Augmented Reality Technologies for the Visualisation of SLAM Systems

Simultaneous Localisation and Mapping (SLAM) is a popular and important autonomous mapping and navigation technique in mobile robotics. Due to the probabilistic nature and the real world uncertainties in which the robot operates, SLAM development and testing is challenging. The difficulty exists because of a lack of perceptual and cognitive overlap between the robot and the human developer, i.e. understanding what the robot is seeing and processing. The most promising way of achieving this overlap is through the use of visualisations, because the human visual system is highly perceptive with advanced pattern recognition abilities. The research presented in this thesis investigates the application of Augmented Reality (AR) for SLAM visualisations, with the goal of assisting SLAM development and testing. AR is well suited for this application as it provides a real world view of the robot and its physical environment. A literature survey in SLAM, visualisation and AR showed that while SLAM visualisations are lacking AR has been applied in other areas of mobile robotics. An anonymous web based survey of SLAM developers confirmed that SLAM development is challenging and that graphical visualisations of SLAM are essential but lacking. SLAM algorithms were analysed in order to identify parameters needing to be visualised for error detection and correction, and visualisation requirements were formulated from this analysis. An AR visualisation system for SLAM was developed and implemented, presenting novel visualisation techniques for the Extended Kalman Filter SLAM and the Rao-Blackwellized Particle Filter FastSLAM. An evaluation was designed and carried out, investigating the effectiveness of the AR visualisation system in assisting SLAM debugging, and its performance and registration accuracy characteristics. The findings show that the existing state of the art visualisations are more effective, and should be used, for detecting fault effects. The findings further show that the novel feature correlation and data association AR visualisations are more effective and more preferred, and thus should be used, for detecting fault causes. Fault correction findings seem to indicate that the novel feature correlations and colour-mapping AR visualisations are more effective and more preferred. Qualitative user feedback showed