivable surfaces as well as possible, which can be used for adapting the classification in areas in which other dynamic obstacles are found.The second extension is the process developed to create maps from 3D laser scans. New methods are developed, which allows the robot to maintain an existing map over a longer period of time, and to expand, even if the robot has been powered down and restarted. In addition, a new idea will be developed and tested to improve the robustness of the mapping, which maintains several fundamentally different hypotheses about the course of the trajectory. Especially not now not meant to be considered only one probable trajectory. Thus, it is possible to make better decisions about the data association.