Developing efficient on-the-edge Deep Learning (DL) applications is a challenging and non-trivial task, as first different DL models need to be explored with different trade-offs between accuracy and complexity, second, various optimization options, frameworks and libraries are available that need to be explored, third, a wide range of edge devices are available with different computation and memory constraints. As such, trade-offs arise among inference time, energy consumption, efficiency (throughput/watt) and value (throughput/dollar). To shed some light in this problem, a case study is delivered where seven Image Classification (IC) and six Object Detection (OD) State-of-The-Art (SOTA) DL models were used to detect face masks on the following commercial off-the-shelf edge devices: Raspberry PI 4, Intel Neural Compute Stick 2, Jetson Nano, Jetson Xavier NX, and i.MX 8M Plus. First, a full end-toend video pipeline face mask wearing detection architecture is developed. Then, the thirteen DL models were optimized, evaluated and compared on the edge devices, in terms of accuracy and inference time. To leverage the computational power of the edge devices, the models have been optimized, first, by using the SOTA optimization frameworks (TensorFlow Lite, OpenVINO, TensorRT, eIQ) and, second, by evaluating/comparing different optimization options, e.g., different levels of quantization. Note that the five edge devices are evaluated and compared too, in terms of inference time, value and efficiency. Last, we obtain insightful observations on which optimization frameworks, libraries and options to use and on how to select the right device depending on the target metric (inference time, efficiency and value). For example, we show that Jetson Xavier NX platform is the best in terms of latency and efficiency (FPS/Watt), while Jetson Nano is the best in terms of value (FPS/$).

No abstract

Detecting vital signs by using a contactless camera-based approach can provide several advantages over traditional clinical methods, such as lower financial costs, reduced visit times, increased comfort, and enhanced safety for healthcare professionals. Specifically, Eulerian Video Magnification (EVM) or Remote Photoplethysmography (rPPG) methods can be utilised to remotely estimate heart rate and respiratory rate biomarkers. In this paper two contactless camera-based health monitoring architectures are developed using EVM and rPPG, respectively; to this end, two different CNNs, (Mediapipe’s BlazeFace and FaceMesh) are used to extract suitable regions of interest from incoming video frames. These two methods are implemented and deployed on four off-the-shelf edge devices as well as on a PC and evaluated in terms of latency (in each stage of the application’s pipeline), throughput (FPS), power consumption (Watt), efficiency (throughput/Watt), and value (throughput/cost). This work provides important insights about the computational costs and bottlenecks of each method on each hardware platform, as well as which platform to use depending on the target metric. One of our insights shows that the Jetson Xavier NX platform is the best platform in terms of throughput and efficiency, while Raspberry Pi 4 8 GB is the best platform in terms of value.

Various methods of statistical and harmonic infor mation analysis are widely used in medical diagnosis. These methods include cardiac rhythm variability (CRV) analysis methods and spectral methods [1 4]. CRV meth ods provide integral evaluation of the state of mechanisms of regulation of physiological functions of the body. For example, the CRV method provides analysis of general activity of regulatory mechanisms, neurohumoral regula tion of cardiac activity, and sympathetic and parasympa thetic nervous systems. Spectral analysis of cardiac rhythm provides useful information about frequency and energy characteristics of the cardiovascular system. Because Fourier transform is widely used for detecting periodic components of complex signals, the quality of evaluation of these parameters depends on periodic fac tors including CRV.The goal of this work was to suggest a method for preliminary processing of sphygmographic data. This can be used to evaluate the effect of CRV on reliability of spectral analysis of the cardiac cycle. Let the model of sphygmograph signal A(t) be represented as a sequence of functions A i (ϕ i ). Each function is determined within time interval 0 T i (ϕ i ranges from 0 to T i ), where T i = _ T + ∆t i is duration of i th period of cardiac rhythm; _ T is mean dura tion of cardiac rhythm period; ∆t i is deviation from mean duration of i th period of cardiac rhythm. If the process is periodic (∆t i = 0 and A 1 (ϕ 1 ) = A i (ϕ ι ) for each i value), A(t) is a sum of harmonics:where A 0 , A n are amplitudes of harmonics; ψ n are initial phases of harmonics; n is number of harmonic. Discrete spectrum and low contribution of higher harmonics allow sphygmograph signal to be represented as finite number of Fourier series components.In the presence of CRV, A(t) function is aperiodic. The spectrum of an aperiodic function is generally a con tinuous function. Fourier transform fails to provide reli able frequency and energy information about such spec tral functions. Deterioration of spectral analysis quality is due to broadening of spectral bands and overlap between wings of neighboring bands.To reduce the effect of CRV on broadening of spec tral bands and absolute values of harmonic amplitudes, the method of variability elimination was used. Spectra of functions A i (ϕ i ) were suggested to contain the same har monics. This suggestion was supported by the results of analysis of experimental data. It follows from this analysis that mean value of correlation coefficient between A i (ϕ i ) and A j (ϕ j ) (i ≠ j) for 60 periods of cardiac rhythm was 0.8. Preliminary analysis showed that mean value of correla tion coefficient between spectral functions A i (ϕ i ) and A j (ϕ j ) was 0.9.The method of variability elimination is implement ed as follows. The sequence A 1 (ϕ 1 ) is cut from initial sig nal A(t) using time window ∆t = t 2 -t 1 (time interval between maximal amplitudes of sphygmograms of the first and second cardiac pulses):The point t 1 corresponds to the last local maximum (point В in Fig. 1) ...

Spectroscopic sensor imaging of food samples meta-processed by deep machine learning models can be used to assess the quality of the sample. This article presents an architecture for estimating microbial populations in meat samples using multispectral imaging and deep convolutional neural networks. The deep learning models operate on embedded platforms and not offline on a separate computer or a cloud server. Different storage conditions of the meat samples were used, and various deep learning models and embedded platforms were evaluated. In addition, the hardware boards were evaluated in terms of latency, throughput, efficiency and value on different data pre-processing and imaging-type setups. The experimental results showed the advantage of the XavierNX platform in terms of latency and throughput and the advantage of Nano and RP4 in terms of efficiency and value, respectively.