Sensor Fusion for Indoor Localisation : Sensor Configuration and its Impact on Localisation Performance

Sammanfattning

Localisation is a foundational capability for autonomous systems, and its accuracy and robustness depend strongly on the configuration of the sensors involved. This thesis investigates how different sensor configurations influence the performance of a multi-sensor fused system for indoor localisation and SLAM. The Rawseeds dataset, recorded with a stereo camera, 2D-LiDAR, IMU, and wheel encoders, was used to evaluate how combinations of these sensors affect estimation accuracy and system robustness in a static indoor environment. Two mapping and localisation frameworks were used. ORB-SLAM3, used as a visual odometry baseline, and RTAB-Map, used for both SLAM and localisation. Through systematic experiments on two test sequences, each representing distinct motion characteristics and visual conditions, the contribution of each sensor modality was isolated and analysed. The evaluation was based on Absolute Trajectory Error (ATE) with respect to ground truth trajectories. The results confirm that the IMU plays an important role in maintaining robustness by preventing tracking loss during rapid motion, or in sub par visual conditions, while contributing less to translational accuracy when stability is already achieved. In contrast, wheel encoder odometry provides reliable motion estimations that reduces drift, rendering additional sensors less critical for maintaining track. Robustness is, however, reduced as a consequence of using wheel encoder odometry. Wheel encoders are inherently sensitive to disturbances such as slipping, something that brings robustness down regardless of whether such disturbances occur. The 2DLiDAR, meanwhile, complements both IMU and camera data by providing constraints that further refine pose estimates. The comparison between SLAM and localisation highlights the dependence on map quality when evaluating performance. Localisation within a previously made map generally results in lower translational errors, though the precision is limited by the quality of the map itself. The behaviour of different sensor configurations is also more consistent during localisation, indicating that once a good map is available, the system becomes less sensitive to which sensors are used. Overall, the findings demonstrate that the best localisation performance does not arise from using all available sensors, but from selecting the configuration best suited to the motion and environmental conditions that can be expected. In the cases that have been tested as part of this thesis, the use of stereo camera and IMU has been seen to be a good baseline that suits different motion types and visual conditions while providing good accuracy and robustness.