The nameedge intelligence, also known asEdge AI, is a recent term used in the past few years to refer to the confluence of machine learning, or broadly speaking artificial intelligence, with edge computing. In this article, we revise the concepts regarding edge intelligence, such as cloud, edge, and fog computing, the motivation to use edge intelligence, and compare current approaches and analyze application scenarios. To provide a complete review of this technology, previous frameworks and platforms for edge computing have been discussed in this work to provide the general view of the basis for Edge AI. Similarly, the emerging techniques to deploy deep learning models at the network edge, as well as specialized platforms and frameworks to do so, are review in this article. These devices, techniques, and frameworks are analyzed based on relevant criteria at the network edge, such as latency, energy consumption, and accuracy of the models, to determine the current state of the art as well as current limitations of the proposed technologies. Because of this, it is possible to understand the current possibilities to efficiently deploy state-of-the-art deep learning models at the network edge based on technologies such as artificial intelligence accelerators, tensor processing units, and techniques that include federated learning and gossip training. Finally, the challenges of Edge AI are discussed in the work, as well as the future directions that can be extracted from the evolution of the edge computing and Internet of Things approaches.

There have been significant advances regarding target detection in the autonomous vehicle context. To develop more robust systems that can overcome weather hazards as well as sensor problems, the sensor fusion approach is taking the lead in this context. Laser Imaging Detection and Ranging (LiDAR) and camera sensors are two of the most used sensors for this task since they can accurately provide important features such as target´s depth and shape. However, most of the current state-of-the-art target detection algorithms for autonomous cars do not take into consideration the hardware limitations of the vehicle such as the reduced computing power in comparison with Cloud servers as well as the reduced latency. In this work, we propose Edge Computing Tensor Processing Unit (TPU) devices as hardware support due to their computing capabilities for machine learning algorithms as well as their reduced power consumption. We developed an accurate and small target detection model for these devices. Our proposed Multi-Level Sensor Fusion model has been optimized for the network edge, specifically for the Google Coral TPU. As a result, high accuracy results are obtained while reducing the memory consumption as well as the latency of the system using the challenging KITTI dataset.

Data acquisition and treatment are key issues for any Deep Learning (DL) technique, especially in computer vision tasks. A big effort must be done for the creation of labeled datasets, due to the time this task requires and its complexity in cases where different sensors must be used. This is the case of radar imaging applications, where radar data are difficult to analyze and must be labeled manually. In this paper, a semi-automatic framework to generate labels for range Doppler maps (radar images) is proposed. This technique is based on a sensor fusion approach with radar and camera sensors. The proposed scheme operates in two steps: The first step is the environment features extraction, in which the radar data is preprocessed and filtered to remove ghost targets and detect clusters, and camera data are used to extract the information of the targets. In the second step, a rule-based system that considers the extracted features fuses the information to generate labels for the radar data. By using the proposed framework, the experimentation performed suggests that the time required to label the data is reduced as well as the possibility of human error during the labeling task. Our results show that the proposed technique can improve the final model accuracy with regards the traditional labeling method, carried out by human experts.

The interfaces between users and systems are evolving into a more natural communication, including user gestures as part of the interaction, where air-writing is an emerging application for this purpose. The aim of this work is to propose a new air-writing system based on only one array of ultrasonic transceivers. This track will be obtained based on the pairwise distance of the hand marker with each transceiver. After acquiring the track, different deep learning algorithms, such as long short-term memory (LSTM), convolutional neural networks (CNN), convolutional autoencoder (ConvAutoencoder), and convolutional LSTM have been evaluated for character recognition. It has been shown how these algorithms provide high accuracy, where the best result is extracted from the ConvLSTM, with 99.51% accuracy and 71.01 milliseconds of latency. Real data were used in this work to evaluate the proposed system in a real scenario to demonstrate its high performance regarding data acquisition and classification.

There is a growing interest in pushing computation to the edge, especially the problem-solving abilities of artificial neural networks (ANNs). This article presents a simplified method to obtain a ternary neural network based on the multilayer perceptron. The method is focused on resource-constrained devices, where memory, computing power, and battery are some of the most relevant constraints. A dynamic threshold is estimated to perform ternarization, and a new pruning technique is proposed to obtain a drastic reduction in the ANN’s size, with the corresponding decrease in resource utilization and power consumption of the resulting hardware. In addition, a support framework has been developed to automate hardware design exploration and generation from the network trained in software. Experimental results show that the proposed method and architecture, when implemented in a field-programmable gate array (FPGA), provide excellent figures in power (0.11–0.13 W) and efficiency (1225–1448 kfps/W) with respect to state of the art, being its efficiency double than the maximum one reported previously.