Visual-feature-assisted mobile robot localization in a long corridor environment

Abstract

Localization plays a vital role in the mobile robot navigation system and is a fundamental capability for autonomous movement. In an indoor environment, the current mainstream localization scheme uses two-dimensional (2D) laser light detection and ranging (LiDAR) to build an occupancy grid map with simultaneous localization and mapping (SLAM) technology; it then locates the robot based on the known grid map. However, such solutions work effectively only in those areas with salient geometrical features. For areas with repeated, symmetrical, or similar structures, such as a long corridor, the conventional particle filtering method will fail. To solve this crucial problem, this paper presents a novel coarse-to-fine paradigm that uses visual features to assist mobile robot localization in a long corridor. First, the mobile robot is remote-controlled to move from the starting position to the end along a middle line. In the moving process, a grid map is built using the laser-based SLAM method. At the same time, a visual map consisting of special images which are keyframes is created according to a keyframe selection strategy. The keyframes are associated with the robot’s poses through timestamps. Second, a moving strategy is proposed, based on the extracted range features of the laser scans, to decide on an initial rough position. This is vital for the mobile robot because it gives instructions on where the robot needs to move to adjust its pose. Third, the mobile robot captures images in a proper perspective according to the moving strategy and matches them with the image map to achieve a coarse localization. Finally, an improved particle filtering method is presented to achieve fine localization. Experimental results show that our method is effective and robust for global localization. The localization success rate reaches 98.8% while the average moving distance is only 0.31 m. In addition, the method works well when the mobile robot is kidnapped to another position in the corridor.

摘要

定位在移动机器人导航系统中起着至关重要的作用, 是自主移动的基本能力. 在室内环境中, 当前主流的定位方案使用2D激光雷达, 利用即时定位和建图(SLAM)技术来构建占据栅格地图; 然后, 基于已知的地图来定位. 然而, 此类方案仅在具有显著几何特征的区域有效. 对于重复、 对称或类似结构的区域, 例如长走廊, 常规粒子过滤方法将失效. 为解决这一问题, 本文提出一种从粗到细的模式, 该模式使用视觉特征辅助长走廊中的移动机器人定位. 首先, 移动机器人被远程控制, 沿着中线从起始位置移动到终点. 在移动过程中, 使用基于激光的SLAM方法建图. 同时, 根据关键帧选择策略创建关键帧图像组成的视觉地图. 关键帧通过时间戳与机器人的姿势相关联. 其次, 基于提取的激光扫描距离特征, 提出一种移动策略, 确定初始粗略位置. 这对于移动机器人来说至关重要, 因为它给出了机器人需要移动到哪里才能调整姿势的指令. 然后, 移动机器人根据移动策略以适当的视角捕捉图像, 并将其与图像地图进行匹配, 以获得粗略的定位. 最后, 提出一种改进的粒子滤波方法来实现精细定位. 实验结果表明, 该方法对全局定位是有效和鲁棒的. 定位成功率达98.8%, 平均移动距离仅0.31米. 此外, 当移动机器人被绑架到走廊中的另一个位置时, 该方法依然有效.