From Compression of Wearable-based Data to Effortless Indoor Positioning

In recent years, wearable devices have become ever-present in modern society. They are typically defined as small, battery-restricted devices, worn on, in, or in very close proximity to a human body. Their performance is defined by their functionalities as much as by their comfortability and convenience. As such, they need to be compact yet powerful, thus making energy efficiency an extremely important and relevant aspect of the system. The market of wearable devices is nowadays dominated by smartwatches and fitness bands, which are capable of gathering numerous sensor-based data such as temperature, pressure, heart rate, or blood oxygen level, which have to be processed in real-time, stored, or wirelessly transferred while consuming as little energy as possible to ensure long battery life. Implementing compression schemes directly at the wearable device is one of the relevant methods to reduce the volume of data and to minimize the number of required operations while processing them, as raw measurements include plenty of redundancies that can be removed without damaging the useful information itself.

This thesis presents a number of contributions in the field of compression of wearable-based data, mainly in areas of lossy compression techniques designated for the time series sensor-based data and positioning. In the scope of this work, two novel time-series compression techniques are proposed, namely Direct Lightweight Temporal Compression (DLTC) and Altered Symbolic Aggregate Approximation (ASAX), which are specifically designed to address relevant challenges of modern wearable systems. As many of the modern wearables also possess localization capabilities critical for navigation, tracking, and monitoring applications, reducing the computational and storage demands for indoor positioning applications is the second addressed challenge. Performing the positioning task quickly and efficiently on all connected devices, including wearables, becomes crucial in industrial applications, eHealth, or security. As the localization technique of choice in Global Navigation Satellite System (GNSS) signal-obscured scenarios, positioning via fingerprinting proves a reliable and efficient solution, while arising new challenges to be solved. Improving the efficiency of the fingerprinting-based system by applying lossy compressions onto the training radio map is realized by proposing, implementing, and evaluating various novel dimensionality-reduction techniques.

This thesis proposes Element-Wise cOmpression using K-means (EWOK), a bitlevel compression based on element-wise k-means clustering, radio Map compression Employing Signal Statistics (MESS), a sample-wise compression that extracts signal statistics based on their locations, as well as evaluates feature-wise methods Principal Component Analysis (PCA) and Auto-Encoder (AE) that transform fingerprints into low-dimensional representation.

The evaluation in the thesis shows the effectiveness of each compression scheme on 26 different datasets and provides the results achieved by combining the individual schemes together, accomplishing multi-dimensional radio map compression that sustains high positioning accuracy of the dataset, despite manyfold size reduction.

The processing requirements of the positioning system are further addressed by proposing a cascade of models that reduces the required search space of the algorithm. By combining numerous Machine Learning (ML) architectures, it is capable of further reducing the positioning time (and thus, positioning effort), while improving the positioning performance. The thesis further includes the introduction of an indoor positioning dataset collected by the author, denoted TUJI 1, a novel performance metric to evaluate the latency caused by the lossy compression, and several crucial
adjustments to the distance metric calculations, generalizing their applicability.

The thesis provides novel insights into the compression of sensor-based, timeseries data and into reducing the computational effort of the fingerprinting positioning schemes while introducing a relevant number of novel and efficient solutions beyond the State-of-the-Art.