Minimal Infrastructure Radio Frequency Home Localisation Systems

The ability to track the location of a subject in their home allows the provision of a number of location based services, such as remote activity monitoring, context sensitive prompts and detection of safety critical situations such as falls. Such pervasive monitoring functionality offers the potential for elders to live at home for longer periods of their lives with minimal human supervision. The focus of this thesis is on the investigation and development of a home roomlevel localisation technique which can be readily deployed in a realistic home environment with minimal hardware requirements. A conveniently deployed Bluetooth ® localisation platform is designed and experimentally validated throughout the thesis. The platform adopts the convenience of a mobile phone and the processing power of a remote location calculation computer. The use of Bluetooth ® also ensures the extensibility of the platform to other home health supervision scenarios such as wireless body sensor monitoring. Central contributions of this work include the comparison of probabilistic and nonprobabilistic classifiers for location prediction accuracy and the extension of probabilistic classifiers to a Hidden Markov Model Bayesian filtering framework. New location prediction performance metrics are developed and signicant performance improvements are demonstrated with the novel extension of Hidden Markov Models to higher-order Markov movement models. With the simple probabilistic classifiers, location is correctly predicted 80% of the time. This increases to 86% with the application of the Hidden Markov Models and 88% when high-order Hidden Markov Models are employed. Further novelty is exhibited in the derivation of a real-time Hidden Markov Model Viterbi decoding algorithm which presents all the advantages of the original algorithm, while producing location estimates in real-time. Significant contributions are also made to the field of human gait-recognition by applying Bayesian filtering to the task of motion detection from accelerometers which are already present in many mobile phones. Bayesian filtering is demonstrated to enable a 35% improvement in motion recognition rate and even enables a floor recognition rate of 68% using only accelerometers. The unique application of time-varying Hidden Markov Models demonstrates the effect of integrating these freely available motion predictions on long-term location predictions