In recent years, real-time semantic segmentation on embedded devices has become increasingly popular, largely driven by the growing interest in smart vehicles and robots. The rising of autonomous driving has brought about new challenges for these systems, such as the need for low latency and computation-intensive operations, which can lead to excessive energy consumption and computing power. To address these challenges, this paper focuses on the critical task of semantic segmentation, which is essential for accurate environment perception, and proposes an implementation that achieves high accuracy and low complexity using a U-Net as base architecture. The goal is to enable real-time semantic segmentation on low-power cores while preserving performance, which is crucial for the success of autonomous vehicles and robots. The lightweight U-Net architectures have been implemented in a STM32 microcontroller, namely STM32L4R5, as a severe benchmark to meet the low-power, low-cost requirements.

AM-FM decomposition techniques have been successfully used for extracting significative features from a large variety of signals, helping realtime signal monitoring and pattern recognition, since they represent signals as a simultaneous composition of amplitude modulation and frequency modulation, where the carriers, amplitude envelopes, and the instantaneous frequencies are the features to be estimated. Human activities often involve repetitive movements, such as in running or cycling, where sinusoidal AM-FM decompositions of signals have already demonstrated to be useful to extract compact features to aid monitoring, classification, or detection. In this work we thus present the challenges and results of implementing the iterated coherent Hilbert decomposition (ICHD), a particularly effective algorithm to obtain an AM-FM decomposition, within a resource-constrained and low-power ARM Cortex-M4 microcontroller that is present in a wearable sensor we developed. We apply ICHD to the gyroscope data acquired from an inertial measurement unit (IMU) that is present in the sensor. Optimizing the implementation allowed us to achieve real-time performance using less then 16 % of the available CPU time, while consuming only about 5.4 mW of power, which results in a run-time of over 7 days using a small 250 mAh rechargeable cell.

Cardiac arrhythmia is one of the most important cardiovascular diseases (CVDs), causing million deaths every year. Moreover it is difficult to diagnose because it occurs intermittently and as such requires the analysis of large amount of data, collected during the daily life of patients. An important tool for CVD diagnosis is the analysis of electrocardiogram (ECG), because of its non-invasive nature and simplicity of acquisition. In this work we propose a classification algorithm for arrhythmia based on recurrent neural networks (RNNs) that operate directly on ECG data, exploring the effectiveness and efficiency of several variations of the general RNN, in particular using different types of layers implementing the network memory. We use the MIT-BIH arrhythmia database and the evaluation protocol recommended by the Association for the Advancement of Medical Instrumentation (AAMI). After designing and testing the effectiveness of the different networks, we then test its porting to an embedded platform, namely the STM32 microcontroller architecture from ST, using a specific framework to port a pre-built RNN to the embedded hardware, convert it to optimized code for the platform and evaluate its performance in terms of resource usage. Both in binary and multiclass classification, the basic RNN model outperforms the other architectures in terms of memory storage (∼117 KB), number of parameters (∼5 k) and inference time (∼150 ms), while the RNN LSTM-based achieved the best accuracy (∼90%).

The detection of cracks is a key aspect for assessing the condition of in-service structures such as road, bridges or dams. Therefore it assumes a great importance for civil infrastructures monitoring, road maintenance and traffic safety. Intelligent detection methods based on convolutional neural networks (CNNs) have been widely applied to the field of crack detection in recently years. In this work we propose a vision system based on two lightweight and accurate CNN models implemented in a low-cost, low-power platform, namely the OpenMV Cam H7 Plus, to monitor and to detect concrete cracks in real-time, suitable to realize a prototype of early warning system. In order to be useful, such a system must provide a very high accuracy, so as not to give false alarms, and be parsimonious enough on computational resources to be embedded into low-power, portable systems that can be deployed on the field. To reach this goal, firstly we analyze different state-of-the-art CNNs applied to the concrete crack detection task in order to discover the smallest network in terms of memory storage and number of parameters. Then, we compare the performance, in terms of memory occupancy and accuracy, of the proposed CNN architectures with the smallest network in the investigated literature, LeNet, all trained on two different image datasets, the Concrete Crack Images for Classification dataset and the SDNET2018 dataset, and implemented on the embedded system OpenMV Cam H7 Plus. The proposed CNN architectures perform nicely on this platform, using only a small fraction, between 6% to 26%, of the memory required by LeNet, and always providing better accuracy in all the tested cases and on both the datasets tried, with only a marginal increase in inference time.

The encephalographic (EEG) signal is an electrical signal that measures the brain activity. Due to its noninvasive acquisition process, it is often used to investigate the presence of Alzheimer’s disease (AD) or other common forms of neurodegerative disorders due to brain changes, that occur most frequently in older adults. Early detection of prodromal stages of AD, in which an individual has mild but measurable cognitive deficiencies with no significant effect on the functional activity of daily living, may help to reduce mortality and morbidity. This paper proposes an investigation of the classification of AD from EEG signal using robust-principal component analysis (R-PCA) feature extraction algorithm. Four widely used machine learning algorithms such as k-nearest neighbor (kNN), decision tree (DT), support vector machine (SVM), and naive Bayes have been implemented and compared by using a custom dataset composed of 13 subjects healthy or affected by AD in order to asses their classification performance.

The evaluation of the effectiveness of different machine learning algorithms on a publicly available database of signals derived from wearable devices is presented with the goal of optimizing human activity recognition and classification. Among the wide number of body signals we choose a couple of signals, namely photoplethysmographic (optically detected subcutaneous blood volume) and tri-axis acceleration signals that are easy to be simultaneously acquired using commercial widespread devices (e.g. smartwatches) as well as custom wearable wireless devices designed for sport, healthcare, or clinical purposes. To this end, two widely used algorithms (decision tree and k-nearest neighbor) were tested, and their performance were compared to two new recent algorithms (particle Bernstein and a Monte Carlo-based regression) both in terms of accuracy and processing time. A data preprocessing phase was also considered to improve the performance of the machine learning procedures, in order to reduce the problem size and a detailed analysis of the compression strategy and results is also presented.

This paper presents a system for the acquisition of in-house parameters, such as temperature, pressure, humidity and so on, that can be used for the intelligent control of a building. The main objective of this work is to determine an environmental model of an in-house room using machine learning techniques. The system is based on a low data-rate network of sensing and control nodes to acquire the data, realized with a new protocol, called ToLHnet, that is able to employ both wired and wireless communication on different media. Several standard machine learning techniques, namely linear regression, classification and regression tree algorithm, support vector machine, have been used for the regression of the input-output thermal model. Additionally, a recently proposed new technique named particle-Bernstein polynomial has been successfully applied. Experimental results show that this technique outperforms the previous techniques, for both accuracy and computation time.

The present work aims at the evaluation of the effectiveness of different machine learning algorithms on a variety of clinical data, derived from small, medium, and large publicly available databases. To this end, several algorithms were tested, and their performance, both in terms of accuracy and time required for the training and testing phases, are here reported. Sometimes a data preprocessing phase was also deemed necessary to improve the performance of the machine learning procedures, in order to reduce the problem size. In such cases a detailed analysis of the compression strategy and results is also presented.

Speaker identification plays a crucial role in biometric person identification as systems based on human speech are increasingly used for the recognition of people. Mel frequency cepstral coefficients (MFCCs) have been widely adopted for decades in speech processing to capture the speech-specific characteristics with a reduced dimensionality. However, although their ability to decorrelate the vocal source and the vocal tract filter make them suitable for speech recognition, they greatly mitigate the speaker variability, a specific characteristic that distinguishes different speakers. This paper presents a theoretical framework and an experimental evaluation showing that reducing the dimension of features by applying the discrete Karhunen-Loève transform (DKLT) to the log-spectrum of the speech signal guarantees better performance compared to conventional MFCC features. In particular with short sequences of speech frames, with typical duration of less than 2 s, the performance of truncated DKLT representation achieved for the identification of five speakers are always better than those achieved with the MFCCs for the experiments we performed. Additionally, the framework was tested on up to 100 TIMIT speakers with sequences of less than 3.5 s showing very good recognition capabilities.

This paper presents a technique for parametric model estimation of the motor unit action potential (MUAP) from the surface electromyography (sEMG) signal by using homomorphic deconvolution. The cepstrum-based deconvolution removes the effect of the stochastic impulse train, which originates the sEMG signal, from the power spectrum of sEMG signal itself. In this way, only information on MUAP shape and amplitude were maintained, and then, used to estimate the parameters of a time-domain model of the MUAP itself. In order to validate the effectiveness of this technique, sEMG signals recorded during several biceps curl exercises have been used for MUAP amplitude and time scale estimation. The parameters so extracted as functions of time were used to evaluate muscle fatigue showing a good agreement with previously published results.

Accurate heart rate (HR) estimation from photoplethysmography (PPG) recorded from subjects' wrist when the subjects are performing various physical exercises is a challenging problem. This paper presents a framework that combines a robust algorithm capable of estimating HR from PPG signal with subjects performing a single exercise and a physical exercise identification algorithm capable of recognizing the exercise the subject is performing. Experimental results on subjects performing two different exercises show that an improvement of about 50% in the accuracy of HR estimation is achieved with the proposed approach.

This paper presents a technique for learning hidden Markov model (HMM) state sequences from phonemes, that combined with modified discrete cosine transform (MDCT), is useful for speech synthesis. Mel-cepstral spectral parameters, currently adopted in the conventional methods as features for HMM acoustic modeling, do not ensure direct speech waveforms reconstruction. In contrast to these approaches, we use an analysis/synthesis technique based on MDCT that guarantees a perfect reconstruction of the signal frame feature vectors and allows for a 50% overlap between frames without increasing the data rate. Experimental results show that the spectrograms achieved with the suggested technique behave very closely to the original spectrograms, and the quality of synthesized speech is conveniently evaluated using the well known Itakura-Saito measure.

Sport, fitness, as well as rehabilitation activities, often require the accomplishment of repetitive movements. The correctness of the exercises is often related to the capability of maintaining the required cadence and muscular force. Failure to maintain the required force, also known as muscle fatigue, is accompanied by a shift in the spectral content of the surface electromyography (EMG) signal toward lower frequencies. This paper presents a novel approach for simultaneously obtaining exercise repetition frequency and evaluating muscular fatigue, as functions of time, by only using the EMG signal. The mean frequency of the amplitude spectrum (MFA) of the EMG signal, considered as a function of time, is directly related to the dynamics of the movement performed and to the fatigue of the involved muscles. If the movement is cyclic, MFA will display the same pattern and its average will tend to decrease. These two effects have been simultaneously modeled by a two-component AM-FM model based on the Hilbert transform. The method was tested on signals recorded using a wireless system applied to healthy subjects performing dumbbell biceps curls, dumbbell lateral rises, and bodyweight squats. Experimental results show the excellent performance of the proposed technique.

In this paper an effective technique to train an acoustic model from large and unsynchronized audio and text chunks is presented. Given such a speech corpus, an algorithm to automatically segment each chunk into smaller fragments and to synchronize those to the corresponding text is defined. These smaller fragments are more suitable to be used in standard model training algorithms for usage in automatic speech recognition systems. The proposed approach is particularly suitable to bootstrap language models without relying neither on specialized training material nor borrowing from models trained for other similar languages. Extensive experimentation using the CMU Sphinx 4 recognizer and the SphinxTrain model generator in a setting designed for large-vocabulary continuous speech recognition shows the effectiveness of the approach.

In this paper an effective unsupervised statistical identification technique for nonstationary nonlinear systems is presented. This technique extracts from the system outputs the multivariate relationships of the system natural modes, by means of the separation property of the Karhunen-Loève transform (KLT). Then, it applies a Self-Organizing Map (SOM) to the KLT output vectors in order to give an optimal representation of data. Finally, it exploits an optimized Expectation Maximization (EM) algorithm to find the optimal parameters of a Gaussian mixture model. The resulting statistical system identification is thus based on the estimation of the multivariate probability density function (PDF) of system outputs, whose convergence towards that computed by kernel estimation has also been proved by verifying the asymptotically vanishing of Kullback-Leibler divergences. A large number of simulations on ECG signals demonstrated the validity and the excellent performance of this technique along with its applicability to noninvasive diagnosis of a large class of medical pathologies originated by unknown, unpractical to measure, physiological factors.

This paper addresses nonlinear nonstationary system identification from stimulus-response data, a problem concerning a large variety of applications, in dynamic control as well as in signal processing, communications, physiological system modelling and so on. Among the different methods suggested in the vast literature for nonlinear system modelling, the ones based on the Volterra series and the Neural Networks are the most commonly used. However, a strong limitation for the applicability of these methods lies in the necessary property of stationarity, an assumption that cannot be considered as valid in general and strongly affecting the validity of results. Another weakness of the approaches currently used is that they refer to differential systems, thus being unsuitable to model systems described by integral equations. A computational intelligence technique that exploits the potentialities of both the Karhunen-Loeve Transform (KLT) and Neural Networks (NNs) representation and without any of the limitations of the previous approaches is suggested in this paper. The technique is suitable for modelling the wide class of systems described by nonlinear nonstationary functionals, thus including both differential and integral systems. It takes advantage of the KLT separable kernel representation that is able to separate the dynamic and static behaviours of the system as two distinct components, and the approximation property of NNs for the identification of the nonlinear no-memory component. To validate the suggested technique comparisons with experimental results on both nonlinear nonstationary differential and integral systems are reported.

This paper presents a methodology for statistical simulation of non-linear integrated circuits affected by device mismatch. This simulation technique is aimed at helping designers maximize yield, since it can be orders of magnitude faster than other readily available methods, e.g. Monte Carlo. DC, AC, and transient analyses are performed by means of the stochastic modified nodal analysis, using a piecewise linearization technique with respect to the stochastic sources, around a few automatically selected points.

The learning capability of neural networks is equivalent to modeling physical events that occur in the real environment. Several early works have demonstrated that neural networks belonging to some classes are universal approximators of input-output deterministic functions. Recent works extend the ability of neural networks in approximating random functions using a class of networks named stochastic neural networks (SNN). In the language of system theory, the approximation of both deterministic and stochastic functions falls within the identification of nonlinear no-memory systems. However, all the results presented so far are restricted to the case of Gaussian stochastic processes (SPs) only, or to linear transformations that guarantee this property. This paper aims at investigating the ability of stochastic neural networks to approximate nonlinear input-output random transformations, thus widening the range of applicability of these networks to nonlinear systems with memory. In particular, this study shows that networks belonging to a class named non-Gaussian stochastic approximate identity neural networks (SAINNs) are capable of approximating the solutions of large classes of nonlinear random ordinary differential transformations. The effectiveness of this approach is demonstrated and discussed by some application examples.

This paper presents, on the basis of a rigorous mathematical formulation, a multicomponent sinusoidal model that allows an asymptotically exact reconstruction of nonstationary speech signals, regardless of their duration and without any limitation in the modeling of voiced, unvoiced, and transitional segments. The proposed approach is based on the application of the Hilbert transform to obtain an amplitude signal from which an AM component is extracted by filtering, so that the residue can then be iteratively processed in the same way. This technique permits a multicomponent AM-FM model to be derived in which the number of components (iterations) may be arbitrarily chosen. Additionally, the instantaneous frequencies of these components can be calculated with a given accuracy by segmentation of the phase signals. The validity of the proposed approach has been proven by some applications to both synthetic signals and natural speech. Several comparisons show how this approach almost always has a higher performance than that obtained by current best practices, and does not need the complex filter optimizations required by other techniques

In this paper a novel modeling technique for speech signals, based on the source-filter model of speech production and on orthogonal transform theory, is presented. The proposed approach models the impulse response of such filter, by projection onto a basis of damped Bessel functions, which have been chosen for their similarity to the signal to be modeled. In such a way an orthogonal transform pair is defined which provides a simple and effective methodology for the extraction of model parameters, and its effectiveness in the case of voiced speech has been demonstrated by synthesizing natural sounding speech signals with the aid of only a few extracted parameters.

The Karhunen-Loeve transform, being able to represent stochastic processes under appropriate conditions, is a powerful signal processing tool. But the high computational cost incurred in the modeling of long signals has limited its use in the recognition of speech segmented at the word level. In this paper, we present a novel algorithm that significantly reduces the computational cost when the number of signals to be treated is small in comparison to their samples.

In this paper a simulator for the statistical analysis of analog CMOS integrated circuits affected by technological tolerance effects, including device mismatch, is presented. The tool, able to perform dc, ac, and transient analyses, is based on a rigorous formulation of circuit equations starting from the modified nodal analysis and including random current sources to take into account technological tolerances. Statistical simulation of specific circuits shows that the simulator requires a simulation time several orders of magnitude lower than that required by Monte Carlo analysis, while ensuring a good accuracy.

This paper presents a simulator for the statistical analysis of MOS integrated circuits a ected by mismatch e ect. The tool is based on a rigorous formulation of circuit equations including random current sources to take into account technological tolerances. The simulator requires a simulation time of several orders of magnitude lower than that required by Montecarlo analysis, while ensuring a good accuracy.