Trajectory Improvement of Railway Mobile Mapping System using Monocular Visual Inertial SLAM

Abstract

Nowadays, Mobile Mapping Systems (MMS) have been widely used in railway, with integrated Inertial Navigation System (INS)/Global Navigation Satellite System (GNSS) system as the common approach. For environments where GNSS signals become unavailable, additional aiding sources need to be considered to preserve the quality of measurement. Typically, in most applications, a Distance Measurement Indicator (DMI) is combined with the INS/GNSS system to ensure the desired accuracy. Nevertheless, as this approach is physically less feasible for railway, alternative solutions would be preferred. To bridge this gap, we investigate the applicability of Simultaneous Localization and Mapping (SLAM) integration with INS/GNSS for the case of a railway MMS. In particular, we aim to propose a solution for adapting a Monocular Visual (Inertial) SLAM method for railway application and further evaluate our solution using real-world data based on the RILA system (Fugro’s rail MMS). Accordingly, this dissertation consists of the following three research activities.
Firstly, we have conducted a literature review on the existing monocular Visual Inertial SLAM methods to identify the technique which fulfils the key requirements of the rail application. Considering the results of the study, we selected the ORB-SLAM3 method and proposed an end-to-end pipeline that covers all the phases to adapt it for RILA system.
Secondly, the impact of adapting ORB-SLAM3 technique on performance of the trajectory estimation has been evaluated using two criteria: Absolute Position Error (APE) and Relative Position Error (RPE). Accordingly, a case study using RILA dataset was developed for the experimental evaluations. In this case study, we have simulated a scenario in which the train entered the station, stayed there stationary for 1 minute, and then left the station slowly and manually inserted GNSS blockages before and after the station. Furthermore, a ground-truth trajectory was generated to evaluate the quality of the estimated SLAM-based trajectory. The results revealed that both APE and RPE increases with significant fluctuations in the first 15 seconds due to the lack of SLAM initialization time. Therefore, we introduced an strategy for fine-tuning the accuracy by allocating sufficient time for SLAM initialization phase. The APE and RPE were significantly reduced after fine-tuning and the expected estimated error at each time was equal to 4.3% of the travelled path length.
Finally, we presented the effect of using Zero Velocity Update (ZUPT) signals as aiding information on the accuracy of the estimated trajectory in the areas with poor GNSS coverage. Consequently, we manually inserted the extracted ZUPT signals to the INS/GNSS integration process and then compared the positional accuracy of the generated trajectories with and without this information. The results show that the estimated positional accuracy was improved by 30% with only 51 seconds of stationary condition in our case study.
Overall, based on the obtained results from the evaluations, it is possible to claim that our proposed ORB-SLAM3 technique has great potential to improve the estimated positional accuracy and in particular can be used as a standalone ZUPT detector in railway application.