Logo Logo
Switch Language to German

Marouane, Chadly; Gutschale, Robert and Linnhoff-Popien, Claudia (2018): Visual Odometry for Pedestrians Based on Orientation Attributes of SURF. In: Proceedings of Sai Intelligent Systems Conference (Intellisys) 2016, Vol 2, Vol. 16: pp. 153-174

Full text not available from 'Open Access LMU'.


With the decreasing size and prize of cameras, visual positioning systems for persons are becoming more attractive and feasible. A main advantage of visual methods is that they can be independent of any infrastructure and are therefore applicable in indoor as well as outdoor scenarios. As such, they are an attractive alternative to infrastructure based methods. This paper presents a method that uses visual data to create a two-dimensional trajectory of the pedestrians movement. A camera that is mounted on the persons upper body is used to obtain the image data. With the SURF algorithm, feature points that posses specific attributes are extracted from the image frames. Based on these attributes, the method determines the pedestrians steps and estimates the heading at each step. As each determination of a new position is based on previous estimations, the method accumulates errors with increasing distances. An extensive evaluation with different test persons for various scenarios demonstrates that the method achieves a reasonably good overall accuracy for shorter distances. For distances of up to 25m, a mean error of 5.52 m for indoor scenarios and of 7.56m for outdoor scenarios has been determined. Furthermore, the method is also reliably functional at increasing walking and running speeds. An additional evaluation shows the usability of one of the SURF-attributes for the implementation of an activity detector for different movement speeds. The method robustly detects steps with a high accuracy at an error rate of approximately one percent. However, just like other Pedestrian Dead Reckoning methods, the heading estimation proves to be challenging and to be a source of errors.

Actions (login required)

View Item View Item