Outlier Detection for Indoor Vision-based Navigation Applications

In the research domain of indoor navigation or more generally, navigation in a GPS-denied environment, vision is believed to be one of the most promising but challenging technologies so far. Major problem lies in that the accuracy and reliability of vision-based positioning can easily be sabotaged by misidentified visual features. For map-based approaches, errors can also be easily brought into the system in the mapping procedure. In order to tackle this problem, we compare and evaluate different outlier detection strategies in the context of vision-based navigation in this paper, and propose a multi-level operation scheme of outlier detection for such applications. The approach is implemented on our earlier built vision-based positioning system. Experiments prove that a combined use of different outlier detection methods in a multi-level strategy can treat gross errors from different sources of a vision-based navigation system (e. g. image measurement, mismatches, ground control survey) specifically, and thus the accuracy and reliability of the system can be improved.