移动计算与边缘设备中的人工智能部署

Deployment of AI in MobileComputing and Edge Devices Abstract— Artificial intelligence (AI) has emerged as a critical force in the technology industry,improving efficiency, productivity, and decision-making across a wide range of sectors. Theapplication of AI is rapidly spreading beyond desktop and cloud computing, offering not onlyperformance benefits but also substantial cost-saving potential. Often overlooked domains are the integration of AI in mobile computing and edge devices.Examples include Advanced Driver Assistance Systems in autonomous vehicles, mappingcapabilities in robotic vacuum cleaners, biometric data analysis in wearable fitness trackers,and natural language processing in smart home devices like Alexa or Google Home. Theadvantages of using AI on edge devices, include the portability and scalability of their usage inmany industries, but also reduced latency, and privacy. However, it is challenging to implementbecause of the limited computational power and memory, few developers experienced in AIoptimization for low-power devices, and the selection of the right hardware and models to use. In this paper, we offer solutions for what to do when faced with the challenges of deploymentof AI in embedded devices, and lead by example on how we overcame these challenges forour specific edge device deployment. Our edge deployment uses a YOLOv8 nano model on aJetson Nano (reComputer J1020v2, 4GB) for object detection in a factory setting. The systemachieves 75% mAP (mean Average Precision) while processing up to 7 full HD images persecond, demonstrating that low-power edge devices can deliver reliable, real-time AIperformance in practical applications. 1.Introduction Artificial Intelligence (AI) is increasingly being leveragedacross a wide range of industries to improve automation,decision-making, and overall technological efficiency [1].Recent advances have shown that edge devices runningAI can outperform their traditional algorithm-basedcounterparts in tasks requiring local, real-timeintelligence [2]. One common method in Internet of Things (IoT) is to rely on the remote device to gatherdata, and then process it inside a cloud computingserver. However, in situations where real-time dataprocessing is required, cloud computing is not suitable,due to a lot of latency, limited connectivity, bandwidthshortage, and energy consumption. Additionally, transmitting large amounts of data to andfrom cloud servers raises operational costs and privacyconcerns. For these situations, edge computing is abetter option, since it can obtain and process datasimultaneously. For example, for devices running objectdetection algorithms, taking an image and immediatelyprocessing it is a significant advantage over relying ondata transmission to a cloud server. For instance, ObjectDetection algorithms are often developed and executedusing Graphics Processing Units (GPUs), which arecommonly found in desktop PCs and cloud servers.Using a GPU in a gadget with limited power sources(battery) can significantly reduce autonomy. Thesolution is usage of devices with built in low poweredhardware accelerators like the Jetson Nano. •Low latency:Certain systems like real-time qualitycontrol in manufacturing, need timing precisions tothe millisecond or even lower, which only on-device computation can provide. •Data secrecy:Phone apps handling personal datalike biometric systems or medical diagnosticbenefit from keeping data local, providing privacyand compliance with regulations. •Offline operation:In remote environments, such asagricultural fields, edge devices can functionindependently without relying on cloudinfrastructure. This also means that if the application only requiressome of the bullet points listed above, it also haspotential to benefit from the rest of the benefits thatedge devices bring. [3] A. Paper Overview The paper is structured as follows: The introduction toedge AI deployments and how to select a model thatfits a device are described in sections 1 to 3. Thesolutions and counter measures for common problemsin edge AI deployments are explained in sections 4 and5. Our approach as well as the results achieved, arepresented in Section 6. Finally, conclusions are drawnin Section 7. The advantages are especially noticeable in places withlimited or inconsistent connectivity, such as remoteindustrial locations, rural areas like crops in agriculture,or mobile platforms like self-driving cars. With theirmobility and autonomy, edge-based AI capabilities areideal for Internet of Things (IoT) deployments, wheredispersed intelligence enables scalability and reactivityof systems. Edge devices reduce reliance on cloudinfrastructure, reduce bandwidth consumption, andimprove user data privacy by enabling on-device dataprocessing, which keeps sensitive data on the device. 2. Circumstances for using an edgedevice Some key elements that one should look at to decide ifan edge device is needed are listed below. Does yourapplicat