PaperlayerView exact revision on arXiv

Source-linked AI summary

An Edge-Computing Based Architecture for Mobile Augmented Reality

Jinke Ren, Yinghui He, Guan Huang, Guanding Yu, Yunlong Cai, Zhaoyang Zhang

arXiv:1810.02509v2cs.ITmath.IT

TL;DR

Mobile AR's long processing delays and high energy consumption on hardware-limited devices motivate the paper's approach. It proposes a hierarchical user-edge-cloud framework and operation mechanism; simulations report lower latency and energy consumption than conventional schemes.

  • Problem

    Hardware-limited mobile devices incur long processing delays and high energy consumption for AR, while existing local techniques cannot keep pace with demand.

    Source: §I · p. 1, C. Existing Technologies · p. 3

  • Method

    A hierarchical user-edge-cloud architecture inserts an edge layer that integrates communication, computation, and control to support collaborative offloading of AR tasks.

    Source: §I · p. 1, A. Hierarchical Computation Architecture · pp. 3–4

  • Results

    At 36 devices, edge computing reduces latency by 41.44% versus local and 12.85% versus cloud, and saves energy 73.71% versus local and 65.34% versus cloud.

    Source: §V · p. 7

  • Takeaways & Limitations

    The tests report reduced processing latency and energy consumption and describe associated QoE benefits including high image resolution, long battery autonomy, portability, and real-time performance.

    Source: §V · p. 7

Abstract

from arXiv · show

In order to mitigate the long processing delay and high energy consumption of mobile augmented reality (AR) applications, mobile edge computing (MEC) has been recently proposed and is envisioned as a promising means to deliver better quality of experience (QoE) for AR consumers. In this article, we first present a comprehensive AR overview, including the indispensable components of general AR applications, fashionable AR devices, and several existing techniques for overcoming the thorny latency and energy consumption problems. Then, we propose a novel hierarchical computation architecture by inserting an edge layer between the conventional user layer and cloud layer. Based on the proposed architecture, we further develop an innovated operation mechanism to improve the performance of mobile AR applications. Three key technologies are also discussed to further assist the proposed AR architecture. Simulation results are finally provided to verify that our proposals can significantly improve the latency and energy performance as compared against existing baseline schemes.

I. INTRODUCTION

Mobile AR is constrained by long processing delay and high energy use on resource-limited devices, while existing fixes do not keep pace with demand. The paper reviews AR and proposes a hierarchical user-edge-cloud framework with an operation mechanism and supporting technologies.

I. INTRODUCTION

  • Mobile AR faces long processing delays and high energy consumption on hardware-limited devices, while existing techniques cannot keep pace with growing demand and can increase device costs.These constraints are identified as major hurdles to implementation and wider commercial use.Source: §I · p. 1
  • The framework inserts an edge layer between user and cloud layers, enabling AR tasks to be offloaded to both edge and cloud servers for collaborative processing.The edge layer integrates communication, computation, and control functions.Source: §I · p. 1
  • The proposed MEC-based AR framework combines a hierarchical computation architecture, an AR operation mechanism, and three supporting technologies, with simulations reporting improved energy and delay performance over baseline schemes.The three technologies address joint communication and computation resource allocation, collaborative cloud-edge computing, and content-based image retrieval.Source: §I · pp. 1–2

II. OVERVIEW OF AUGMENTED REALITY

  • The overview covers indispensable components of general AR applications, mobile AR devices, and existing approaches to latency and energy challenges.It frames the discussion around energy efficiency and latency optimization.Source: §II · p. 2

A. AR Application Components

General AR applications rely on five components that work together from video capture through display. The overview also contrasts devices' operating modes and practical trade-offs, including power use and portability.

A. AR Application Components

  • Five components are indispensable to almost all AR applications: video source, tracker, mapper, object recognizer, and renderer.They work in sequence to capture and track video, establish virtual coordinates, identify objects, retrieve augmentation, and display the result.Source: A. AR Application Components · p. 2

B. Fashionable AR Devices

  • Google Glass, HoloLens, and ODG R-9 operate independently, while Meta 2 and Recon Jet rely on external equipment such as PCs or smartphones.The devices support different use cases, but the independent devices described also face high energy consumption, and Meta 2 is non-portable for outdoor use.Source: B. Fashionable AR Devices · p. 2
  • The described devices provide immersive AR, yet generally face long processing delays and high energy consumption, the overview's key performance concerns.Google Glass, HoloLens, and ODG R-9 are described as unable to work for long periods; Meta 2's reliance on powerful computers limits outdoor portability.Source: B. Fashionable AR Devices · p. 2

C. Existing Technologies

The overview describes hardware, approximation, and partial-frame updating as ways to reduce AR energy use or delay, but notes that local computation of intensive tasks remains a constraint.

C. Existing Technologies

  • Advanced hardware can reduce energy use through low-frequency, low-voltage multicore CPUs and dynamic voltage and frequency scaling.The passage contrasts these with single-core structures and links scaling to specific requirements.Source: C. Existing Technologies · p. 2
  • Approximate computing can trade computational accuracy against energy or delay efficiency when task accuracy requirements permit it.Location sensing is given as an example of a low-precision task suitable for approximate processing.Source: C. Existing Technologies · pp. 2–3
  • Partial video-frame updating avoids redundant computation by processing newly generated data rather than repeatedly processing unchanged video.The passage motivates this approach by the large volume of video data produced by frequent device movement.Source: C. Existing Technologies · p. 3
  • These techniques improve end-to-end latency and energy only to some extent because tracking, mapping, and object recognition remain local, motivating computation offload to MEC servers.The passage presents MEC as a way to address mobile-device resource shortages by offloading computation-intensive tasks to network-edge servers.Source: C. Existing Technologies · p. 3

III. MEC-BASED AR FRAMEWORK

The framework adds an edge layer to a multi-user user-edge-cloud architecture and coordinates communication, computation, and control there. Its edge modules process offloaded AR tasks and coordinate resources across the layers.

III. MEC-BASED AR FRAMEWORK

  • The framework targets simultaneous execution of multiple AR applications using user, edge, and cloud layers, with devices wirelessly connected to the edge and cloud resources supplying storage and computation.The cloud layer is described as having an enormous database and abundant computational resources.Source: A. Hierarchical Computation Architecture · p. 3, §III · p. 3

A. Hierarchical Computation Architecture

  • The architecture's distinguishing feature is an edge layer between user and cloud layers, potentially located at cellular base stations or WiFi access points.The proposed design allows edge and cloud computing to collaborate for performance enhancement.Source: A. Hierarchical Computation Architecture · p. 3
  • Three SDN-based edge modules—communication unit, operation platform, and virtualized controller—integrate communication, computation, and control functions.The communication unit carries data between layers; the operation platform processes offloaded tasks; and the controller coordinates edge-layer functions.Source: A. Hierarchical Computation Architecture · pp. 3–4
  • The communication unit receives offloaded data and returns computation results to users, while the operation platform caches inputs and uses computing and graphics modules for AR tasks.The platform's small edge database stores frequently accessed object information, avoiding repeated retrieval from the remote cloud database and reducing end-to-end latency.Source: A. Hierarchical Computation Architecture · p. 4
  • The virtualized controller coordinates network, computing, storage, and system controllers, including resource allocation to AR tasks with specific requirements.It also evaluates task priorities and collaboration properties and manages edge-database memory and updates.Source: A. Hierarchical Computation Architecture · p. 4

B. MEC-based AR Operation Mechanism

The operation mechanism divides AR processing across user, edge, and cloud tiers. Video capture and rendering remain local, while computation-intensive operations are handled at the edge or cloud.

B. MEC-based AR Operation Mechanism

  • The mechanism assigns local video capture and rendering to the user tier and offloads computation-intensive tracking, mapping, and object recognition to edge and cloud servers.The user tier senses the environment, captures gestures, and sends video streams and operation instructions uplink.Source: B. MEC-based AR Operation Mechanism · pp. 4–5
  • At the edge, tracking and mapping reconstruct virtual coordinates, while frame classification and image retrieval reuse work across users observing the same object.An image found in the edge database yields its annotation there; otherwise, its frame is sent to the cloud for searching.Source: B. MEC-based AR Operation Mechanism · p. 5
  • The cloud searches for images unavailable at the edge, returns requested information for combination with the virtual map, and sends results to devices for local rendering.The local renderer mixes the returned data with the original video for display.Source: B. MEC-based AR Operation Mechanism · p. 5

IV. KEY IMPLEMENT TECHNOLOGIES

The section identifies three technologies intended to improve energy efficiency and reduce processing latency in the proposed AR architecture.

IV. KEY IMPLEMENT TECHNOLOGIES

  • Joint communication and computation resource allocation is one of the three technologies proposed to improve energy efficiency and reduce processing latency.The section also identifies collaborative cloud and edge computing and content-based image retrieval.Source: §IV · p. 5
  • Collaborative cloud and edge computing is another proposed technology for improving energy efficiency and reducing processing latency.It is listed alongside resource allocation and content-based image retrieval.Source: §IV · p. 5
  • Content-based image retrieval is the third identified technology for improving energy efficiency and reducing processing latency.The section presents it as support for the proposed AR architecture.Source: §IV · p. 5

A. Joint Communication and Computation Resource Allocation

Because AR tasks share wireless and edge/cloud computing resources, the section considers centralized and distributed allocation schemes. Their applicability depends on what information the edge can obtain.

A. Joint Communication and Computation Resource Allocation

  • Joint allocation of wireless communication and edge/cloud computation resources is essential because AR devices share those resources.The section distinguishes centralized and distributed allocation schemes.Source: A. Joint Communication and Computation… · p. 5
  • Centralized allocation lets edge-layer controllers use user priority, channel state, and task requirements to adapt resource policies to user demands.The controllers can optimize objectives such as end-to-end latency under resource constraints or mobile energy efficiency under an offloading-latency constraint.Source: A. Joint Communication and Computation… · p. 5
  • When the edge cannot obtain that information, distributed systems can use game-theoretical algorithms based on past network and channel information.The passage says the resulting allocation achieves Nash equilibrium and can deliver satisfactory latency and energy consumption experiences.Source: A. Joint Communication and Computation… · p. 5

B. Collaborative Cloud and Edge Computing

Cloud-edge collaboration combines cloud capacity with edge proximity, addressing the limits of each alone. The section describes storage and computation allocation across the two layers.

B. Collaborative Cloud and Edge Computing

  • Cloud-edge collaboration pairs the cloud’s computation capacity with edge servers’ proximity to users, despite remote-cloud propagation delay and limited edge capacity.The passage presents collaboration as a promising paradigm for better AR performance.Source: B. Collaborative Cloud and Edge… · p. 5
  • Popular-object information is cached at the edge, while additional information for unpopular objects remains in the cloud to ease edge-storage burden.Historical access records and a threshold are used to judge popularity; the passage says this can avoid massive edge recognition operations and reduce AR processing latency.Source: B. Collaborative Cloud and Edge… · pp. 5–6
  • When multiple edge servers share a cloud server, cloud computation should be allocated according to workload and local capacity.The passage proposes allocating more cloud resources to edge servers with less capacity and heavier workload to balance uneven resources and workloads.Source: B. Collaborative Cloud and Edge… · p. 6

C. Content-based Image Retrieval

The proposed image-retrieval pipeline extracts image features, narrows matching to a shortlist, and verifies the best match. It can use edge and cloud resources together, with cloud search as a fallback when verification fails.

C. Content-based Image Retrieval

  • The pipeline extracts salient interest points from an input image and uses them to estimate similarity with standard images in edge and cloud databases.SIFT and SURF are examples of feature descriptors.Source: C. Content-based Image Retrieval · p. 6
  • Feature matching first filters candidate images using indexed stored features, then applies slower pairwise matching only to the shortlist.This is intended to substantially reduce image-retrieval delay.Source: C. Content-based Image Retrieval · p. 6
  • Geometric verification checks whether the best-matched image differs from the input only through geometric and photometric distortions before establishing a match.If verification is incorrect, the input image and its features are sent to the cloud for further searching.Source: C. Content-based Image Retrieval · p. 6
  • Joint use of edge and cloud computation and storage improves recognition accuracy while reducing end-to-end latency and energy consumption.The cloud server searches further when edge-side verification fails.Source: C. Content-based Image Retrieval · p. 6

V. PERFORMANCE EVALUATION

In the evaluated setting, the edge computing scheme outperforms local and cloud computing on latency and energy consumption. Its advantage is reported across device counts and delay tolerances.

V. PERFORMANCE EVALUATION

  • The simulation compares local computing, cloud computing, and edge computing for multiple devices served by one base station, with remote-cloud distance ignored.The edge and cloud capacities are 5 × 10^10 and 3 × 10^11 CPU cycles/s, respectively; the local scheme keeps tracking and mapping local and offloads recognition to the cloud.Source: §V · p. 6
  • Edge computing achieves the best latency among the three schemes, while latency rises with device count; local computing has the worst latency.The edge scheme's latency advantage over cloud narrows as cloud computing capacity increases.Source: §V · pp. 6–7
  • At 36 devices, edge computing reduces latency 41.44% vs local and 12.85% vs cloud, and energy consumption 73.71% vs local and 65.34% vs cloud.The energy comparison uses a 450 ms maximum delay tolerance.Source: Figure 4 · p. 7, §V · p. 7
  • With 30 devices, mobile energy consumption decreases as delay tolerance increases, and edge computing remains superior in the comparison.The paper attributes higher energy under stricter delay requirements to the additional computation resources needed to meet them.Source: §V · p. 7

VI. CONCLUSION

The conclusion presents a three-layer MEC-based framework that coordinates edge and cloud computation for AR applications. Simulations show substantial performance improvement over existing schemes.

VI. CONCLUSION

  • The architecture places an edge layer between user and cloud layers and integrates communication, computation, and control at the edge.AR tasks can be offloaded to edge and cloud servers for collaborative computation.Source: §VI · p. 7
  • A mechanism is developed to support multiple AR applications for different mobile subscribers, alongside three technologies intended to further reduce latency and energy consumption.Source: §VI · p. 7
  • Simulation results demonstrate substantial performance improvement over existing schemes for the proposed architecture and mechanism.Source: §VI · p. 7
Loading 1810.02509v2…