Low-power Platform Research Articles

Rotating Kernel CNN Optimization for Efficient IoT Surveillance on Low-Power Devices

This study presents an innovative method to optimize convolutional neural networks (CNNs) tailored for the prevalent low-power platforms in IoT and embedded devices, which are integral to smart city infrastructures. Our approach introduces a novel “Rotating Convolutional Kernel” technique, designed to significantly reduce the computational load of CNNs while maintaining high accuracy, an essential feature for the constrained processing capabilities of devices produced by Hisilicon, MediaTek, and Novatek, among others. By leveraging the intelligent video engine (IVE) capabilities inherent in low-end CPUs, our optimized YOLOv3-Tiny model, with fewer than 100 K parameters and specifically fine-tuned for pedestrian and vehicle detection, demonstrates impressive processing speeds of 52 ms/frame on the HI3516EV200 chip and 66 ms/frame on the MSC313E chip, with minimal accuracy compromise. Despite a substantial reduction in parameter count compared to that in MobileNetV2, our model’s top-1 accuracy only slightly decreases by 2.6%, showcasing the effectiveness of our optimization technique. Our findings highlight the potential and applicability of our method in enhancing the performance and utility of IoT and embedded devices within smart cities. By achieving an optimal balance between computational efficiency and detection accuracy, our approach offers a promising avenue for advancing the capabilities of low-power devices in urban surveillance, traffic management, and other smart city applications, thereby contributing to the development of more intelligent, efficient, and responsive urban environments.

Privacy preservation in sensor-based Human Activity Recognition through autoencoders for low-power IoT devices

Human activity recognition is increasingly recognized as a key task in many applications. However, gathering data from the variety of sensors commonly available on end devices risks compromising user’s privacy when signals are transmitted to more powerful computing units for inference offloading. It is therefore important to design and implement strategies that could prevent privacy breaches without impairing the capability of the system of recognizing activity patterns, and by taking into account the energy constraints of low-power devices. In this work, we propose an energy-aware approach aimed at preserving the privacy of users during inference of human activities. The proposed method is based on a deep learning autoencoder trained to process the signal in order to remove the most sensitive information regarding privacy attributes, without significantly impacting classification accuracy. We also perform a thorough architecture’s parameter tuning of the designed system to enable its implementation on a low-power platform, which we also characterize in terms of energy expenditure. Experimental results show that this system is capable of effectively transforming the signal in order to prevent the inference of sensitive attributes (i.e. weight, height, age, and gender) and it can be conveniently implemented on a constrained embedded system at different levels of the trade-off between accuracy and energy consumption. Indeed, a complete obfuscation of sensitive attributes can be achieved at the cost of a marginal reduction in classification accuracy (5% at most), with an expenditure of around 165 mJ for an execution time of around 30ms needed during the signal transformation step.

Research on Indoor 3D Reconstruction Technology Based on Semantic Visual Simultaneous Localization and Mapping

In response to the challenge that traditional visual simultaneous localization and mapping (SLAM) systems, based on the assumption of a static environment, struggle to achieve real-time indoor 3D reconstruction in complex dynamic scenes, this paper proposes a real-time indoor 3D reconstruction algorithm based on semantic visual SLAM. By leveraging object detection to obtain 2D semantic information and providing prior information for geometric methods, the fusion of the two effectively suppresses dynamic features, reduces reliance on deep learning methods, and ensures the algorithm's real-time performance. Experimental results on dynamic scenes in the TUM RGB-D dataset show that our algorithm maintains nearly unchanged real-time performance while achieving an average performance improvement of approximately 97.56% and 97.31% on the TUM dataset and Bonn dataset, respectively, compared to the ORB-SLAM2 system. Moreover, our algorithm can reconstruct more intuitive indoor global Octo-map and semantic metric maps compared to sparse point cloud maps, effectively enhancing the scene perception capability of mobile robots and laying the foundation for performing advanced tasks. Furthermore, our algorithm demonstrates a 3.5-10.5 times improvement in real-time performance compared to other mainstream semantic SLAM systems. Experimental results on the NVIDIA Jetson AGX Xavier confirm that our algorithm can run in real time on low-power platforms such as mobile robots or drones. However, the drawbacks of our algorithm include lower reconstruction accuracy in low-texture and large-scale scenes and ineffective suppression of dynamic features in low-dynamic scenes. Future work will consider replacing and improving deep learning methods and integrating IMU and other sensors to enhance system usability.

A smart approach to EMG envelope extraction and powerful denoising for human–machine interfaces

Electromyography (EMG) is widely used in human–machine interfaces (HMIs) to measure muscle contraction by computing the EMG envelope. However, EMG is largely affected by powerline interference and motion artifacts. Boards that directly provide EMG envelope, without denoising the raw signal, are often unreliable and hinder HMIs performance. Sophisticated filtering provides high performance but is not viable when power and computational resources must be optimized. This study investigates the application of feed-forward comb (FFC) filters to remove both powerline interferences and motion artifacts from raw EMG. FFC filter and EMG envelope extractor can be implemented without computing any multiplication. This approach is particularly suitable for very low-cost, low-power platforms. The performance of the FFC filter was first demonstrated offline by corrupting clean EMG signals with powerline noise and motion artifacts. The correlation coefficients of the filtered signals envelopes and the true envelopes were greater than 0.98 and 0.94 for EMG corrupted by powerline noise and motion artifacts, respectively. Further tests on real, highly noisy EMG signals confirmed these achievements. Finally, the real-time operation of the proposed approach was successfully tested by implementation on a simple Arduino Uno board.

Acoustic receiving on glider-type autonomous underwater vehicles: Advantages and challenges

An ocean glider is a specific type of Autonomous Underwater Vehicle (AUV) that operates using a buoyancy engine rather than traditional propellors or thrusters. This method of propulsion enables a glider to be deployed for weeks or months and is relatively quiet, making the ocean glider a desirable platform for persistent acoustic monitoring. A glider typically remains submerged for several hours, during which time it can be challenging to localize the vehicle precisely. Gliders are low-power platforms that operate in the mid-water column, and therefore subsea navigation technologies implemented on other types of AUVs, such as inertial navigation systems aided by Doppler velocity logs, may not be suitable. Acoustic signals from fixed broadband sources have been used for subsea localization of Seaglider, a commercially available glider platform. Measurements of the multipath acoustic arrival structure received on Seagliders at ranges up to hundreds of kilometers from transmitting sources were used for vehicle localization in both temperate and polar underwater sound propagation environments. Methodology and results for subsea localization of the Seaglider platform will be presented and placed in the context of a broader discussion of the advantages and challenges specific to the glider as a moving acoustic receiving platform.

A Low-Power mmWave Platform for the Internet of Things

With the advancement of the Internet of Things (IoT), billions of devices will be connected to the Internet, enabling new applications such as digital twin, augmented reality, and smart home. These applications have placed a huge strain on today's wireless network. mmWave technology is promising to solve this problem by providing a large bandwidth over the very-high-frequency spectrum band. However, most mmWave radios and platforms have much higher power consumption than what IoT devices and their applications can afford. Hence, mmWave networks cannot be utilized in most IoT applications today. In this work, we present a novel low-power mmWave platform, which brings this technology to IoT applications. Our approach to design this platform is to take a holistic view and optimize the whole wireless system by considering practical challenges in mmWave communication. Our lowcost and low-power platform not only brings mmWave communication to IoT applications, but also enables researchers that do not have hardware background to work on mmWave research.

Low-Power, Multi-Transduction Nanosensor Array for Accurate Sensing of Flammable and Toxic Gases.

Toxic and flammable gases pose a major safety risk in industrial settings; thus, their portable sensing is desired, which requires sensors with fast response, low-power consumption, and accurate detection. Herein, a low-power, multi-transduction array is presented for the accurate sensing of flammable and toxic gases. Specifically, four different sensors are integrated on a micro-electro-mechanical-systems platform consisting of bridge-type microheaters. To produce distinct fingerprints for enhanced selectivity, the four sensors operate based on two different transduction mechanisms: chemiresistive and calorimetric sensing. Local, in situ synthesis routes are used to integrate nanostructured materials (ZnO, CuO, and Pt Black) for the sensors on the microheaters. The transient responses of the four sensors are fed to a convolutional neural network for real-time classification and regression of five different gases (H2 , NO2 , C2 H6 O, CO, and NH3 ). An overall classification accuracy of 97.95%, an average regression error of 14%, and a power consumption of 7mW per device are obtained. The combination of a versatile low-power platform, local integration of nanomaterials, different transduction mechanisms, and a real-time machine learning strategy presented herein helps advance the constant need to simultaneously achieve fast, low-power, and selective gas sensing of flammable and toxic gases.

FAPI-Net: A lightweight interpretable network based on feature augmentation and prototype interpretation.

With the increasing application of deep neural networks, their performance requirements in various fields are increasing. Deep neural network models with higher performance generally have a high number of parameters and computation (FLOPs, Floating Point Operations), and have the black-box characteristic. This hinders the deployment of deep neural network models on low-power platforms, as well as sustainable development in high-risk decision-making fields. However, there is little work to ensure the interpretability of the model in the research on the lightweight of the deep neural network model. This paper proposed FAPI-Net (feature augmentation and prototype interpretation), a lightweight interpretable network. It combined feature augmentation convolution blocks and the prototype dictionary interpretability (PDI) module. The feature augmentation convolution block is composed of lightweight feature-map augmentation (FA) modules and a residual connection stack. The FA module could effectively reduce network parameters and computation without losing network accuracy. The PDI module can realize the visualization of model classification reasoning. FAPI-Net is designed regarding MobileNetV3's structure, and our experiments show that the FAPI-Net is more effective than MobileNetV3 and other advanced lightweight CNNs. Params and FLOPs on the ILSVRC2012 dataset are 2 and 20% lower than that on MobileNetV3, respectively, and FAPI-Net with a trainable PDI module has almost no loss of accuracy compared with baseline models. In addition, the ablation experiment on the CIFAR-10 dataset proved the effectiveness of the FA module used in FAPI-Net. The decision reasoning visualization experiments show that FAPI-Net could make the classification decision process of specific test images transparent.

Improved Secure Encryption with Energy Optimization Using Random Permutation Pseudo Algorithm Based on Internet of Thing in Wireless Sensor Networks

The use of wireless and Internet of Things (IoT) devices is growing rapidly. Because of this expansion, nowadays, mobile apps are integrated into low-cost, low-power platforms. Low-power, inexpensive sensor nodes are used to facilitate this integration. Given that they self-organize, these systems qualify as IoT-based wireless sensor networks. WSNs have gained tremendous popularity in recent years, but they are also subject to security breaches from multiple entities. WSNs pose various challenges, such as the possibility of numerous attacks, their innate power, and their unfeasibility for use in standard security solutions. In this paper, to overcome these issues, we propose the secure encryption random permutation pseudo algorithm (SERPPA) for achieving network security and energy consumption. SERPPA contains a major entity known as a cluster head responsible for backing up and monitoring the activities of the nodes in the network. The proposed work performance is compared with other work based on secure IoT devices. The calculation metrics taken for consideration are energy, overheads, computation cost, and time consumption. The obtained results show that the proposed SERPPA is very significant in comparison to the existing works, such as GKA (Group Key Agreement) and MPKE (Multipath Key Establishment), in terms of data transfer rate, energy consumption and throughput.

A CNN-based image detector for plant leaf diseases classification

Identifying diseases from images of plant leaves is one of the most important research areas in precision agriculture. The aim of this paper is to propose an image detector embedding a resource constrained convolutional neural network (CNN) implemented in a low cost, low power platform, named OpenMV Cam H7 Plus, to perform a real-time classification of plant disease. The CNN network so obtained has been trained on two specific datasets for plant diseases detection, the ESCA-dataset and the PlantVillage-augmented dataset, and implemented in a low-power, low-cost Python programmable machine vision camera for real-time image acquisition and classification, equipped with a LCD display showing to the user the classification response in real-time. Experimental results show that this CNN-based image detector can be effectively implemented on the chosen constrained-resource system, achieving an accuracy of about 98.10%/95.24% with a very low memory cost (718.961 KB/735.727 KB) and inference time (122.969 ms/125.630 ms) tested on board for the ESCA and the PlantVillage-augmented datasets respectively, allowing the design of a portable embedded system for plant leaf diseases classification. Source files are available at https://doi.org/10.17605/OSF.IO/UCM8D.

Neural network-based prediction of the secret-key rate of quantum key distribution

Numerical methods are widely used to calculate the secure key rate of many quantum key distribution protocols in practice, but they consume many computing resources and are too time-consuming. In this work, we take the homodyne detection discrete-modulated continuous-variable quantum key distribution (CV-QKD) as an example, and construct a neural network that can quickly predict the secure key rate based on the experimental parameters and experimental results. Compared to traditional numerical methods, the speed of the neural network is improved by several orders of magnitude. Importantly, the predicted key rates are not only highly accurate but also highly likely to be secure. This allows the secure key rate of discrete-modulated CV-QKD to be extracted in real time on a low-power platform. Furthermore, our method is versatile and can be extended to quickly calculate the complex secure key rates of various other unstructured quantum key distribution protocols.

Multichannel SDR Platforms for Mobile PCL Applications

This article presents a brief overview of hardware platforms with multiple RF input channels that can be employed in the area of passive coherent location (PCL) and passive synthetic aperture radar. Two families of the platforms are considered, i.e., field-programmable gate array based development boards and general-purpose software-defined radio receivers. Practical solutions that have been validated in field experiments are described: a high-performance four-channel platform and its lightweight version that perfectly suits the needs of low-power moving installations. The preliminary results of target detections are presented, which are obtained from the signals acquired with the lightweight PCL platform. Our results prove that, besides its simplicity, the small and low-power platform can be successfully used in PCL applications that use digital audio broadcasting transmitters as the source of illumination.

Online learning for orientation estimation during translation in an insect ring attractor network

Insect neural systems are a promising source of inspiration for new navigation algorithms, especially on low size, weight, and power platforms. There have been unprecedented recent neuroscience breakthroughs with Drosophila in behavioral and neural imaging experiments as well as the mapping of detailed connectivity of neural structures. General mechanisms for learning orientation in the central complex (CX) of Drosophila have been investigated previously; however, it is unclear how these underlying mechanisms extend to cases where there is translation through an environment (beyond only rotation), which is critical for navigation in robotic systems. Here, we develop a CX neural connectivity-constrained model that performs sensor fusion, as well as unsupervised learning of visual features for path integration; we demonstrate the viability of this circuit for use in robotic systems in simulated and physical environments. Furthermore, we propose a theoretical understanding of how distributed online unsupervised network weight modification can be leveraged for learning in a trajectory through an environment by minimizing orientation estimation error. Overall, our results may enable a new class of CX-derived low power robotic navigation algorithms and lead to testable predictions to inform future neuroscience experiments.

CED-Net: A more effective DenseNet model with channel enhancement.

In recent years, deep convolutional neural network (CNN) has been applied more and more increasingly used in computer vision, natural language processing and other fields. At the same time, low-power platforms have more and more significant requirements for the size of the network. This paper proposed CED-Net (Channel enhancement DenseNet), a more efficient densely connected network. It combined the bottleneck layer with learned group convolution and channel enhancement module. The bottleneck layer with learned group convolution could effectively increase the network's accuracy without too many extra parameters and computation (FLOPs, Floating Point Operations). The channel enhancement module improved the representation of the network by increasing the interdependency between convolutional feature channels. CED-Net is designed regarding CondenseNet's structure, and our experiments show that the CED-Net is more effective than CondenseNet and other advanced lightweight CNNs. Accuracy on the CIFAR-10 dataset and CIFAR-100 dataset is 0.4 and 1% higher than that on CondenseNet, respectively, but they have almost the same number of parameters and FLOPs. Finally, the ablation experiment proves the effectiveness of the bottleneck layer used in CED-Net.

A Lightweight CNN-Based Vision System for Concrete Crack Detection on a Low-Power Embedded Microcontroller Platform

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.

Optical Microelectromechanical Systems Technologies for Spectrally Adaptive Sensing and Imaging

AbstractEvolving from past black‐and‐white images, through present red‐green‐blue spectral colors, future remote imaging systems promise spectroscopic functionalities extending well beyond the visible wavelengths. This allows real‐time spectral information to be gathered from multiple wavelength bands that is applicable to numerous remote sensing spectroscopy/imaging applications and aids target recognition. This paper reviews the wavelength tunable microelectromechanical systems (MEMS) optical filter technologies developed for the important infrared and the emerging terahertz wavelength bands of the electromagnetic spectrum with the fabrication effort being enabled by the Western Australian Node of the Australian National Fabrication Facility. A low size, weight and power (SWaP) platform solution is demonstrated delivering mechanically robust, field‐portable, spectroscopic, chem/bio sensing suitable for deployment in remote sensing and imaging applications.

A Low-Cost, Low-Power and Real-Time Image Detector for Grape Leaf Esca Disease Based on a Compressed CNN

Esca is one of the most common grape leaf diseases that seriously affect grape yield, causing a loss of global production in the range of 20%–40%. Therefore, a timely and effective identification of the disease could help to develop an early treatment approach to control its spread while reducing economic losses. For this purpose the use of computer vision and machine learning techniques for recognizing plant diseases have been extensively studied in recent years. The aim of this paper is to propose an image detector based on a high-performance convolutional neural network (CNN) implemented in a low cost, low power platform, to monitor the Esca disease in real-time. To meet the severe constraints typical of an embedded system, a new low-rank CNN architecture (LR-Net) based on CANDECOMP/PARAFAC (CP) tensor decomposition has been developed. The compressed CNN network so obtained has been trained on a specific dataset and implemented in a low-power, low-cost Python programmable machine vision camera for real-time classification. An extensive experimentation has been conducted and the results achieved show the superiority of LR-Net with respect to the state-of-the-art networks both in terms of inference time and memory occupancy.

The impact and resolution of the GPS week number rollover of April 2019 on autonomous geophysical instrument platforms

Abstract. Instrument platforms the world over often rely on GPS or similar satellite constellations for accurate timekeeping and synchronization. This reliance can create problems when the timekeeping counter aboard a satellite overflows and begins a new epoch. Due to the rarity of these events (19.6 years for GPS), software designers may be unaware of such circumstance or may choose to ignore it for development complexity considerations. Although it is impossible to predict every fault that may occur in a complicated system, there are a few “best practices” that can allow for graceful fault recovery and restorative action. These guiding principles are especially pertinent for instrument platforms operating in space or in remote locations like Antarctica, where restorative maintenance is both difficult and expensive. In this work, we describe how these principles apply to a communications failure on autonomous adaptive low-power instrument platforms (AAL-PIP) deployed in Antarctica. In particular, we describe how code execution patterns were subtly altered after the GPS week number rollover of April 2019, how this led to Iridium satellite communications and data collection failures, and how communications and data collection were ultimately restored. Finally, we offer some core tenets of instrument platform design as guidance for future development.

A Prim–Dijkstra Algorithm for Multihop Calibration of Networked Embedded Systems

The development of large-scale systems of networked embedded devices with sensing capabilities relies on the availability of low-cost and resource-constrained components However, the reduced precision and accuracy of low-cost sensors on board of low-power platforms risks to impair the overall reliability of these systems, thus preventing their potential diffusion, especially in deployments with several (e.g., hundreds or more) nodes. Hence, ensuring the required quality of measurements along the lifetime of a sensor network represents a key challenge, which is often tackled also by means of calibration techniques. In this article, we propose a novel approach to multihop calibration, targeting the derivation of a spanning tree that encompasses the optimization of a biobjective problem. Indeed, since minimum spanning trees can be related to the energy budget of a network and shortest path trees can be used as a model for the minimization of the cumulative calibration errors, the search for a spanning tree that simultaneously optimizes the two metrics represents a useful direction toward the design of effective and efficient calibration strategies. To this aim, we introduce a method based on the Prim-Dijkstra algorithm, which represents an effective heuristics for effective search of solutions that could represent a tradeoff between the accuracy of multihop calibration and the energy expenditure needed to calibrate sensors. The proposed approach allows fast derivation of different solutions by means of a single parameter, thus enabling the efficient exploration of the design space even in large-scale scenarios as confirmed by numerical results obtained for validation.

Chimera

The Internet of Things (IoT) is being deployed in an ever-growing range of applications, from industrial monitoring to smart buildings to wearable devices. Each of these applications has specific computational requirements arising from their networking, system security, and edge analytics functionality. This diversity in requirements motivates the need for adaptable end-devices, which can be re-configured and re-used throughout their lifetime to handle computation-intensive tasks without sacrificing battery lifetime. To tackle this problem, this article presents Chimera, a low-power platform for research and experimentation with reconfigurable hardware for the IoT end-devices. Chimera achieves flexibility and re-usability through an architecture based on a Flash Field Programmable Gate Array (FPGA) with a reconfigurable software stack that enables over-the-air hardware and software evolution at runtime. This adaptability enables low-cost hardware/software upgrades on the end-devices and an increased ability to handle computationally-intensive tasks. This article describes the design of the Chimera hardware platform and software stack, evaluates it through three application scenarios, and reviews the factors that have thus far prevented FPGAs from being utilized in IoT end-devices.