This paper presents a novel approach to noninvasively assess different levels of perfusion using remote photoplethysmography (rPPG) from an RGB camera. In a proof-of-concept study involving 20 healthy participants, three degrees of stenosis were induced in the upper extremity by applying external pressure. The rPPG signal from the affected arm was correlated with a reference signal from the well-perfused control arm. The results show that correlation analysis of rPPG signals can distinguish between mild, moderate, and full stenosis. This supports the potential of rPPG as a cost-effective tool for detecting perfusion disorders such as peripheral artery disease (PAD).

Sleep apnea syndrome (SAS) is a sleep-related breathing disorder characterized by repetitive breathing interruptions during sleep, resulting in daytime drowsiness, concentration difficulties, and increased risk of cardiovascular diseases. SAS is diagnosed in specialized sleep laboratories via polysomnography (PSG). PSG involves a high number of contact-based sensors, and it may cause patient discomfort and bias in measurement results. Therefore, contactless alternatives to PSG are a promising way to overcome these issues. This work introduces a novel feature-based method for detecting and classifying apneic events in terms of event amplitude (apneas and hypopneas) and event source (obstructive and central) by using multispectral imaging in the near-infrared (NIR) and far-infrared (FIR) spectra. In the NIR spectrum, remote photoplethysmography signals at 780 and 940 nm are extracted, while in the FIR spectrum, a respiratory airflow signal is extracted. The method is based on the extraction of explainable and medically significant features and the fusion of multiple data modalities (multispectral images, demographic patient data, interspectral correlation analysis, and time-series analysis). The classification accuracy between normal breathing, hypopneas, and apneas is 99.5%, while the differentiation between obstructive and central apneas achieves an accuracy of 98.8%. The estimations of the apnea-hypopnea index (AHI), obstructive apnea index (oAI), and central apnea index (cAI) result in a Pearson correlation of 0.9981, 0.9989, and 0.9950 respectively. A correct SAS stage prediction for 19 symptomatic patients in our dataset is accomplished. The results show that this method may be used as a PSG substitute for apnea and hypopnea detection and classification.

Tiny ML, defined by its ability to integrate AI into the smallest devices, unlocks the potential for always-on AI solutions at the edge. This capability is essential in predictive maintenance, enabling smart sensors within the Industrial In-ternet of Things (1IoT). In factory automation, the ability to detect anomalies and adapt those anomaly detection models to varying operating conditions is crucial for the efficiency and longevity of machinery. However, the challenge of semi-supervised anomaly detection with domain adaptation across diverse working conditions has not been sufficiently addressed in TinyML. This paper presents Shared Encoder Domain Adaptation (SEDA) to overcome these limitations. SEDA is a multi-target domain adaptation method for anomaly detection tailored to TinyML applications. Our approach facilitates the transfer of knowledge from existing domains to new ones, ensuring reliable anomaly detection in all domains with minimal data under new operational conditions. This is particularly relevant for IloT applications, where sensors must perform condition monitoring in a self-learning manner. The effectiveness of our method is evaluated through an extensive parameter study on the Toy-ADMOS 2 data set and compared to various baseline methods. The results show an improvement of up to 0.212 on the Area Under the Receiver Operating Characteristic Curve (ROC-AUC) score. It demonstrates that our method significantly enhances performance in multi-target domain adaptation. Importantly, it does this while remaining efficient for resource-constrained systems. This method has the potential to enable self-learning sensor systems with adaptation strategies to different working conditions while requiring only a few data points from these new domains.

Edge Artificial Intelligence (AI) relies on the integration of Machine Learning (ML) into even the smallest embedded devices, thus enabling local intelligence in real-world applications, e.g. for image or speech processing. Traditional Edge AI frameworks lack important aspects required to keep up with recent and upcoming ML innovations. These aspects include low flexibility concerning the target hardware and limited support for custom hardware accelerator integration. Artificial Intelligence for Embedded Systems Framework (AIfES) has the goal to overcome these challenges faced by traditional edge AI frameworks. In this paper, we give a detailed overview of the architecture of AIfES and the applied design principles. Finally, we compare AIfES with TensorFlow Lite for Microcontrollers (TFLM) on an ARM Cortex-M4-based System-on-Chip (SoC) using fully connected neural networks (FCNNs) and convolutional neural networks (CNNs). AIfES outperforms TFLM in both execution time and memory consumption for the FCNNs. Additionally, using AIfES reduces memory consumption by up to 54% when using CNNs. Furthermore, we show the performance of AIfES during the training of FCNN as well as CNN and demonstrate the feasibility of training a CNN on a resource-constrained device with a memory usage of slightly more than 100 kB of RAM.

In recent years, many studies have been published on blood pressure (BP) estimation from photoplethysmogram (PPG) signals to overcome limitations of current BP devices, such as high risk of complications in invasive methods or limited accuracy, temporal resolution and comfort for cuff-based systems. However, most of these studies suffer from methodological drawbacks regarding data handling, leading to overly positive evaluation of such methods. We present an approach on beat-to-beat blood pressure estimation from raw PPG signals that is evaluated extensively on data of unseen test subjects. The proposed method obtains an mean absolute error ± standard deviation of 8.07 ± 6.86 mmHg for DBP and 8.73 ± 7.36 mmHg for SBP. To increase transparency of the model's decision-making, we examine layer activation of the employed convolutional neural network. Moreover, we analyze the impact of fine tuning for personalization of the model and derive strategies to enhance the personalization process.

The emerging Edge Artificial Intelligence (AI) paradigm is a key driver of innovation in many areas such as audio or industrial sensor data processing. Edge AI relies on the integration of machine and deep learning-based signal processing into even the smallest embedded devices thus enabling local intelligence in real-world applications. Nevertheless, classical signal processing techniques such as the Fast Fourier Transform (FFT) still remain an essential component of Edge AI systems, especially in the context of data preprocessing. However, their optimization with respect to resource requirements is often neglected although it can increase the performance of the overall system. In this paper, we present a new approximate FFT (AFFT) approach that enables the end-to-end resource optimization of digital signal processing (DSP) pipelines in Edge AI systems. The approach combines an aggressive FFT parameter quantization with error mitigation techniques to trade off accuracy and performance. To evaluate the methodology, a flexible code generation framework is implemented. We analyze the accuracy of the AFFT, demonstrate its capabilities in an audio classification use-case involving a convolutional neural network (CNN) and benchmark the approach on an ARM Cortex-M4-based System-on-Chip (SoC). Our methodology maintains the classification accuracy of the CNN close to the full-precision level while outperforming the ARM CMSIS FFT both in execution time and energy consumption.

The development in the field of autonomous driving goes hand in hand with ever new developments in the field of image processing and machine learning methods. In order to fully exploit the advantages of deep learning, it is necessary to have sufficient labeled training data available. This is especially not the case for omnidirectional fisheye cameras. As a solution, we propose in this paper to use synthetic training data based on Unity3D. A five-pass algorithm is used to create a virtual fisheye camera. This synthetic training data is evaluated for the application of free space detection for different deep learning network architectures. The results indicate that synthetic fisheye images can be used in deep learning context.

In the context of Ambient Assisted Living (AAL), the detection of daily activities is an active field of research. In this study, we present an algorithm for the performed Activities of Daily Living (ADLs) related to personal hygiene, which is based on the evaluation of a person’s proximity to objects and pose information. To this end, we have employed a person detection algorithm that provides a person’s position within a room. By fusing the obtained position with the objects’ position, we were able to deduce whether the person was occupied with a certain object and to draw conclusions about the performed ADLs. One prerequisite for a reliable modelling of human activities is the knowledge about the accuracy of the person detection algorithm. We have, therefore, analysed the algorithm with regard to its accuracy under different, application-specific conditions. The results show that the considered algorithm ensures high accuracy for our AAL application and that it is even suitable for environments, in which objects are very close to each other. On the basis of these findings, tests with video sequences have been conducted in an AAL environment. This evaluation confirmed that the reasoning algorithm can reliably recognise activities related to personal hygiene.

An emerging topic in the field of elderly care is the determination and tracking of vital parameters, such as the heart rate. This parameter provides important information about a person’s current health status. Within the last years, various research focussed on this topic. The recognition of vital parameters is increasingly relevant for our ageing society. This paper presents a method to remotely determine the human heart rate with a camera. At this point, we suggest to use independent component analysis (ICA) and adaptive filtering for a robust detection. In our processing chain, we used different image processing techniques, e. g. face detection, and signal processing techniques, e. g. FFT and bandpass filtering, in this study. An evaluation with several probands, illuminations, frame rates and different heart rate levels showed that we could achieve a mean error of 4.36 BPM, which corresponds to CAND of 94.45 %, and a speed of 35 fps.