1 Real-Time Dynamic Map with Crowdsourcing Vehicles in Edge Computing

2025-04-28 0 0 1.31MB 10 页 10玖币

侵权投诉

Real-Time Dynamic Map with Crowdsourcing

Vehicles in Edge Computing

Qiang Liu, Member, IEEE, Tao Han, Senior Member, IEEE,

Jiang (Linda) Xie, Fellow, IEEE, and BaekGyu Kim, Member, IEEE,

Abstract—Autonomous driving perceives surroundings with

line-of-sight sensors that are compromised under environmental

uncertainties. To achieve real time global information in high deﬁ-

nition map, we investigate to share perception information among

connected and automated vehicles. However, it is challenging

to achieve real time perception sharing under varying network

dynamics in automotive edge computing. In this paper, we

propose a novel real time dynamic map, named LiveMap to detect,

match, and track objects on the road. We design the data plane

of LiveMap to efﬁciently process individual vehicle data with

multiple sequential computation components, including detection,

projection, extraction, matching and combination. We design the

control plane of LiveMap to achieve adaptive vehicular ofﬂoading

with two new algorithms (central and distributed) to balance the

latency and coverage performance based on deep reinforcement

learning techniques. We conduct extensive evaluation through

both realistic experiments on a small-scale physical testbed and

network simulations on an edge network simulator. The results

suggest that LiveMap signiﬁcantly outperforms existing solutions

in terms of latency, coverage, and accuracy.

Index Terms—Dynamic Map, Edge Computing, Autonomous

Driving

I. INTRODUCTION

AUTONOMOUS driving and advanced driving assistance

system (ADAS) are being evolved with the development

of modern machine learning and pervasive parallel comput-

ing. Vehicles leverage a variety of sensors, e.g., camera and

LiDAR, to perceive surroundings, and use onboard computers

to understand the collected raw data in real time, e.g., semantic

segmentation and object recognition. With the high-deﬁnition

(HD) map, advanced vehicular control algorithms accurately

relocalize the vehicle and can tackle road situations with the

perceived environmental context, e.g., pedestrians and lanes.

Achieving highly reliable and safe driving, however, is very

challenging, based on non-real-time HD map and individual

vehicle perception. On the one hand, the HD map [2], in-

cluding geometric, semantic, and map-prior layer, has no real

time road information, e.g., pedestrian and vehicles, in the

time scale of subseconds. On the other hand, the perceptions

of individual vehicles are limited and might be compromised

Qiang Liu is with the School of Computing, University of Nebraska-

Lincoln. E-mail: qiang.liu@unl.edu

Tao Han is with the Department of Electrical and Computer Engineering,

New Jersey Institute of Technology. E-mail: tao.han@njit.edu

Jiang (Linda) Xie is with the Department of Electrical and Com-

puter Engineering, University of North Carolina at Charlotte. E-mail:

linda.xie@uncc.edu

BaekGyu Kim is with the Department of Information and Communication

Engineering, Daegu Gyeongbuk Institute of Science and Technology. E-mail:

bkim@dgist.ac.kr

Partial contents of this article appeared in IEEE International Conference

on Computer Communications 2021 [1].

LiveMap

Transportation Systems

Edge Servers

Radio Access Points

Fig. 1: An example of automotive edge computing.

under a variety of environmental uncertainties such as weather

and occlusion [3]. For example, existing line-of-sight vehicle

sensors are with limited sensing ranges, which indicates that

they cannot perceive information in occluded areas [4]. Con-

sidering a car follows a truck that blocks the car’s front sensor,

passing the truck without the information about the opposite

lane is unsafe.

Connected and automated vehicles (CAVs) emerge in re-

cent years to connect vehicles [5], [6] via advanced wire-

less technologies, e.g., 5G and beyond, with pervasive edge

computing infrastructures [7], [8], e.g., edge servers in radio

access networks (RAN). The Automotive Edge Computing

Consortium estimates that more than 50% of all cars on the

road in the United States will have connected features by

2025 [9]. Various onboard sensors of vehicles, e.g., cameras

and LiDAR, can be leveraged to construct global information

via crowdsourcing. By using edge servers as the hub, the

information perceived by individual vehicles is seamlessly

collected, processed, and shared among vehicles and infras-

tructures with ultra-low latency.

However, it is non-trivial to share perception data among

CAVs because of the constrained network infrastructures

and resources (e.g., spectrum and servers). For example,

the perception of vehicles may have duplicated information

due to their heavily overlapped sensing ranges in a dense

urban scenario. In addition, the uplink transmission of vehicle

perception data, e.g., point clouds, demands a tremendous

data rate which may overwhelm mobile networks [10]. Edge

servers, that support hundreds of vehicles if not more, experi-

ence fast-changing trafﬁc and workloads under varying vehicle

trajectories. Therefore, it is imperative to design intelligent

network management solutions to achieve real time perception

sharing under constrained network resources in automotive

edge computing.

In this paper, we propose LiveMap, a new real-time dynamic

map as shown in Fig. 1. LiveMap achieves the detection,

arXiv:2210.05034v1 [cs.DC] 10 Oct 2022

Object

Detection

Feature

Extraction

Combination

& Tracking

Object

Matching

detections

Global Map

features

RGB-D

image

Data

Acquisition

Service

Request

Object

Projection

configs

object

images

historical objects

Vehicle Server

Centralized scheme

Data

Plane

Control

Plane

objects

Local Map

objects

D-HEAD

Algorithm HEAD Algorithm

Sensors

Fig. 2: The overview of LiveMap. The data plane is to process sensor data for detecting, matching and tracking objects. The control plane

is to manage networks for accelerating transmission and computation of vehicle ofﬂoadings.

matching, and tracking of objects on the road in the time

scale of subseconds via crowdsourcing data from CAVs.

We design LiveMap to achieve an efﬁcient data plane for

vehicle data processing and an intelligent control plane for

vehicle ofﬂoading decisions. The data plane is composed of

object detection, object projection, feature extraction, object

matching, and object combination. In particular, we design to

improve the object detection with new neural network pruning

techniques, build concise feature extraction with variational

autoencoder techniques, optimize the feature matching with

a novel location-aware distance function, and increase the

combination accuracy with a new conﬁdence-weighted com-

bination method. The control plane enables adaptive vehicle

ofﬂoading, e.g., ofﬂoading the computations from vehicles

to servers, under varying network dynamics. We design two

algorithms, that apply in central and distributed scenarios,

to minimize the latency of ofﬂoadings while satisfying the

requirement of map coverage. We design these two algorithms

based on deep reinforcement learning (DRL) that optimize the

vehicle scheduling and ofﬂoading decision of individual CAVs.

In addition, we implement LiveMap on a small-scale physical

testbed with multipleJetRacers (Nvidia Jetson Nano), a 5GHz

WiFi router, and an edge server with Nvidia GPU.

The main contributions of this paper are listed:

•We design a new real time dynamic map (LiveMap) via

crowdsourcing sensor data of CAVs in automotive edge

computing networks.

•We develop an efﬁcient data plane with sequential process-

ing of sensor data for reducing the processing delay and

improving detection accuracy.

•We design an intelligent control plane with two new al-

gorithms that improve the latency performance without

compromising the map coverage.

•We develop an edge network simulator and prototype

LiveMap on a small-scale physical testbed.

•We evaluate LiveMap via both experiments and simulations,

and the results validate its superior performance.

II. LiveMap OVERVIEW

In Fig. 2, we overview the architecture of LiveMap, which

includes the data plane, i.e., sensor data processing for de-

tecting, matching and tracking objects, and the control plane,

i.e., network management for accelerating transmission and

computation.

The data plane is composed of several sequential process-

ing components. The acquisition component retrieves RGB-

D images from CAV sensors and its relocalization. The

detection component detects possible objects in RGB images

by exploiting state-of-the-art object detection framework, i.e.,

YOLOv3 [11]. The projection component projects the detected

objects from pixel coordinates to world coordinates based

on the depth information and camera-to-world transformation

matrix. The extraction component extracts visual features from

cropped object images by using a variational autoencoder. The

matching component matches detected objects in either the

local or global map according to their visual features and geo-

locations. The combination component combines multi-viewed

objects by integrating a variety of attributes, e.g., conﬁdence

and geo-location. Note that all components, except acquisition

and combination, can be ﬂexibly executed in either CAVs or

edge servers, according to the control plane. Finally, the global

map will be updated, where new updates will be broadcasted

to all vehicles for updating their local maps.

The control plane includes a central and a distributed

scheme. In the central scheme, a vehicle sends a service

request along with its local state to the edge server. The HEAD

algorithm optimizes the scheduling of this vehicle under the

current map coverage, and determines the ofﬂoading decision

with a central DRL agent under the global state if the vehicle

is scheduled. In the distributed scheme, the vehicle invokes the

D-HEAD algorithm independently to optimize its scheduling

and ofﬂoading decision according to the current local state.

The vehicle starts the data plane according to the ofﬂoading

decision if it is scheduled.

III. THE DESIGN OF DATA PLANE

The data plane is designed to efﬁciently process vehicle data

in terms of processing delay and detection accuracy.

A. Data Acquisition

We develop the acquisition component to acquire the sensor

data, i.e., LiDAR and RGB-D images. Without loss of gen-

erality, we consider the RGB and depth images from RGB-

D cameras, e.g., Intel RealSense D435. Besides, it obtains

the accurate vehicle location in world coordinates, which de-

pends on either high-accuracy GPS or advanced relocalization

algorithms such as ORB SLAM2 [12]. The accurate vehicle

location is necessitated for combining the multi-viewed objects

detected by multiple vehicles.

B. Object Detection

We design the detection component to detect transportation

objects from RGB images, e.g., trucks and pedestrians, where

detection results include the object classes, probability, and 2D

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1Real-TimeDynamicMapwithCrowdsourcingVehiclesinEdgeComputingQiangLiu,Member,IEEE,TaoHan,SeniorMember,IEEE,Jiang(Linda)Xie,Fellow,IEEE,andBaekGyuKim,Member,IEEE,AbstractAutonomousdrivingperceivessurroundingswithline-of-sightsensorsthatarecompromisedunderenvironmentaluncertainties.Toachieverealtimegl...

展开>> 收起<<

1 Real-Time Dynamic Map with Crowdsourcing Vehicles in Edge Computing.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Real-Time Dynamic Map with Crowdsourcing Vehicles in Edge Computing

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: