【计算机科学】【2019】基于视觉SLAM和对象检测的多目标映射与路径规划

本文为加拿大滑铁卢大学（作者：Ami Woo）的硕士论文，共104页。

自主机器人的路径规划是移动机器人智能穿越环境所需的关键任务之一。机器人路径通常是通过利用在一定时间内可访问的地图来规划的，例如最小化旅行距离或时间。本文提出了一种采用图优化和目标检测的基于同时定位和映射（SLAM）多目标路径规划方法。该方法的目的不仅在于找到一条使行驶距离最小化的路径，而且还在于最小化路径中的障碍物数量。本文以视觉SLAM（VSLAM）为基础生成全局路径规划图。VSLAM产生轨迹网络，通常以备用图的形式（如果基于度量）或相对于机器人里程碑估计的概率关系。并行地运行对象检测算法，以提供由VSLAM生成的轨迹网络图的附加信息，并将此用于多目标路径规划。VSLAM、目标检测和路径规划领域通常是独立研究的，但是本文将这些领域联系起来解决多目标路径规划问题。

论文的第一部分介绍了利用VSLAM和目标检测生成轨迹网络图的联系和方法。当VSLAM中需要一个新的关键帧时，将节点插入到图中。节点之间的移动距离是最小化的第一个标准，并在遍历时进行计算。与VSLAM相并行，目标检测组件量化节点之间检测到的对象数量。本文只对需要检测的预训练对象进行量化，训练对象为轿车和卡车。对象数是添加到图形中的两个附加边缘信息。随后，本文提出了基于生成图的多目标路径规划方法。图上路径规划的目标不仅是最小化行驶距离，而且最小化通过的汽车和卡车数量。所提出的设计使用KITTI数据集进行测试，该数据集专门用于自主驾驶，并由许多汽车和卡车组成。该设计不仅限于自动驾驶应用，还可以应用于其他领域，如监视、救援等，以及更多不同目标检测的场景。

Path planning of the autonomous robots isone of the crucial tasks that need to be achieved for mobile robots to navigatethrough the environment intelligently. The robot paths are typically plannedutilizing map that is accessible at the time with a certain optimizationobjective such as to minimizing the travel distance, or time. This thesisproposes a multi-objective path planning approach by integrating SimultaneousLocalization And Mapping (SLAM) with a graph based optimization approach and anobject detection algorithm. The proposed approach aims not only to find a paththat minimizes travel distance but also to minimize the number of obstacles inthe path to be followed. This thesis uses Visual SLAM (VSLAM) as the basis togenerate graphs for global path planning. VSLAM generates a trajectory networkwhich is usually in the form of a spare graph (if odometry based) orprobabilistic relations on landmark estimates relative to the robot. An objectdetection algorithm is run in parallel to provide additional information ontrajectory network graphs generated by the VSLAM, to be used in multi-objectivepath planning. The VSLAM, object detection, and path planning fields aretypically studied independently, but this thesis links the these fields tosolve the multi-objective path planning problem. The first part of the thesispresents the connections and methodology on using the VSLAM and objectdetection to generate trajectory network graphs. The nodes are inserted to thegraph when a new keyframe is needed in VSLAM. The distance travelled betweenthe nodes is the first criterion to minimize and is computed while traversing.In parallel to VSLAM, the object detection component quantifies the number ofobjects detected between the nodes. Only the pre-trained objects to detect arequantified and the trained objects in the thesis are cars and trucks. Thenumber of objects are the two additional edge information added to the graph.Later in the thesis, the multi-objective path planning on the generated graphsis presented. The objective of path planning on graph is not just on minimizingthe distance to travel but also on minimizing the number of cars and trucks itpasses. The proposed design is tested using KITTI dataset which is specializedfor autonomous driving and consists of many cars and trucks. The design is notlimited to autonomous driving applications, but can be applied to other fieldssuch as surveillance, rescuing, and many more with different objects to detect.

1 引言
2 项目背景与文献回顾
3 总体设计
4 VSLAM设计
5 对象检测
6 基于位姿SLAM的多目标路径规划
7 仿真与实验测试
8 结论与未来工作展望

更多精彩文章请关注公众号：