基于视觉的室内移动机器人定位与导航关键技术研究

The popularization and application of service robots are subject to the high cost of laser sensors. In order to solve this problem, further improve the intelligence of robots and apply them in practice under complex indoor environment, there is an urgent need for the vision-based scheme of mobile robots with long-term continuous autonomous ability. This project focuses on the challenges of long-term autonomous navigation for mobile robots in complex indoor conditions such as dynamic and crowed environments. With visual sensors as core perception device, this project aims to study the fundamental principle and methodology of vision-based simultaneous localization and mapping (vSLAM), trajectory planning, and motion/behavior control. This project, which takes advantages of latest techniques on machine learning, localizability estimation, visual seroving, and so forth, contributes to fully excavating big data resources formed by long-term navigation, as well as establishing a long-term mechanism of continuous rolling optimization among perception, planning and control. The project dedicates to improving autonomy of mobile robots and both accuracy and robustness in localization and navigation. The project will eventually develop a mobile robot prototype and conduct experimental validation of its long-term localization and navigation capacity in the physical indoor environments. The validation results will be carefully assessed, based on which the prototype will be further optimized. In summary, vision-based localization and navigation techniques for mobile robots in the indoor environments enjoy a prosperous application prospect especially in the field of household, elderly assistance and logistics. With the help of these techniques, mobile robots would perform better in adaptability and autonomy when executing long-term work in the complex indoor environments.

目前服务机器人的推广应用受制于激光传感器高昂的成本，为了解决这一问题并进一步提高机器人智能化水平，使其在室内复杂环境中得到实际应用，迫切需要具备长期连续自主作业能力的移动机器人视觉解决方案。本项目针对移动机器人在室内动态、拥挤等复杂环境下长期自主导航所面临的技术挑战，以视觉传感器为核心感知手段，研究基于视觉的同步定位与建图、运动规划、行为与运动控制的基础理论与方法，利用机器学习、定位能力估计、视觉伺服等先进技术，充分挖掘长期导航所形成的丰富大数据资源，建立感知、规划和控制之间连续滚动优化的长期机制，着力于提高移动机器人的长期自主性以及定位与导航的精确性和鲁棒性。最终开发移动机器人实验样机，在真实室内环境中开展长期定位与导航测试，并进行评价和优化。上述关键技术在家居、助老、物流等领域具有广泛的应用前景，将显著提高移动机器人在室内复杂环境下的长期适用性和自主性。

随着移动式服务机器人应用领域的不断拓展，室内环境中的行人、家具、光照等复杂动态对机器人的感知和控制提出了挑战，迫切需要具备长期连续自主作业能力的移动机器人视觉解决方案。本项目围绕移动机器人在室内动态、拥挤等复杂环境下长期自主导航需求，以视觉传感器为核心感知手段，研究基于视觉的同步定位与建图、运动规划、行为与运动控制的基础理论与方法。.本项目基于机器学习、位姿估计、机器视觉的最新进展，充分挖掘长期导航形成的大数据资源，建立了一套面向移动机器人感知、规划与控制系统的长期持续滚动优化机制。在视觉建图与定位技术方面，提出了基于特征点贝叶斯存续性滤波的全局地图预测的长期视觉SLAM、动态场景下鲁棒增量式长期视觉拓扑定位、高效增量式的层次森林视觉惯性SLAM闭环检测算法，提升了复杂高动态环境和长期动态环境下的定位精确性和鲁棒性；在运动规划与自主导航技术方面，提出了基于时空地图预测的长期导航方法、拥挤行人中基于安全强化学习的无地图导航方法，以及异构多机器人系统的协作任务与路径规划算法，提升了复杂动态环境下的导航效率和安全性。基于智能轮椅和服务机器人在真实场景的实验结果表明了所提方法的有效性，展现了在居家服务、助老助残和物流机器人领域的广阔应用前景。