室内复杂光照环境视觉定位数据集构建研究

Data Set Construction in the Complex Illumination Environment for Visual Localization

摘要: 在具有复杂光照特征的室内环境中，由于缺少公开统一的视觉定位算法验证基准，机器人的视觉定位任务难以进行验证和评估。因此，为了打破数据拥有者和使用者之间的鸿沟，解决面向复杂光照的视觉数据的获取成本高、更新慢的问题，建立可持续更新的机器人视觉定位数据集是十分重要的。本文聚焦复杂光照的室内环境条件，探索标准定位数据集组织架构和数据集构建流程，并根据环境特点进行硬件选型和系统构建，构建了具有复杂照明特点的视觉定位数据集，而且基于本数据集评测了多种视觉定位算法。本文为所有的视觉序列提供真实地面位置，精度达到厘米级，可以用于测试视觉定位算法处理光照变化的能力。本文提出的构建流程可以为其他场景特点数据集的构建提供可迁移的经验，本文提出的数据集可服务于视觉定位验证任务，实现了现实场景和视觉定位算法提升之间的衔接，能给复杂光照环境下的机器人视觉定位系统提供更多的度量方法。

Abstract: In the indoor environment with complex illumination characteristics, it is difficult to verify and evaluate the visual localization tasks of robots due to the lack of publicly available unified bench-marks for the verification of visual localization algorithms. Therefore, to break the gap between data owners and users and solve the problems of high acquisition cost and slow update of vision data for complex illumination, it is essential to establish a continuously updatable robot vision localization dataset. Focusing on complex illumination environment, this paper explores the standard framework and data set construction process of visual positioning data sets, and conducts hardware selection and system construction according to the characteristics of the environment, and con-structs a visual positioning data set with complex lighting characteristics. The dataset evaluates various visual localization algorithms. This paper provides ground-truth locations for all visual sequences with centimeter-level accuracy, which can be used to test the ability of visual localization algorithms to handle lighting changes. The construction process proposed in this paper can provide transferable experience for the construction of other scene characteristic datasets. This dataset serves visual positioning verification task, realizes the connection between the real scene and the improvement of the visual positioning algorithm, and can provide more measurement methods for robot visual positioning system in complex illumination environment.