| Summary: | In recent times, there has been a continuous increase in the ubiquity, processing power and sensing capabilities of modern smartphones. This has made possible the emergence of new technologies that allows users to keep track of information regarding their health, activities and location, even in indoor places were GPS signal is not available. These technologies typically rely on fusing and processing information coming from multiple sensors, such as the accelerometer or the magnetometer. This thesis proposes a framework for indoor location and activity recognition from new source of information: the sound perceived through the device’s microphone. It does so by extracting information relative to the user’s position and activities through machine learning and audio processing techniques. In the context of indoor location, the proposed SoundSignature algorithm allows the device to learn from labeled data and predict the location it is in. These locations may be different rooms or distinct regions of large places, such as open spaces. Another proposed algorithm, SoundSimilarity, further refines this positioning by comparing the sound signals from two or more devices in real time. A novel audio similarity metric identifies if the devices are close to one another, mitigating the potential errors of the SoundSignature algorithm. This also has many other use cases, such as detecting proximity between the user and devices. Finally, the Activity Monitoring algorithm allows the device to learn from labeled data to recognize the activity the user is performing. This information may be also used to further refine the location algorithm by recognizing location-dependent activities such as the closing of doors or watching television.
|
|---|