In the last decade, research activities on outdoor tracking technology have seen an explosion of advances. It is currently possible to localise pedestrians outdoors to decimetre precision using mass-market raw global satellite navigation system receivers; however, localising pedestrians indoors to a similar accuracy is not yet possible. These limitations led us to start an investigation on how the indoor localisation systems can be improved.
The main contribution of this thesis is a novel method based on the pose graph optimisation problem. The method uses the iterative Non-Linear Least Squares (NLLS) solver to solve the localisation problem and integrates a better usage of sensor models, which are pre-compute values instead of time-consuming on-going estimations. This process decreases run-time complexity and allows the integration of sensor models directly into the localisation solver.
To build sensor models from a training set, the machine learning method Gaussian Process was used. As this requires good ground truth estimates, the common approach is to use an additional localisation system that provides more accurate observations as ground truth. These systems are either more expensive or difficult to set up. Therefore, the second major contribution of this thesis is a proposed method that use existing sensors in combination with known landmarks to compensate for long-term sensor drift. The extended NLLS solver is simultaneously used to correct the pedestrian trajectory, to estimate on-going sensor drift and to allow for slight deviations in the landmark's absolute location.
In addition to proposed methods and algorithms, the entire indoor localisation system was built and deployed in different environments, which allowed us to collect hours of real-world data used in further analyses and to built machine learning models.
|
