Neural Compute Stick Research Articles

Development of intelligent suite for malaria pathogen detection in microscopy images

The identification of malaria infection using microscope images of blood smears is considered as a ‘gold standard’. The diagnosis of malaria needs expert microscopists which are scarce in remote areas where malaria is endemic. Therefore, it is desirable to automate the repetitive task of pathogen detection in the blood samples received as microscope images. This study provides an easy to use and deploy method for implementing a malaria pathogen detection software- the Intelligent Suite. The Intelligent Suite features a graphical user interface (GUI) implemented using ‘cvui’ library to interact with the OpenVINO’s inference engine for model optimisation and deployment across several inference devices. The intelligent Suite uses a custom YOLO-mp-3l model trained on Darknet framework for detection of malaria pathogen in thick smear microscope images. Moreover, the Intelligent Suite provides user interface for inference device/mode selection, alter model parameters, and generate detection reports along with the model performance metrics. The Intelligent Suite was executed on a CPU computer with model inference running on a plug-and-play Neural Compute Stick (NCS2) and performance reported.

Single-Event Effects on Commercial-Off-the-Shelf Edge-Processing Artificial Intelligence ASICs

Single-event effects test results are presented for heavy ions and high-energy protons on two commercial edge-processing artificial intelligence application-specific integrated circuits, particularly the Google Coral Accelerator Module, which employs a Google Coral Edge Tensor Processing Unit, and the Intel Neural Compute Stick 2, which employs the Intel Movidius Myriad X. The observed SEE signatures, probable causes within the devices, and single-event functional interrupt cross-sections are discussed. Ultimately, when SEUs or SEFIs are observed, there are negligible changes in the measured precision or accuracy of the algorithm, and rarely do the errors result in incorrect identification or classification.

Catalysing assistive solutions by deploying light-weight deep learning model on edge devices

ABSTRACT Nowadays, real-time object detection, which is a crucial task, is being performed through image processing and deep learning techniques. As there are several high-performance computing edge devices available, selecting the best-fit device for a particular problem is a tough task and keeping in mind the cost, performance, and weight of the device in mind. One faces several challenges while performing this task in real-time such as a lack of resources in terms of power and mobility. We have provided an insight into the computation power of devices in terms of Frames per Second (FPS) by deploying object detection models on them. This paper will provide insight into selecting the appropriate combination of device and object detection models for real-time applications. Raspberry Pi 3 (RPi3), Raspberry Pi 4 (RPi4), Intel Neural Compute Stick 2 (NCS2), and Nvidia Jetson NANO are popular devices with high computation power used for real-time applications. The memory constraints of devices along with the deployment of different You Only Look Once (YOLO) and Single-Shot Detector (SSD) are the two object detection models that have been explained in this paper. A deep learning inference optimiser, TensorRT, has been used in NANO to achieve high throughput in the performance of object detection. The precision, recall, and F1 score achieved on deploying each tested model have been presented. After observing the devices during experimentation, RPi4+NCS2 showed the best execution with the blend of factors i.e. speed, portability, and user-friendliness.

Robust Fastener Detection Based on Force and Vision Algorithms in Robotic (Un)Screwing Applications.

This article addresses how to tackle one of the most demanding tasks in manufacturing and industrial maintenance sectors: using robots with a novel and robust solution to detect the fastener and its rotation in (un)screwing tasks over parallel surfaces with respect to the tool. To this end, the vision system is based on an industrial camera with a dynamic exposure time, a tunable liquid crystal lens (TLCL), and active near-infrared reflectance (NIR) illumination. Its camera parameters, combined with a fixed value of working distance (WD) and variable or constant field of view (FOV), make it possible to work with a variety of fastener sizes under several lighting conditions. This development also uses a collaborative robot with an embedded force sensor to verify the success of the fastener localization in a real test. Robust algorithms based on segmentation neural networks (SNN) and vision were developed to find the center and rotation of the hexagon fastener in a flawless condition and worn, scratched, and rusty conditions. SNNs were tested using a graphics processing unit (GPU), central processing unit (CPU), and edge devices, such as Jetson Javier Nx (JJNX), Intel Neural Compute Stick 2 (INCS2), and M.2 Accelerator with Dual Edge TPU (DETPU), with optimization parameters, such as the unsigned integer (UINT) and float (FP), to understand their performance. A virtual program logic controller (PLC) was mounted on a personal computer (PC) as the main control to process the images and save the data. Moreover, a mathematical analysis based on the international standard organization (ISO) and patents of the manual socket wrench was performed to determine the maximum error allowed. In addition, the work was substantiated using exhaustive evaluation tests, validating the tolerance errors, robotic forces for successfully completed tasks, and algorithms implemented. As a result of this work, the translation tolerances increase with higher sizes of fasteners from 0.75 for M6 to 2.50 for M24; however, the rotation decreases with the size from 5.5° for M6 to 3.5° for M24. The proposed methodology is a robust solution to tackle outliers contours and fake vertices produced by distorted masks present in non-constant illumination; it can reach an average accuracy to detect the vertices of 99.86% and the center of 100%, also, the time consumed by the SNN and the proposed algorithms is 73.91 ms on an Intel Core I9 CPU. This work is an interesting contribution to industrial robotics and improves current applications.

Implementation of a Modified U-Net for Medical Image Segmentation on Edge Devices

Deep learning techniques, particularly convolutional neural networks, have shown great potential in computer vision and medical imaging applications. However, deep learning models are computationally demanding as they require enormous computational power and specialized processing hardware for model training. To make these models portable and compatible for prototyping, their implementation on low-power devices is imperative. In this work, we present the implementation of Modified U-Net on Intel Movidius Neural Compute Stick 2 (NCS-2) for the segmentation of medical images. We selected U-Net because, in medical image segmentation, U-Net is a prominent model that provides improved performance for medical image segmentation even if the dataset size is small. The modified U-Net model is evaluated for performance in terms of dice score. Experiments are reported for segmentation task on three medical imaging datasets: BraTs dataset of brain MRI, heart MRI dataset, and Ziehl-Neelsen sputum smear microscopy image (ZNSDB) dataset. For the proposed model, we reduced the number of parameters from 30 million in the U-Net model to 0.49 million in the proposed architecture. Experimental results show that the modified U-Net provides comparable performance while requiring significantly lower resources and provides inference on the NCS-2. The maximum dice scores recorded are 0.96 for the BraTs dataset, 0.94 for the heart MRI dataset, and 0.74 for the ZNSDB dataset.

Edge of Medical Things Implementation for Deep Learning-Based Cognitive Task Recognition

The emergence of edge devices for the Internet of Medical Things (IoMT) has enabled the integration of low-power resource-limited hardware units with communication technologies for smart healthcare applications including, brain-computer interface (BCI) and mental health management. Implementation of BCI systems on edge devices can provide high clinical decision accuracy with low latency and power consumption while reducing cloud-based training and communication. Hence, this article presents the edge of medical things (EoMT) implementation for cognitive task recognition (CTR) from electroencephalogram (EEG) signals in portable BCI systems. The proposed implementation exploits deep learning-based CTR on different computing platforms for edge devices including Arduino Due and Portenta H7 micro-controllers, Intel Movidius Neural Compute Stick 2 (NCS2), and Mimas v2 Spartan 6 field-programmable gate array (FPGA) board, along with IoMT communication technologies toward transmission of CTR information to caregivers. The real-time CTR performance results on different public and self-recorded databases indicate that the EoMT implementation scheme achieves high accuracies of 99.30 percent and 99.02 percent for public databases, and 96 percent and 89.50 percent for self-recorded databases, with low power consumption and latency, which makes it suitable for the IoMT context.

FastStereoNet: A Fast Neural Architecture Search for Improving the Inference of Disparity Estimation on Resource-Limited Platforms

Convolutional neural networks (CNNs) provide the best accuracy for disparity estimation. However, CNNs are computationally expensive, making them unfavorable for resource-limited devices with real-time constraints. Recent advances in neural architectures search (NAS) promise opportunities in automated optimization for disparity estimation. However, the main challenge of the NAS methods is the significant amount of computing time to explore a vast search space [e.g., <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.6\times 10^{29}$ </tex-math></inline-formula> ] and costly training candidates. To reduce the NAS computational demand, many proxy-based NAS methods have been proposed. Despite their success, most of them are designed for comparatively small-scale learning tasks. In this article, we propose a fast NAS method, called FastStereoNet, to enable resource-aware NAS within an intractably large search space. FastStereoNet automatically searches for hardware-friendly CNN architectures based on late acceptance hill climbing (LAHC), followed by simulated annealing (SA). FastStereoNet also employs a fine-tuning with a transferred weights mechanism to improve the convergence of the search process. The collection of these ideas provides competitive results in terms of search time and strikes a balance between accuracy and efficiency. Compared to the state of the art, FastStereoNet provides <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$5.25\times $ </tex-math></inline-formula> reduction in search time and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$44.4\times $ </tex-math></inline-formula> reduction in model size. These benefits are attained while yielding a comparable accuracy that enables seamless deployment of disparity estimation on resource-limited devices. Finally, FastStereoNet significantly improves the perception quality of disparity estimation deployed on field-programmable gate array and Intel Neural Compute Stick 2 accelerator in a significantly less onerous manner.

Intelligent Real-Time Face-Mask Detection System with Hardware Acceleration for COVID-19 Mitigation.

This paper proposes and implements a dedicated hardware accelerated real-time face-mask detection system using deep learning (DL). The proposed face-mask detection model (MaskDetect) was benchmarked on three embedded platforms: Raspberry PI 4B with either Google Coral USB TPU or Intel Neural Compute Stick 2 VPU, and NVIDIA Jetson Nano. The MaskDetect was independently quantised and optimised for each hardware accelerated implementation. An ablation study was carried out on the proposed model and its quantised implementations on the embedded hardware configurations above as a comparison to other popular transfer-learning models, such as VGG16, ResNet-50V2, and InceptionV3, which are compatible with these acceleration hardware platforms. The ablation study revealed that MaskDetect achieved excellent average face-mask detection performance with accuracy above 94% across all embedded platforms except for Coral, which achieved an average accuracy of nearly 90%. With respect to detection performance (accuracy), inference speed (frames per second (FPS)), and product cost, the ablation study revealed that implementation on Jetson Nano is the best choice for real-time face-mask detection. It achieved 94.2% detection accuracy and twice greater FPS when compared to its desktop hardware counterpart.

Few-Shot User-Definable Radar-Based Hand Gesture Recognition at the Edge

Technological advances and scalability are leading Human-Computer Interaction (HCI) to evolve towards intuitive forms, such as through gesture recognition. Among the various interaction strategies, radar-based recognition is emerging as a touchless, privacy-secure, and versatile solution in different environmental conditions. Classical radar-based gesture HCI solutions involve deep learning but require training on large and varied datasets to achieve robust prediction. Innovative self-learning algorithms can help tackling this problem by recognizing patterns and adapt from similar contexts. Yet, such approaches are often computationally expensive and hardly integrable into hardware-constrained solutions. In this paper, we present a gesture recognition algorithm which is easily adaptable to new users and contexts. We exploit an optimization-based meta-learning approach to enable gesture recognition in learning sequences. This method targets at learning the best possible initialization of the model parameters, simplifying training on new contexts when small amounts of data are available. The reduction in computational cost is achieved by processing the radar sensed data of gestures in the form of time maps, to minimize the input data size. This approach enables the adaptation of simple convolutional neural network (CNN) to new hand poses, thus easing the integration of the model into a hardware-constrained platform. Moreover, the use of a Variational Autoencoders (VAE) to reduce the gestures’ dimensionality leads to a model size decrease of an order of magnitude and to half of the required adaptation time. The proposed framework, deployed on the Intel <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">®</sup> Neural Compute Stick 2 (NCS 2), leads to an average accuracy of around 84% for unseen gestures when only one example per class is utilized at training time. The accuracy increases up to 92.6% and 94.2% when three and five samples per class are used.

A Smart System for Personal Protective Equipment Detection in Industrial Environments Based on Deep Learning at the Edge

Real-time object detection is currently used to automate various tasks in industrial environments. One of the most important tasks is to improve the safety of workers by monitoring the correct use of Personal Protective Equipment (PPE) in dangerous areas. In this context, usually, a monitoring system analyzes the stream of videos from surveillance cameras to assess PPE usage in real time. When a worker not wearing the appropriate PPE is detected, an acoustic or visual alarm is triggered automatically to raise attention and awareness. The solutions proposed so far are mostly cloud-based systems: images from the site are continuously offloaded to the cloud for analysis. This centralized architecture requires significant network bandwidth to transmit the video feeds through an internet connection that must be reliable, as a network outage would disrupt the service. In this work, we propose a system for real-time PPE detection based on video streaming analysis and Deep Neural Network (DNN). We adopt the edge computing model in which the application for image analysis and classification is deployed on an embedded system installed in proximity of the camera and directly connected to it. The system does not require continuous image transmission towards a cloud system, thus ensuring bandwidth efficiency, reliability, and workers’ privacy. A prototype of the proposed system is developed exploiting a low-cost commercial embedded system, i.e. a Raspberry PI, equipped with an Intel Neural Compute Stick 2. We tested the system with five different pre-trained convolutional neural networks (CNNs), fine-tuned to detect different PPEs, namely helmets, vests, and gloves. In our experimental evaluation, we first compared the five CNNs in terms of classification performance and inference latency. Then, we deployed each CNN on the real system and evaluated the system’s throughput regarding the number of video frames analyzed each second.

Anatomy of Deep Learning Image Classification and Object Detection on Commercial Edge Devices: A Case Study on Face Mask Detection

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-to-end 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/$).

Ag-YOLO: A Real-Time Low-Cost Detector for Precise Spraying With Case Study of Palms.

To date, unmanned aerial vehicles (UAVs), commonly known as drones, have been widely used in precision agriculture (PA) for crop monitoring and crop spraying, allowing farmers to increase the efficiency of the farming process, meanwhile reducing environmental impact. However, to spray pesticides effectively and safely to the trees in small fields or rugged environments, such as mountain areas, is still an open question. To bridge this gap, in this study, an onboard computer vision (CV) component for UAVs is developed. The system is low-cost, flexible, and energy-effective. It consists of two parts, the hardware part is an Intel Neural Compute Stick 2 (NCS2), and the software part is an object detection algorithm named the Ag-YOLO. The NCS2 is 18 grams in weight, 1.5 watts in energy consumption, and costs about $66. The proposed model Ag-YOLO is inspired by You Only Look Once (YOLO), trained and tested with aerial images of areca plantations, and shows high accuracy (F1 score = 0.9205) and high speed [36.5 frames per second (fps)] on the target hardware. Compared to YOLOv3-Tiny, Ag-YOLO is 2× faster while using 12× fewer parameters. Based on this study, crop monitoring and crop spraying can be synchronized into one process, so that smart and precise spraying can be performed.

In-Network Flow Classification With Knowledge Distillation

Recent research has incorporated machine learning with software-defined networking to support intelligent traffic engineering. However, most frameworks only enable machine learning in remote controllers, which introduce significant signaling overhead and data forwarding costs. In this work, we present a new architecture called in-network inference (INI) to realize local learning in Neural Compute Stick (NCS), a portable device that can be connected to a programmable switch via a USB port. While NCS can flexibly extend the computing power of a switch, its limited capacity however cannot afford real-time inference for enormous traffic demands. To develop a practical local learning architecture, we design a two-phase learning framework that combines local learning with knowledge distillation and remote learning to achieve lightweight but accurate traffic classification. We further design an inference model deployment and adaptation algorithm to utilize multiple NCS devices equipped with different switches to share the inference workload of a network. Our testbed experiments show that the two-phase learning framework reduces the inference rejection rate by 46.5% and maintains the inference accuracy of 98.10%. The trace-driven simulations verify that the proposed adaptive model placement scheme considers load balancing and, hence, better utilizes the computing resources of NCS to serve dynamic inference requests.

Analysis of Edge-Optimized Deep Learning Classifiers for Radar-Based Gesture Recognition

The increasing significance of technology in daily lives led to the need for the development of convenient methods of human-computer interaction (HCI). Given that the existing HCI approaches exhibit various limitations, hand gesture recognition-based HCI may serve as a more intuitive mode of human-machine interaction in many situations. In addition, the system has to be deployable on low-power devices for applicability in broadly defined Internet of Things (IoT) and smart home solutions. Recent advances exhibit the potential of deep learning models for gesture classification, whereas they are still limited to high-performance hardware. Embedded neural network accelerators are constrained in terms of available memory, central processing unit (CPU) clock speed, graphics processing unit (GPU) performance, and a number of supported operations. The aforementioned problems are addressed in this paper by namely two approaches - simplifying the signal processing pipeline to avoid recurrent structures and efficient topological design. This paper employs an intuitive scheme allowing for the generation of the data in the compressed form from the sequence of range-Doppler images (RDI). Thus, it allows for the design of a neural classifier avoiding the usage of recurrent layers. The proposed framework has been optimized for Intel® Neural Compute Stick 2 (Intel® NCS 2), at the same time achieving promising classification accuracy of 97.57%. To confirm the robustness of the proposed algorithm, five independent persons have been involved in the algorithm testing process.

ANNETTE: Accurate Neural Network Execution Time Estimation With Stacked Models

With new accelerator hardware for DNN, the computing power for AI applications has increased rapidly. However, as DNN algorithms become more complex and optimized for specific applications, latency requirements remain challenging, and it is critical to find the optimal points in the design space. To decouple the architectural search from the target hardware, we propose a time estimation framework that allows for modeling the inference latency of DNNs on hardware accelerators based on mapping and layer-wise estimation models. The proposed methodology extracts a set of models from micro-kernel and multi-layer benchmarks and generates a stacked model for mapping and network execution time estimation. We compare estimation accuracy and fidelity of the generated mixed models, statistical models with the roofline model, and a refined roofline model for evaluation. We test the mixed models on the ZCU102 SoC board with DNNDK and Intel Neural Compute Stick 2 on a set of 12 state-of-the-art neural networks. It shows an average estimation error of 3.47% for the DNNDK and 7.44% for the NCS2, outperforming the statistical and analytical layer models for almost all selected networks. For a randomly selected subset of 34 networks of the NASBench dataset, the mixed model reaches fidelity of 0.988 in Spearman's rank correlation coefficient metric. The code of ANNETTE is publicly available at https://github.com/embedded-machine-learning/annette.

Hawk-Eye: An AI-Powered Threat Detector for Intelligent Surveillance Cameras

With recent advances in both AI and IoT capabilities, it is possible than ever to implement surveillance systems that can automatically identify people who might represent a potential security threat to the public in real-time. Imagine a surveillance camera system that can detect various on-body weapons, masked faces, suspicious objects and traffic. This system could transform surveillance cameras from passive sentries into active observers which would help in preventing a possible mass shooting in a school, stadium or mall. In this paper, we present a prototype implementation of such systems, Hawk-Eye, an AI-powered threat detector for smart surveillance cameras. Hawk-Eye can be deployed on centralized servers hosted in the cloud, as well as locally on the surveillance cameras at the network edge. Deploying AI-enabled surveillance applications at the edge enables the initial analysis of the captured images to take place on-site, which reduces the communication overheads and enables swift security actions. At the cloud side, we built a Mask R-CNN model that can detect suspicious objects in an image captured by a camera at the edge. The model can generate a high-quality segmentation mask for each object instance in the image, along with the confidence percentage and classification time. The camera side used a Raspberry Pi 3 device, Intel Neural Compute Stick 2 (NCS 2), and Logitech C920 webcam. At the camera side, we built a CNN model that can consume a stream of images directly from an on-site webcam, classify them, and displays the results to the user via a GUI-friendly interface. A motion detection module is developed to capture images automatically from the video when a new motion is detected. Finally, we evaluated our system using various performance metrics such as classification time and accuracy. Our experimental results showed an average overall prediction accuracy of 94% on our dataset.

NEURAL NETWORK COMPUTER FOR RECOVERING LOST INFORMATION FROM STANDARD SENSORS OF THE ON-BOARD SYSTEM FOR CONTROL AND DIAGNOSTICS OF TV3-117 AIRCRAFT ENGINE

The subject matter of the article is TV3-117 aircraft engine and methods for control and diagnostics its technical condition. The goal of the work is the development of a neural network computer for recovering lost information from standard sensors of the on-board control and diagnostics system of TV3-117 aircraft engine technical state in real time. The following tasks were solved in the article: recovering of lost information by an auto-associative neural network in case of a single sensor failure, recovering of lost information by an «optimal» auto-associative neural network in case of single sensor failures of the on-board control and diagnostic system, recovering the lost information by an auto-associative neural network and an on-board control and diagnostic system from the gas temperature registration sensor before the turbine compressor in case of its failure. The following methods were used: methods of probability theory and mathematical statistics, methods of neuroinformatics, methods of the theory of information systems and data processing. The following results were obtained: The urgent task of recovering lost information from standard sensors in real time has been solved. Various computer architectures and recovery algorithms were investigated. An engineering technique for recovering lost information using a neurocomputer is proposed. As a result of the use of the neurocomputer, effective and high-quality information recovery from standard sensors was ensured under the conditions of the on-board control and diagnostics system of TV3-117 aircraft engine. Conclusions: The use of an auto-associative neural network in the on-board control and diagnostics system for information recovery makes it possible to ensure fault tolerance of the measuring channels of control systems, in particular, the TV3-117 aircraft engine. The main advantage of using neural networks as part of an on-board control and diagnostics system is the possibility of training and learning in real time, taking into account the individual characteristics of a particular engine. Information recovery in case of sensor failure using an auto-associative neural network provides data recovery error of no more than 0.45 % for single failures and not more than 0.6 % for double failures. At the same time, the time of one data recovery cycle is 1589.544 ns for the Raspberry Pi NanoPi M1 Plus calculator and 196.246 ns for the specialized Intel Neural Compute Stick 2 neuroprocessor, which meets the requirements of onboard implementation as part of an onboard control and diagnostic system.

EdgeDRNN: Recurrent Neural Network Accelerator for Edge Inference

Low-latency, low-power portable recurrent neural network (RNN) accelerators offer powerful inference capabilities for real-time applications such as IoT, robotics, and human-machine interaction. We propose a lightweight Gated Recurrent Unit (GRU)-based RNN accelerator called EdgeDRNN that is optimized for low-latency edge RNN inference with batch size of 1. EdgeDRNN adopts the spiking neural network inspired delta network algorithm to exploit temporal sparsity in RNNs. Weights are stored in inexpensive DRAM which enables EdgeDRNN to compute large multi-layer RNNs on the most inexpensive FPGA. The sparse updates reduce DRAM weight memory access by a factor of up to 10x and the delta can be varied dynamically to trade-off between latency and accuracy. EdgeDRNN updates a 5 million parameter 2-layer GRU-RNN in about 0.5ms. It achieves latency comparable with a 92W Nvidia 1080 GPU. It outperforms NVIDIA Jetson Nano, Jetson TX2 and Intel Neural Compute Stick 2 in latency by 5X. For a batch size of 1, EdgeDRNN achieves a mean effective throughput of 20.2GOp/s and a wall plug power efficiency that is over 4X higher than the commercial edge AI platforms.

Traffic light and moving object detection for a guide‐dog robot

Guide dogs are helpful for visually impaired people for navigating through the streets. However, it is expensive and time consuming to train a guide dog. In addition, a guide dog cannot decide when and where to cross a street safely, and it is up to the human to decide. Here, the authors propose a framework for creating a guide dog robot by using artificial intelligence and other technologies. The proposed framework is based on an Intel UP squared board, together with a Neural Compute Stick Movidius to process the images gathered from a GoPro camera. MobileNet single shot detector (SSD) is the main framework to detect the moving objects in the environment. The final decision is made after fusing the information gathered from all the sources. The authors also apply the Amazon Alexa device for the voice communication between the guide dog robot and the visually impaired person. A prototype of the proposed system is implemented and tested. Experimental results show that the proposed framework can process the information at a traffic intersection scene and can guide a blind person to cross the street safely.

License plate segmentation and recognition system using deep learning and OpenVINO

As Intelligent Transportation Systems applications increase in prevalence, Automatic License Plate Recognition solutions must be made continually faster and more accurate. The authors propose an embedded system for fast and accurate license plate segmentation and recognition using a modified single shot detector (SSD) with a feature extractor based on depthwise separable convolutions and linear bottlenecks. The feature extractor requires less parameters than the original SSD + VGG implementation, enabling fast inference. Tested on the Caltech Cars dataset, the proposed model achieves 96.46% segmentation and 96.23% recognition accuracy. Tested on the UCSD-Stills dataset, the proposed model achieves 99.79% segmentation and 99.79% recognition accuracy. The authors achieve a per-plate (resized to 300 × 300 px) processing time of 59 ms on an Intel Xeon CPU with 12 cores (2.60 GHz per core), 14 ms using the same CPU and OpenVINO (a neural network acceleration platform), and 66 ms using the proposed low-cost Raspberry Pi 3 and Intel Neural Compute Stick 2 with OpenVINO embedded system.