Many studies have shown the significant benefits of using specialized hardware, such as Graphics Processing Units (GPUs), to accelerate Deep Learning tasks that require efficient large-scale data processing while maintaining high accuracy. Numerous software tools also exist that simplify model deployment. Given the large variety of hardware and software alternatives available, achieving optimal performance requires selecting and adapting the deployment to best suit each application. As highlighted in this article, performance depends not only on the hardware itself but also on the efficient design and implementation of the underlying code, to fully leverage available computational resources. To assess the impact of different implementation strategies on Deep Learning models and to better understand the tradeoffs between implementation complexity and application requirements, this work analyzes and compares two different approaches for deploying the EfficientAD visual anomaly detection model on GPUs. Notably, by increasing software development complexity to a limited and affordable extent, performance can be largely improved, achieving around 60% reduction in total execution latency.

Nowadays segmentation Convolutional Neural Networks (CNNs) are used in industrial quality control systems to accurately measure the shape and size of production defects. However, these models are quite big and require quite powerful computing systems to be deployed in real applications. In order to obtain fast, power-efficient, and affordable implementations of segmentation CNNs for industrial quality control, in this paper execution time, memory usage, and power consumption measurements are carried out in different modern embedded GPU SoC devices, using different optimization strategies. Results show that segmentation CNNs can be deployed in embedded devices using FP16 or INT8 quantization with negligible accuracy loss and high throughput, up to 240fps.

Field-programmable system-on-chip (FPSoC) devices, combining high-performance processors and FPGA fabric in the same chip, are currently a leading technology in the design of complex digital systems. Since design times are longer than those of systems based on graphic processing units or standalone processors, many efforts are being devoted to develope efficient compilers from high-level languages. Even though, efficient processor-FPGA communication is still an important open issue. To contribute to this area, this article presents an extensive characterization of the processor-FPGA communication delays in Zynq-7000 devices. Although partial analyses of communication performance in these devices have been reported, this is the first work to address very important issues such as the use of DMA for data transfers or the effect of L2 cache controller settings and external RAM controller settings. As a result, data transfer rates are analyzed considering all parameters that influence them. The performance of Zynq-7000 devices is also compared to that of Cyclone V devices, hence covering the two most important current families that dominate the FPSoC market. This information is of utmost importance for designers to optimize processor-FPGA communication and, in turn, the performance of their FPSoC-based systems.

Fast Deep Neural Network (DNN) inference is nowadays possible thanks to the recent significant advancements in computing platforms and DNN frameworks. The advent of the Internet of Things, among other factors, leads to the massive deployment of edge computing devices, whose features and performance are very diverse. In order to help designers identify the hardware platforms and DNNs that best suit their target embedded applications, this paper presents a DNN inference performance analysis for state-of-the-art edge devices. These include Graphical Processing Units, Tensor Processing Units and Field-Programmable Systems-on-Chip. Different versions of the MobileNet and Inception DNNs and different frameworks are considered in the analysis.

This paper presents a case study and the underlying methodology of controller design for nonlinear multiple input multiple output systems. Deep neural networks are used to represent the controller policy and a reinforcement learning technique, namely Double Deep Q-Learning, is used to train it. The controller is designed from a simplified virtual model of the system and directly ported to the real system. This frees the designer from extracting an accurate model of the system (classic control theory approach), translating expert knowledge (fuzzy controllers), or acquiring large sets of data (supervised training of neural networks). Without loss of generality, the case study corresponds to the design of a 2-input 2-output line follower controller for unmanned ground vehicles (UGVs). Results show that the virtually-designed controller can drive real UGVs through the target trajectories with deviations within acceptable margins.

This paper presents a generic method to process mid-wave infrared (MWIR) images in laser-based manufacturing processes. The background noise of the camera is used as a source of information for correcting different problems that affect MWIR cameras. The mean of the noise distribution is used to correct the background drift due to sensor heat-up, whereas the standard deviation is used to generate dynamic thresholds for subsequent algorithms. The proposed method is portable, robust, and independent of the background, scale, and optical and electronic aberrations of the camera. The method is validated in the calculation of melt pool cooling rate. Performance is tested in a real scenario using a Field Programmable System-on-Chip platform. Results show that the system is capable of sending to a remote computer MWIR images at 1,435 frames per second and cooling rate information at a rate of 10,680 samples per second.

This paper presents a camera-based localization system for controlling unmanned ground vehicles in a 2D space using Field Programmable System-on-Chip devices, whose architecture combines powerful hard processors and FPGA fabric in the same chip. The FPGA captures and preprocesses images in real time while the hard processor is in charge of most image processing tasks. In these architectures, efficient data exchange between the hardware and software subsystems is a key feature to achieve high performance. In this case, versatile image writers have been designed to efficiently share images between the FPGA and the processor. The proposed free-software solution has been tested in Cyclone V devices, and compared with a previous pure hardware version of the system, as well as with other localization systems. Experimental results demonstrate that this system achieves better accuracy and higher frame rate than non-commercial localization systems in the literature, while being less expensive than commercial (closed) ones.

Atrial fibrillation (AF) is the most prevalent type of cardiac arrhythmia. As such, its early detection at a massive population scale is of paramount importance not only to avoid serious risks to people's health but also to ensure the sustainability of health care systems worldwide. Ambulatory monitoring can help mitigate these problems, and wearable devices can play a fundamental role in this regard. These devices can be equipped with sensors to measure cardiac activity, for instance from photoplethysmography (PPG) signals. The use of these sensors in wearable devices for AF screening faces several technological challenges. One of them is their sensitivity to interferences, in particular caused by patient's physical activity, which may be indistinguishable from cardiac activity. This paper presents a method to identify the periods of time during which PPG signals can be used because these effects are not present in them. Experimental results obtained from real patients show that by using the proposed method, a significant amount of valid data can be obtained, improving the usability of wearable devices for affordable AF ambulatory screening.

Decreasing profit margins and increasing concerns about animal welfare are boosting the interest for the development of monitoring and analysis technologies specifically targeting the poultry meat production process. In this context, this paper addresses monitoring of poultry activity in breeding farms. Specifically, it analyzes the suitability of different vision systems and image processing algorithms with this purpose. These systems and algorithms have been tested in an actual farm during a breeding cycle. Experimental results are presented demonstrating that density-based computations provide the best results, and that they can be carried out using either video or thermographic images, but the latter are a better option because of practical operating reasons related to varying, low light intensity conditions.

Education in industrial electronics-related subjects is very challenging because of the many different technological topics and application domains to be addressed. Mobile robotics is a very interesting area in this context, because it requires a wide coverage of many diverse topics and, at the same time, it is typically very appealing for engineering students. Therefore, mobile robot systems are very suitable platforms for the development of Project Based Learning solutions. This paper presents UviSpace, a modular, open-source intelligent space with Unmanned Ground Vehicles, which is being successfully applied for hands-on training of students in real-world problem solving in industrial electronics-related subjects. The hardware and software modules of the system are described, and student projects are discussed to demonstrate the usefulness and versatility of UviSpace as educational platform.

Decreasing profit margins and increasing concerns about animal welfare are boosting the interest for the development of monitoring and analysis technologies specifically targeting the poultry meat production process. In this context, this paper addresses one of the most critical steps of such process, the transportation of newborn animals from birth places to breeding farms. Specifically, an intelligent sensor system to monitor ventilation in truck trailers has been designed and validated. Experimental results obtained both in the lab and during real transportation trips are presented to demonstrate the suitability of the proposed system for the target application, in terms of cost, size, power consumption, and ability to operate at low air speeds.

FPGAs have evolved from hardware accelerators to very powerful System-on-Chip platforms, mainly thanks to the availability of increasingly powerful embedded processors. The efficient integration of embedded processors with the FPGA fabric requires fast exchange of information between both sides, not to compromise the performance improvement these architectures should provide. Therefore, detailed performance analyses of different FPGA-processor communication scenarios are needed for designers to be able to evaluate the suitability of current devices for a given application, and to take the most advantage of them in it. This paper presents, to the best of authors' knowledge, the first reported characterization of FPGA-processor communications in Altera Cyclone V SoC devices. In this initial step towards a comprehensive study, the performance of communication mechanisms where the processor acts as master is analyzed.

The ability to perform accurate localization is a fundamental requirement of the navigation systems intended to guide unmanned ground vehicles in a given environment. Currently, the use of vision-based systems is a very suitable alternative for some indoor applications. This paper presents a novel distributed FPGA-based embedded image processing system for accurate and fast simultaneous estimation of the position and orientation of remotely controlled vehicles in indoor spaces. It is based on a network of distributed image processing nodes, which minimize the amount of data to be transmitted through communication networks and hence allow dynamic response to be improved, providing a simple, flexible, low-cost, and very efficient solution. The proposed system works properly under variable or nonhomogeneous illumination conditions, which simplifies the deployment. Experimental results on a real scenario are presented and discussed. They demonstrate that the system clearly outperforms the existing solutions of similar complexity. Only much more complex and expensive systems achieve similar performance.

Currently, frequency is the physical variable that can be measured with the best accuracy. Therefore, sensors based on frequency measurements are used in many industrial applications. QCM sensors are a kind of resonator-based sensors that allow mass variations in the nanogram-picogram range to be detected. This paper presents a high-accuracy frequency meter system for QCM sensors based on a Field-Programmable System-on-Chip architecture. A high-performance custom hardware peripheral integrated in a Nios II-based system has been implemented and experimentally validated. The obtained results demonstrate the suitability of the proposed system for the target application, as well as its potential to be extended to others. Some limitations and possible ways of improvement have also been identified.

Laser cladding is a material processing technique widely used in industry for part reconstruction, coating or rapid prototyping, among other fields. When optical systems are used as sensing elements to measure the cladding area, malfunctions can occur because of interferences of suspended powder or damages to the cameras caused by laser beams. It is therefore important to provide means to detect problems affecting the optical system and, if possible, extend its life cycle by keeping correct performance even when damaged. In previous works, an FPGA-based measurement and control system for laser cladding has been proposed, which significantly outperforms other solutions used in industry, allowing some of their limitations to be overcome. This paper presents an improved FPGA system, capable of providing correct operation in the presence of suspended powder or (to some extent) other defects affecting the optical system. Experimental results obtained in an industrial part reconstruction system are presented that demonstrate the claimed contributions of the work.

This paper describes the design of an FPGA-based SoC system intended to measure the mass resolution of quartz crystal microbalance (QCM) sensors. The implemented system integrates the frequency measurement and the computation of the sensor resolution in a single FPGA to configure a standalone measurement system. A Virtex-4 FPGA from Xilinx was chosen as hardware platform due to its suitable architecture to support SoC and DSP applications. The developed measurement system combines the 32-bit embedded MicroBlazereg soft processor with a specific hardware coprocessor designed to accelerate the computations required for real-time processing. A data flow based on techniques as recursive formulas and jump prediction was proposed in order to reduce the computations and memory requirements of the coprocessor. Finally, the implemented system was validated with a 9 MHz QCM sensor in damping media used for electrochemical applications and the results were compared with a full floating point implementation