PaperlayerView exact revision on arXiv

Source-linked AI summary

Mobile Edge Intelligence and Computing for the Internet of Vehicles

Jun Zhang, Khaled B. Letaief

arXiv:1906.00400v1cs.NIeess.SP

TL;DR

Intelligent IoV must handle massive vehicle-generated data that onboard systems and Cloud computing cannot efficiently store or process because of resource, power, communication, and latency constraints. The paper surveys edge information systems—covering edge caching, edge computing, and edge AI—and their design issues, platforms, methodologies, and intelligent-vehicle applications. It concludes that EIS can support data acquisition, storage, and processing when integrated with onboard processors and Cloud platforms.

  • Problem

    Intelligent IoV generates massive data, while onboard systems face resource and power constraints and Cloud processing faces communication overhead and latency.

  • Method

    The paper surveys edge caching, edge computing, and edge AI, examining platforms, design methodologies, hardware, applications, and open research problems.

  • Results

    The survey presents EIS applications for intelligent-vehicle perception, mapping, localization, mobility analytics, public safety, and privacy-sensitive learning.

  • Takeaways & Limitations

    EIS provides a framework for low-latency content delivery, computation, and localized data acquisition, aggregation, and processing in intelligent IoV.

Abstract

from arXiv · show

The Internet of Vehicles (IoV) is an emerging paradigm, driven by recent advancements in vehicular communications and networking. Advances in research can now provide reliable communication links between vehicles, via vehicle-to-vehicle communications, and between vehicles and roadside infrastructures, via vehicle-to-infrastructure communications. Meanwhile, the capability and intelligence of vehicles are being rapidly enhanced, and this will have the potential of supporting a plethora of new exciting applications, which will integrate fully autonomous vehicles, the Internet of Things (IoT), and the environment. These trends will bring about an era of intelligent IoV, which will heavily depend upon communications, computing, and data analytics technologies. To store and process the massive amount of data generated by intelligent IoV, onboard processing and Cloud computing will not be sufficient, due to resource/power constraints and communication overhead/latency, respectively. By deploying storage and computing resources at the wireless network edge, e.g., radio access points, the edge information system (EIS), including edge caching, edge computing, and edge AI, will play a key role in the future intelligent IoV. Such system will provide not only low-latency content delivery and computation services, but also localized data acquisition, aggregation and processing. This article surveys the latest development in EIS for intelligent IoV. Key design issues, methodologies and hardware platforms are introduced. In particular, typical use cases for intelligent vehicles are illustrated, including edge-assisted perception, mapping, and localization. In addition, various open research problems are identified.

I. INTRODUCTION

Intelligent IoV combines increasingly capable vehicles and connected networks but faces massive data, onboard resource, power, and Cloud latency constraints. The survey presents edge information systems as an integrated framework for caching, computing, analytics, and localized processing.

  • Intelligent IoV builds on V2V and Internet-connected vehicles with communications, processing, storage, and learning capabilities.
  • B. Living On The Edge: Edge information systems integrate caching, computing, and AI at radio access points to support low-latency delivery, computation, and localized data processing.
  • A. Big Data in Intelligent IoV: More than 200 future-vehicle sensors may generate 3 Gbit/s to 40 Gbit/s, while each autonomous vehicle may produce approximately 4,000 GB daily.
  • A. Big Data in Intelligent IoV: Onboard storage and computation remain limited for intelligent IoV data, with sensing-data SSDs potentially filling within hours and processing exhausting vehicle resources.
  • A. Big Data in Intelligent IoV: Real-time autonomous driving requires traffic processing within 100 ms, but GPU power and cooling demands can reduce driving range and fuel efficiency.
  • Cloud processing helps with software updates and model training but is constrained by long latency and massive data transmission.
  • B. Living On The Edge: The survey assigns onboard processing to highly latency-sensitive tasks, edge servers to local computation and storage, and Cloud computing to large-scale training and historical data.

C. Paper Outline

The paper surveys edge information systems for intelligent IoV and organizes the discussion around system technologies, vehicle tasks, and application examples. It introduces edge caching, mobile edge computing, edge AI, and key intelligent-vehicle tasks.

  • The survey covers edge caching, edge computing, and edge AI, including platforms, design methodologies, hardware, use cases, and open research problems.
  • Section II introduces EIS and intelligent-vehicle tasks, while Sections III–V cover edge caching, MEC, and edge AI respectively.
  • The paper frames perception, HD mapping, and SLAM as key application cases for intelligent vehicles.

A. The Edge Information System

EIS serves as an intermediary between onboard processors and the remote Cloud, using edge resources and vehicular communications to acquire, aggregate, and process IoV data. Its architecture includes edge servers, vehicles, and user devices, supporting vehicle-as-client and vehicle-as-server scenarios for intelligent-vehicle tasks.

  • EIS acts as an intermediary between onboard processors and the remote Cloud while acquiring, aggregating, and processing data for intelligent vehicles.
  • Edge servers at base stations or roadside units provide storage, computing, backbone connectivity, and V2I communication within their coverage ranges.Base-station servers offer broader coverage, while roadside-unit servers are closer to vehicles and can support lower latency.
  • Intelligent vehicles combine sensors, communication modules, and onboard computing and storage, communicating through V2V or V2I links.
  • Vehicle as a Client: In the vehicle-as-client scenario, edge servers co-locate data acquisition and processing for applications such as local HD-map construction and road-condition monitoring.
  • Vehicle as a Server: In the vehicle-as-server scenario, vehicles provide services to passengers, recipients, and neighboring vehicles, enabling personalized experiences and cooperative perception or driving.Compared with edge-server approaches, this scenario suffers less from mobility and supports cooperation among neighboring vehicles.
  • Key Tasks of Intelligent Vehicles: The paper focuses on perception, HD mapping, and SLAM, whose outputs support localization, mapping, decision-making, planning, and vehicle control.

1) Perception:

Intelligent-vehicle perception uses heterogeneous onboard sensors for tasks such as object detection, classification, localization, and obstacle detection, but sensor limitations constrain current automation. EIS can augment perception through edge computation, storage, and cooperative sensing.

  • Cameras support object detection, classification, and vision-based localization; LiDAR supports 3D mapping and localization; Radar and Sonar support obstacle detection.Sensor types differ in computational cost and capability, with cameras and stereovision more computationally expensive than active sensors such as LiDAR.
  • Sensor limitations and high costs are cited as a main reason commercial vehicles currently achieve only Level 1 to Level 2 automation.Perception also faces poor weather or lighting and complex urban environments.
  • Edge-Assisted Perception: EIS can enhance perception by improving camera and stereovision sensing accuracy through deep learning and enabling computation-intensive multi-sensor fusion at edge servers.
  • HD Mapping: HD maps provide centimetrelevel localization by comparing onboard-sensor observations with corresponding map data, overcoming limitations of GNSS-based methods.An HD map models roadway features with 10∼20 cm surface accuracy.
  • Perception Challenges: Internal sensors lack guaranteed integrity because environmental noise and cost concerns can limit their capabilities.
  • Perception Challenges: Sensors have limited perception ranges and cannot measure objects outside their field of view or hidden by obstacles.
  • HD Mapping: EIS supports HD-map dissemination, data aggregation, map building, change detection, and geographically coordinated crowdsourced mapping.Keeping and processing mapping data locally is presented as a way to achieve more efficient HD mapping.

3) Simultaneous Localization and Mapping (SLAM):

Map-based localization can lose accuracy after drastic environmental changes, while SLAM jointly estimates vehicle state and constructs an environment map. Autonomous-driving SLAM requires faster, more efficient algorithms for outdoor, variably lit road environments and intensive computation.

  • Drastic environmental changes can reduce map-based localization accuracy enough to affect driving safety.
  • SLAM simultaneously estimates a vehicle’s state and constructs an environmental map without relying heavily on a priori information.
  • SLAM allows vehicles to continuously observe their environment and adapt to new situations, supporting accurate autonomous operation.The paper describes accurate SLAM as necessary for full autonomy and notes its use in DARPA Urban Challenge vehicles.
  • Using both map-based and SLAM-based localization provides redundancy for safer and more consistent road behavior.
  • Many existing SLAM algorithms target indoor, highly structured environments, whereas autonomous vehicles operate outdoors with variable lighting and road-based conditions.
  • Autonomous-driving SLAM therefore requires faster and more efficient algorithms, with computation demand identified as highly intensive.

2) Spatial Locality-Aware Caching:

Spatial locality-aware caching accounts for location-dependent content importance in vehicular networks. Research addresses mobility, coverage, intermittent connectivity, freshness, and joint caching-computing resource allocation across edge and vehicle caches.

  • Spatial locality-aware caching considers different importance or popularity of the same content across geographic regions.
  • Edge caching can reduce delivery delay by placing content at servers in specific areas according to location-dependent demand.One deep-learning caching scheme detects passenger ages and genders when making regional caching decisions.
  • Large-size content such as videos, music, and HD maps is difficult to deliver because vehicles face limited network capacity and intermittent connections.
  • Vehicle mobility prediction using LSTM can support proactive caching through deep reinforcement learning in multi-access EIS.
  • Vehicle caching complements edge-server caching when coverage and vehicle connections are unreliable, exploiting mobility for more cost-effective and utility-enhanced services.
  • Vehicle-cache designs must account for content size, update frequency, freshness, and limited onboard storage to improve caching efficiency and downloading experiences.
  • Caching vehicles can serve moving or static users within communication range, making relationships between caching vehicles and users central to policy design.

4) Parked Vehicles for Caching:

Parked vehicles can serve as cache servers connected to edge infrastructure and moving vehicles. Edge caching supports perception, localization, sensing, content delivery, security, and traffic-control applications in intelligent IoV.

  • Parked Vehicles for Caching: Parked vehicles can act as cache servers because they conveniently connect with edge servers and moving vehicles.Studies analyze parked-vehicle connectivity and develop competition and cooperation mechanisms to encourage participation.
  • Cache-Assisted Perception and Localization: Edge caching assists perception and localization tasks including autonomous overtaking, cooperative collision avoidance, see-through, and bird’s-eye-view services.These functions provide perception content to vehicles to assist driving and improve traffic safety.
  • Other Applications: Edge servers can collect and cache temporary traffic-event videos, such as footage of a minor accident, for delivery to approaching vehicles.The service relies on cooperation among edge servers and vehicles for real-time road surveillance.
  • Other Applications: InfoRank selects and ranks intelligent vehicles for urban sensing, using them as data caches to reduce edge-server burden.The approach completes vicinity monitoring with a small cost.
  • Other Applications: Predicted vehicle–RSU connections and content popularity enable proactive prefetching of requested content at edge servers.The OTT system uses real-world testbed predictions to support content delivery.
  • Other Applications: Edge caching has also been studied for secure information sharing and transportation control, addressing sensor limitations and the relationship between user equilibrium and system optima.Secure sharing targets false-data dissemination and vehicle tracking, while transportation control models communication costs for cache-enabled vehicles.

IV. EDGE COMPUTING FOR INTELLIGENT IOV

Edge computing places processing resources near vehicles to provide low-latency services while reducing mobile energy consumption and improving privacy and security. Intelligent IoV edge-computing research covers platforms, deployment, resource allocation, offloading, and mobility management.

  • Edge Computing for Intelligent IoV: MEC servers near mobile users provide ultra-low latency, reduced mobile energy consumption, and enhanced privacy and security.The section surveys hardware platforms, design problems, and MEC-enabled IoV applications.
  • Edge Computing Platforms: Edge servers equipped with GPUs or Edge TPUs can be deployed at base stations and roadside units using virtualization technologies.Relevant technologies include network virtualization, network-function virtualization, and virtual machines.
  • Edge Computing Platforms: OpenVDAP provides a full-stack edge platform combining onboard computing and communication, secure vehicle operations, an edge-aware application library, and workload offloading and scheduling.Its components address isolation, security, privacy, application support, and resource management.
  • Edge Computing Platforms: CAVBench benchmarks edge-computing platforms for connected and autonomous vehicles across six applications, including SLAM, object detection, tracking, battery diagnostics, speech recognition, and edge video analysis.The benchmark is designed to evaluate different platform capabilities across representative workloads.
  • Deployment and Resource Management: MEC research compares coupled and decoupled server placement and develops stochastic, game-theoretic, contextual, load-balancing, and mobility-aware resource-management methods.Mobility-aware approaches include handover and migration costs, mobility prediction, cooperation among edge servers, and BS–RSU integration.
  • Resource Allocation and Offloading: Resource allocation must account for stochastic wireless channels and task arrivals, shared radio and computing resources, and vehicle mobility affecting offloading and result feedback.These factors directly influence computation latency and offloading decisions.

B. Vehicle as a Server

In the Vehicle-as-a-Server scenario, intelligent vehicles contribute onboard computing and other resources to assist passengers, other vehicles, or MEC infrastructure. This model depends on incentives, task assignment, and mobility-aware resource utilization.

  • Vehicle as a Server: Vehicles can provide computation services to passengers or cooperate with other vehicles, creating a need for resource-sharing incentives.Vehicular Cloud systems similarly leverage onboard connectivity, computation, storage, and sensing, while high mobility remains a major implementation challenge.
  • Vehicle as a Server: NVIDIA DRIVE combines in-vehicle computing, a reference architecture, data-center simulation, and deep-neural-network training for autonomous-driving development.The DRIVE AGX platform is built on NVIDIA Xavier.
  • Resource Sharing and Incentives: Market mechanisms can incentivize nearby vehicles to provide vacant computing power when roadside-unit servers are sparse and radio coverage is constrained.Distributed task-allocation algorithms target cost minimization.
  • Resource Sharing and Incentives: A two-stage approach combines contract theory and matching theory to incentivize vehicles and assign computation tasks supporting base-station MEC servers.The BS first designs resource-sharing contracts, after which participating vehicles are assigned tasks.

1) Transit/Human Mobility Analytics:

Edge computing supports mobility analytics, driver-specific services, traffic estimation, and public-safety video analytics by processing vehicular data near users. Edge AI frameworks extend this approach to collaborative training and joint inference.

  • Transit/Human Mobility Analytics: Trellis, an edge-computing platform deployed on public transportation vehicles, performs human-mobility analytics with lower latency, greater responsiveness, and more efficient network-bandwidth use than Cloud platforms.The platform analyzes transit usage to evaluate routes and schedules and support adjustments.
  • Other Applications: MEC-based in-vehicle data analytics can support driver identification and personalized vehicle settings, automated use logs, insurance, and unauthorized-use detection.These applications target shared vehicles used by multiple drivers.
  • Other Applications: MEC-assisted traffic estimation addresses limited surveillance-camera availability, coarse GPS speed estimates, and bandwidth and latency challenges.The framework uses a front-facing camera and performs estimation near the edge and on vehicles.
  • Public Safety Video Analytics: AutoVAPS integrates body-worn cameras and public-safety sensors through data, model, and access layers.Its access layer supports privacy-preserving data sharing and access, while the model layer provides edge intelligence.
  • Edge AI Frameworks: Edge AI coordinates machine-learning training and joint inference among base-station and roadside-unit servers and onboard vehicle processors.The survey illustrates collaborative training, joint inference, and use cases including perception, mapping, and SLAM.

1) Federated Learning:

Federated learning addresses the communication and privacy challenges of training IoV models from distributed vehicle data, while edge servers can reduce update overhead. Joint device-edge inference extends this principle to computation-intensive vehicle applications, but task-specific deployment remains open.

  • Federated Learning: Federated learning avoids moving privacy-sensitive vehicle data to a central server, reducing the communication burden of distributed model training.It is identified for privacy-sensitive tasks such as driving-assistant speech recognition and infotainment.
  • Federated Learning: Edge servers can intermediate between vehicles and Cloud servers to reduce federated-learning communication overhead and latency.The paper describes aggregating multiple local updates before communicating with the Cloud.
  • Joint Device-Edge Inference: Joint device-edge inference partitions computation between vehicles and edge servers to balance vehicle resource limits against wireless offloading latency.Direct execution consumes substantial computation and energy, whereas offloading can be delayed by time-varying wireless channels and large data transfers.
  • Joint Device-Edge Inference: Layer-granular partitioning frameworks such as Neurosurgeon and Edgent adapt execution to DNN architectures, hardware, wireless conditions, and server load.Neurosurgeon demonstrated latency reduction and mobile-energy savings, while related work also studies accuracy–latency trade-offs.

1) Edge-Assisted Perception:

Edge assistance supports perception, mapping, localization, and SLAM by distributing computation and coordinating information among vehicles. The surveyed approaches use binary or partial offloading, edge aggregation, and cooperative data fusion to address vehicle resource and sensing limits.

  • Edge-Assisted Perception: Edge servers assist perception tasks such as object detection and tracking, especially when vehicles use powerful deep learning models.Offloading can be binary or partial, with decisions depending on channel state, server workload, and computation intensity.
  • Edge-Assisted Perception: Vision-based localization is particularly suitable for joint client-server processing because its highly parallel stages can use abundant edge GPU resources.Examples include feature extraction, disparity-map generation, optical flow, feature matching, and Gaussian blur.
  • Edge-Assisted Perception: Fresh HD-mapping data can be aggregated at nearby edge servers over V2I links before detecting road changes and updating maps.Local aggregation reduces the storage and communication burden associated with area-specific mapping data.
  • Edge-Assisted Perception: Edge-assisted SLAM places expensive map optimization and storage away from the vehicle while retaining lightweight local camera tracking.The C2TAM experiment used a desktop as the nominal Cloud, demonstrating the feasibility of edge-based SLAM.
  • Edge-Assisted Perception: Centralized multi-vehicle SLAM uses an edge server to aggregate and fuse vehicle data before returning results through V2I communications.Vehicles may build sub-maps locally and send them to a nearby edge server for fusion.
  • Edge-Assisted Perception: Vehicle sensing cooperation can broaden environmental awareness by sharing sensing information, perception outputs, or instantaneous 3D views.The surveyed motivation includes cooperative driving and see-through, lifted-seat, satellite, and all-around views.

2) Crowd-Sourced Mapping:

Crowd-sourced mapping uses data from many sensor-equipped vehicles to refresh HD maps, while layered management and cooperative SLAM address storage, communication, and computational demands. Multi-vehicle map integration remains technically challenging and requires further research and testing.

  • Crowd-Sourced Mapping: Crowd-sourced mapping can obtain fresher and broader HD-map data by working with automakers and collecting sensor measurements from many vehicles.The approach addresses the cost, labor, and limited road coverage of constructing maps through dedicated autonomous-vehicle testing.
  • Crowd-Sourced Mapping: Layer-based HD-map management makes additions, deletions, and corrections more efficient and reduces communication overhead by downloading only required layers.New feature layers can be added continuously as intelligent-vehicle requirements expand.
  • Crowd-Sourced Mapping: Decentralized SLAM lets each vehicle build a local map and communicate with other vehicles, but map alignment, pose estimation, uncertainty, complexity, and communication remain difficult.The cited work is largely from mobile robotics rather than intelligent IoV.
  • Crowd-Sourced Mapping: Decentralized SLAM merges aligned local maps into a global map and exchanges representations such as local, graph-based, or topological maps according to bandwidth.Real-time operation also requires handling out-of-sequence measurements caused by latency.
  • Crowd-Sourced Mapping: The survey presents edge caching, edge computing, and edge AI as components of an edge information system for acquiring, storing, and processing intelligent-vehicle data.It positions these capabilities alongside evolving onboard processors and Cloud platforms.
  • Crowd-Sourced Mapping: Intelligent IoV presents substantial opportunities and challenges requiring contributions from researchers, industry, governments, policymakers, and standardization bodies.The conclusion frames these efforts as necessary for developing technologies and policies for intelligent vehicles.
Loading 1906.00400v1…