柳景斌, 郭英晖, 喻文慧. 一种面向大型室内场景的高可用手机视觉全局定位方法[J]. 武汉大学学报 ( 信息科学版). DOI: 10.13203/j.whugis20210602
引用本文: 柳景斌, 郭英晖, 喻文慧. 一种面向大型室内场景的高可用手机视觉全局定位方法[J]. 武汉大学学报 ( 信息科学版). DOI: 10.13203/j.whugis20210602
Liu Jingbin, Guo Yinghui, Yu Wenhui. A Smartphone based visual global localization method with high usability in large indoor spaces[J]. Geomatics and Information Science of Wuhan University. DOI: 10.13203/j.whugis20210602
Citation: Liu Jingbin, Guo Yinghui, Yu Wenhui. A Smartphone based visual global localization method with high usability in large indoor spaces[J]. Geomatics and Information Science of Wuhan University. DOI: 10.13203/j.whugis20210602

一种面向大型室内场景的高可用手机视觉全局定位方法

A Smartphone based visual global localization method with high usability in large indoor spaces

  • 摘要: 基于视觉的手机定位方法是室内定位中的研究热点,但当前存在的面向机场、商场等大型室内环境时存在可靠性差、计算效率低的问题。本文面向这类场景,针对现有的基于室内实景地图和手机视觉匹配的定位方法存在的问题,提出一种基于三维实景地图的“粗定位-精定位”二级定位方法,具体包括:基于WiFi指纹匹配粗定位结果约束匹配图像库范围,子区域分段式建立特征库及利用深度学习的方法去除天花板图像。实验结果表明,大型室内场景下本文所提方法可以将视觉匹配定位精度由1.89m提升至0.45 m,将定位计算效率提升5倍以上。本文方法能够有效降低定位时间,提升特征点云的精度进而提升视觉匹配定位精度,同时能幅度降低特征点匹配错误而造成定位错误的情况,可以实现高可用、亚米级精度的室内视觉全局定位。在特征点提取和匹配的过程中,准确性还有提升空间,因此后续研究将引入线特征等元素提升定位精度。

     

    Abstract: Objectives: Smartphone based visual global localization is a research hotspot in the location based services community. Existing methods suffer from the problems of poor reliability and low computing efficiency especially when they are used for large indoor environments such as airport and shopping mall. M ethods: This paper proposes a “Rough localization to Accurate localization”two-level localization method, it is based on 3d real map and applied to large indoor scenes such as shopping malls. To reduces the location computing time, a method of limiting the scope of image database is proposed: Use WiFi fingerprint matching algorithm to obtain the location results, then limit the image database according to the location results. In order to improve the positioning accuracy, a new method of constructing database is proposed. The whole scene is divided into multiple regions, each region completes the database establishment independently, and then splices different databases. In order to reduce the location error, a scene recognition method is proposed. Deep learning method is used to remove the ceiling image. It can reduce the feature points matching errors. Results: Comparing the location computing time before limiting the scope of image database and after, the proposed method reduced the positioning time from 6s to 0.8s per image. Comparing the location error and positioning precision before removing the ceiling image and after, the scene recognition method reduced the location error from 5.76m to 2.78m and improved the positioning precision from 0.98 m to 0.84 m. The new method of constructing database can improve the positioning precision from 0.82m to 0.45m. All in all, the proposed method can improve the positioning precision from 0.98 m to 0.45 m, reduce the positioning time more than 5 times. Conclusions: Although the proposed method achieves sub-meter accuracy of indoor vision global positioning, the feature points matching errors affects the positioning precision. In future work, feature lines will be used to improve the positioning accuracy.

     

/

返回文章
返回