Sensing and Compression Techniques for Environmental and Human Sensing Applications

In this doctoral thesis, we devise and evaluate a variety of lossy compression schemes for Internet of Things (IoT) devices such as those utilized in environmental wireless sensor networks (WSNs) and Body Sensor Networks (BSNs). We are especially concerned with the efficient acquisition of the data sensed by these systems and to this end we advocate the use of joint (lossy) compression and transmission techniques. Environmental WSNs are considered first. For these, we present an original compressive sensing (CS) approach for the spatio-temporal compression of data. In detail, we consider temporal compression schemes based on linear approximations as well as Fourier transforms, whereas spatial and/or temporal dynamics are exploited through compression algorithms based on distributed source coding (DSC) and several algorithms based on compressive sensing (CS). To the best of our knowledge, this is the first work presenting a systematic performance evaluation of these (different) lossy compression approaches. The selected algorithms are framed within the same system model, and a comparative performance assessment is carried out, evaluating their energy consumption vs the attainable compression ratio. Hence, as a further main contribution of this thesis, we design and validate a novel CS-based compression scheme, termed covariogram-based compressive sensing (CB-CS), which combines a new sampling mechanism along with an original covariogram-based approach for the online estimation of the covariance structure of the signal. As a second main research topic, we focus on modern wearable IoT devices which enable the monitoring of vital parameters such as heart or respiratory rates (RESP), electrocardiography (ECG), and photo-plethysmographic (PPG) signals within e-health applications. These devices are battery operated and communicate the vital signs they gather through a wireless communication interface. A common issue of this technology is that signal transmission is often power-demanding and this poses serious limitations to the continuous monitoring of biometric signals. To ameliorate this, we advocate the use of lossy signal compression at the source: this considerably reduces the size of the data that has to be sent to the acquisition point by, in turn, boosting the battery life of the wearables and allowing for fine-grained and long-term monitoring. Considering one dimensional biosignals such as ECG, RESP and PPG, which are often available from commercial wearable devices, we first provide a throughout review of existing compression algorithms. Hence, we present novel approaches based on online dictionaries, elucidating their operating principles and providing a quantitative assessment of compression, reconstruction and energy consumption performance of all schemes. As part of this first investigation, dictionaries are built using a suboptimal but lightweight, online and best effort algorithm. Surprisingly, the obtained compression scheme is found to be very effective both in terms of compression efficiencies and reconstruction accuracy at the receiver. This approach is however not yet amenable to its practical implementation as its memory usage is rather high. Also, our systematic performance assessment reveals that the most efficient compression algorithms allow reductions in the signal size of up to 100 times, which entail similar reductions in the energy demand, by still keeping the reconstruction error within 4 % of the peak-to-peak signal amplitude. Based on what we have learned from this first comparison, we finally propose a new subject-specific compression technique called SURF Subject-adpative Unsupervised ecg compressor for weaRable Fitness monitors. In SURF, dictionaries are learned and maintained using suitable neural network structures. Specifically, learning is achieve through the use of neural maps such as self organizing maps and growing neural gas networks, in a totally unsupervised manner and adapting the dictionaries to the signal statistics of the wearer. As our results show, SURF: i) reaches high compression efficiencies (reduction in the signal size of up to 96 times), ii) allows for reconstruction errors well below 4 % (peak-to-peak RMSE, errors of 2 % are generally achievable), iii) gracefully adapts to changing signal statistics due to switching to a new subject or changing their activity, iv) has low memory requirements (lower than 50 kbytes) and v) allows for further reduction in the total energy consumption (processing plus transmission). These facts makes SURF a very promising algorithm, delivering the best performance among all the solutions proposed so far.