一种运动恢复结构和航位推算结合的室内行人视觉定位方法

商业和工业领域中,室内行人、车辆、机器人的位置信息正逐渐成为人们关注的热点,并随之产生了大量的室内定位技术和方法,如使用无线信号、地磁、超宽带和超声波等方式进行室内定位。然而,目前的这些室内定位方法大多需要额外辅助设备的支撑,增加了定位成本和硬件开销。视觉定位作为一种目前较为流行的定位方式,具有实施成本低、不依赖任何外界辅助设备等优势。其中,构建带有位置标签的图像数据库是视觉定位方法的关键环节,而传统的构建图像数据库方法人力开销大、时耗长。因此,本文提出一种运动恢复结构（SFM）和航位推算结合的视觉定位方法,能够快速构建图像位置数据库、大大降低人力开销。该方法主要包括2个阶段：离线阶段和在线阶段。离线阶段主要实现图像序列位置的自动标注,通过采集行走路线上的手机内置传感器信息和视频信息,提出一种多约束图像匹配方法用于视频图像的连续匹配,将匹配结果用于SFM方法,可以得到相邻图像间的运动角度,使用行人航位推算（PDR）方法标注图像序列的轨迹坐标。在线阶段使用提出的图像匹配方法计算查询图像与数据库影像间的匹配点数量,将匹配点最多的K个数据库影像位置坐标加权平均作为查询图像的定位结果。最后,分别在2种典型的室内环境下进行实验,结果表明本文方法在离线阶段位置标注的平均误差为0.58 m,在线阶段图像匹配定位的误差范围在0.2~1.4 m。

An Indoor Pedestrian Positioning Approach Based on the Integration of SFM and PDR

Currently, the localization of users, such as people, vehicles or robots, in indoor spaces is a common issue in many commercial and industrial applications. A number of technologies have been proposed for indoor localization based on different principles such as RF (Radio Frequency), magnetic fields, ultra wide band (UWB) and ultrasound. Among these up-to-date indoor positioning technologies, most of them depend on special infrastructures or devices, which limit the commercial application of indoor localization. As one of the state-of-art indoor localization method, visual-based positioning scheme do not rely on any external auxiliary equipment and consequently has the advantage of low cost. However, the construction of geo-tagged image database, one of the most important parts for visual-based localization, is quite labor-intensive and time consuming. The automatic collection of geo-tagged indoor image data is an essential bottleneck for application of visual-based indoor localization systems. This paper proposed a visual-based indoor positioning approach which can automatically collect geo-tagged images based on the integration of structure from motion (SFM) and pedestrian dead rocking (PDR). The main idea of this method is to collect video frames as well as inertial data (by using smartphones) when people are walking in indoor environments. A method is designed to estimate the location (i.e., geo-tags) of images for the construction of geo-tagged image database. There are two phases for this approach: offline phase and online phase. During offline phase, the proposed method is used to estimate the location of the images extracted from video frames. A multi-constrain image matching algorithm was also developed to improve the performance of location estimation. There are three constraints in this multi-constrain image matching algorithm: ratio constraint, symmetry constraint and RANSAC constraint. Based on this image matching algorithm, a SFM process can be conducted to estimate the heading angle of a walking trajectory. After that, the coordinate of sampling points from the walking trajectory can be estimated by using the PDR method and the geo-tagged image database can be constructed. During the online phase, an indoor localization method is proposed to estimate the location of a pedestrian by finding the best matching images of a query image (taken by the pedestrian) from the image database. The multi-constraint image matching algorithm can be used to compute the number of matched key-points between query images and database images. The images with the most matched key-points are selected as the candidates. A weighted average function is used to estimate the location of the query image based on the selected images. In order to evaluate the performance of the visual-based indoor localization approach, two types of indoor environment are selected as the study area: an office building and a hospital. The experimental result showed that the average location estimation error of the geo-tagged images was 0.58 m. Based on the constructed image database, the error of the proposed localization approach ranged from 0.2 to 1.4 m. The performance of this approach indicated that visual information contribute to the construction of geo-tagged image database. This approach can be used in various indoor environments without any infrastructures or extra devices, which can reduce the difficulty in applying this approach to practical use.