Feature Based Object-Related Navigation

本研究项目的目标是开发一个与对象相关的导航系统。该方法的目标不是遵循之前建立的地图中的全局轨迹，而是相对于感知到的物体进行导航（“在第三个十字路口左转，经过一棵树，在右手边的第二所房子前停下来”）。此外，通过将目标特征（显著边缘，纹理等）纳入导航，我们打算增加测量和模型误差方面的鲁棒性，与最先进的方法相比，这些方法通常只使用带有人工参考点的抽象边界盒模型。通过将导航与感知对象相结合，我们的目标一方面是对不可预见事件（如在相邻车道上行驶的车辆突然变道）的快速反应，另一方面是计划的生成，如“跟随车道，超车，……”，以降低计划生成和更改的复杂性，特别是在高动态环境中。我们打算用我们的自动驾驶汽车MuCAR-3来展示我们的研究成果。基于现有的物体检测和分类方法，激光雷达和相机数据，我们计划展示新的导航方法如何应对现实世界中的不同场景，如超车机动或在拥挤的十字路口转弯，以达到预定义的目标，如：“跟随物体X”。发生问题的情况，如道路堵塞，必须自主解决。在进一步的研究步骤中，车辆将探索未知地形，创建拓扑对象相关地图，分类对象作为其节点。但是，该系统的目标不是绘制整个环境，而是仅对与自主导航相关的对象进行理解。车辆打算在这个地图中与对象相关的意义上朝着用户定义的目标行驶。为了显示该方法的效率，我们忽略了GPS或任何全局度量表示的使用。

Goal of this research project is the development of a system for object-related navigation. Instead of following global trajectories in a previously built up map, the aim of the approach is to navigate relative to perceived objects ("Turn left on the third crossroad, pass a tree and stop before the second house on the right hand side"). Further, by incorporating object features (salient edges, textures, etc.) into navigation, we intend to increase the robustness with respect to measurement and model errors, in comparison to state-of-the-art approaches, that often utilize abstract bounding box models with artificial reference points only. By coupling navigation to perceived objects we aim on the one hand at quick reactions to unforseen events, like sudden lane changes of vehicles that drive on the neighbor lane, on the other hand at the generation of plans like: "Follow lane, overtake vecicle, ..." to lower the complexity of plan generation and alteration, especially in high dynamic environments.We intend to demonstrate the results of our research with our autonomous vehicle MuCAR-3. Based on already existing methods for object detection and classification in both, LIDAR and camera data, we plan to show how the new navigation approach copes with different scenarios in the real world, like overtaking maneuvers or turns on heavy crowded crossroads, to reach a predefined goal, like: "Follow object X". Occuring problem situations, like road blockages, have to be resolved autonomously. In a further research step, the vehicle shall explore unknown terrain to create a topological object-related map, with classified objects being its nodes. But instead of mapping the whole environment, the system aims at the comprehension of objects that are relevant for autonomous navigation only. The vehicle is intended to drive within this map in an object-related sense towards a user defined goal. To show the approach's efficiency we neglect the use of GPS or any global metric representation.