面向季节变化的机器人视觉全局定位

定位是移动机器人自主导航的关键技术。传统的定位技术主要以激光传感器为主，存在两方面问题，一方面，激光传感器的数据稀疏，鉴别信息不足，难以全局定位；另一方面，激光传感器成本较高，显著增加了机器人的价格。因此，视觉全局定位成为业界关注的解决方案。然而，自然环境下的视觉全局定位包含诸多挑战：当前图像与地图之间存在显著视角变化，海量地图图像间存在全局或局部相似混淆，自然环境变化导致光照甚至环境内容发生改变等。针对这些问题，本项目以机器人视觉全局定位为研究对象，以环境变化中时间跨度最长、变化因素最多的复杂长期季节变化为场景，以环境变化干扰下的不变信息挖掘为基本思路，模仿人在定位任务中表现出的感知和推理智能，重点突破具有主动关注机制的特征提取、具有多粒度信息表达的特征描述以及高离群点数据的高效假设检验等关键技术，从重复性、不变性和鲁棒性三个维度提升定位性能，最终提升机器人的自主性。

Localization is indispensable for robot navigation. Conventional robot localization solution depends on the laser, which has two problems: 1) global localization is hard due to the limited field of view and data resolution, 2) the cost of the robot with laser may not be accepted by the consumer market. Therefore, the visual global localization technique draws the attention from both industry and academia. There are several challenges to visual global localization in natural environment, including view point change, local and global ambiguity, and illuminative variation. In this project, we set to study the visual global localization across seasons, which contains significant changes covering most factors mentioned before. Motivated by the perception and inference skills performed by humans when localization, we try to discover and understand the invariance in changing environment. We propose an attention mechanism based feature detector, a multi-granular representation for feature description, and an efficient hypothesis validation method against high percentage of outliers, to improve the repeatability, stability and robustness of the localizer, finally bringing better autonomy to robots.