The Global Positioning System (GPS), a space-based system, employs dozens of satellites to provide location determination and navigation services around the world. However, due to the constraints to the power consuming and long-distance transmission, the strength of the GPS signal received on the mobile device is weak. Errors of the detection of the line-of-sight (LOS) propagated components of the signals are expected to be high if the users are in urban areas or in buildings, since obstacles in the surrounding environments could attenuate the LOS propagated components of the GPS signals significantly, but might upfade the multi-path components (constructive multi-path effect). Therefore, GPS should be replaced by other techniques for providing localisation services in urban areas or, especially, in indoor environments.
Among all the candidates, received signal strength (RSS) location fingerprint based positioning systems attract great attentions from both the academia and industry. Usually, a time-consuming and labour-intensive site survey to collect dozens of training samples of RSS from access points (APs) in range on every reference position (RP) in the area of interest is required to build the radio map (RM), before the localisation services could be provided to users. The purpose of the thesis is to reduce the workload involved in the site survey while providing accurate localisation service from two aspects, as shown as follows.
Firstly, the quantity of the training samples collected on each RP is reduced, by taking advantage of the on-line RSS measurements collected by users to calibrate the RM. The on-line RSS measurements are geo-tagged probabilistically by an implementation of particle filter to track the trajectories of the users. The employed particles in estimation of the users’ states are initialised by a supervised clustering algorithm, propagated according to the analysis of the data sourcing from inertial measurement units (IMUs), e.g., walking detection, orientation estimation, step and stepping moments detection, step length detection, etc., and corrected by the wall constraints. Furthermore, the importance weights of the particles are adjusted to reduce the negative influence of the multi-clustered distribution of the particles to the on-line localisation accuracy, by applying the on-line RSS-based localisation results when significant users' body turnings are detected. The final results confirm that the accuracy of the localisation service with the RM calibrated by the method proposed in this thesis is higher than the previously proposed approach taking advantage of expectation maximisation algorithm.
Secondly, a semi-automatic site-survey method which takes advantage of a route-planning algorithm and a walking detection module to recognise automatically the index of the RP for the current site-survey task, inform the system automatically of the start/end of the process of the task on the current RP and switch automatically to the following RPs on the planned route for the following tasks, is proposed. In this way, human beings' intervention to the site-survey process is greatly reduced. As a result, the errors made in the site-survey tasks, such as incorrect recognition of the index of the RP for the current task which is highly likely to occur when the technicians get absent-minded in the work, misoperations to start/end of the task for collecting RSS samples on the current RP at wrong time moments, forgetting to notify the system of the fact that the technician has moved on to the next RP, etc., are avoided. The technicians no longer feel bored or anxious in the process of fulfilment of site-survey tasks, and the working efficiency and robustness of the RM could be also improved.
