1

An Energy-Efficient High Definition Map Data

Distribution Mechanism for Autonomous
Driving
Jinliang Xie, Jie Tang, Member, IEEE and Shaoshan Liu, Senior Member, IEEE

Abstract Autonomous Driving is now the promising future of transportation. As one basis for autonomous driving, High Definition
Map (HD map) provides high-precision descriptions of the environment, therefore it enables more accurate perception and
localization while improving the efficiency of path planning. However, an extremely large amount of map data needs to be
transmitted during driving, thus posing great challenge for real-time and safety requirements for autonomous driving. To this end,
we first demonstrate how the existing data distribution mechanism can support HD map services. Furthermore, considering the
constraints of vehicle power, vehicle speed, base station bandwidth, etc., we propose a HD map data distribution mechanism on
top of Vehicle-to-Infrastructure (V2I) data transmission. By this mechanism, the map provision task is allocated to the selected
RSU nodes and transmits proportionate HD map data cooperatively. Their works on map data loading aims to provide in-time HD
map data service with optimized in-vehicle energy consumption. Finally, we model the selection of RSU nodes into a partial
knapsack problem and propose a greedy strategy-based data transmission algorithm. Experimental results confirm that within
limited energy consumption, the proposed mechanism can ensure HD map data service by coordinating multiple RSUs with the
shortest data transmission time.

Key words—Autonomous driving, HD Maps, Map Data distribution, Energy Efficiency

obstacle avoidance. Besides, HD map supports personalized

I. INTRODUCTION1 driving, including various driving behavior suggestions, the

A UTONOMOUS driving is one of the most promising

applications in artificial intelligence and intelligent
best acceleration point suggestions, the best cornering speed
suggestions, etc. All of them can be used to improve the
safety and comfortability of autonomous driving.
transportation. It liberates human drivers, alleviates traffic
congestion, offers great convenience as well as driving safety, In this paper, we present a mechanism that provides real-
and finally reduces environmental pollution [1]. High time HD MAP services for autonomous driving under tight
Definition Map (HD map) is one basis for autonomous real time and energy budget. While literature is rich with data
driving [2], [3]. Compared with traditional electronic maps, provision designs under Internet of Vehicle networking,
HD map has the characteristics of high-precision data, high- prior efforts are mostly targeting ordinary command data in
dimensional information, and high real-time response. With small volume. For example, early warning data, traffic
the extension of sensing range and the improvement of notification and instruction data. However, we find that the
sensing accuracy, HD map can build up a more HD Map service for autonomous driving operates under
comprehensive autonomous driving guidance in the way of totally different scenarios. It works with great constraints in
perception, positioning, and decision-making [4], [5]. high responsibility and high data volume. For instance, as the
By HD Map service, we can have the centimeter-level supplement of real-time perception, the response of HD map
high-precision positioning for autonomous driving vehicles should be less at least than 80ms on average to fit in
[6]. Using prior knowledge in map, the HD Map can reduce autonomous driving pipeline [8] and its data volume per
the difficulty of environment perception [7]. It can also second can be over 2G per second in some high speed
broadcast various road information in advance, realize more scenarios.
effective driving planning like early vehicle deceleration and With this background, we firstly quantitatively
demonstrate that the existing data distribution services
Xie, J., is with the South China University of Technology, does not fit the scenarios for HD map service (Sec. 3). The
Guangzhou 510641, China. (e-mail: 201821033740@[Link]). in-vehicle disk+memory solution can support the HD map
Tang, J., is with the South China University of Technology, service for only few minutes. Therefore, maps are usually
Guangzhou 510641, China. (e-mail: cstangjie@[Link]). She is the
corresponding author of this paper. being served by online services, and only a nearby small
Liu, S., is with PerceptIn, Santa Clara, CA 95054. (e-mail: areas of the HD Map (called submaps) are downloaded to the
shaoshanliu@[Link]). vehicle when needed. In real-time V2V data transmission,
 theoretically there need to contact up to hundreds of cars to traffic condition, data size can impact the efficiency of the
fill the map data tank and thus posing a great challenge for proposed mechanism (Sec.5). Experimental results show
limited peer-cars and limited bandwidth in real-world that the mechanism performs better in both time and energy.
scenarios. In V2I based solution, with proactive data Meanwhile, the number of RSU nodes can be greatly reduced,
provision, it can deliver the feasible HD Map service. when there is much traffic, the RSU nodes hit is distributed
However, it leaves a question that how the service can be more evenly.
provided with multiple RSU nodes in cooperation. The main contributions of this paper are:
With data transmission scenarios carefully calculated, we 1) We quantitatively show that no single mechanism fits
propose a HD map data distribution mechanism on top of the demands and constraints of HD Map service. The
V2I data transmission (Sec 4.). The proposed solution analysis provides a desirable acceleration target for the HD
considers the energy consumption of vehicles as well as the map service.
bandwidth sharing between them. By this mechanism, the 2) We propose a V2I based HD map distribution
map provision task is allocated to the selected RSU and mechanism. It assigns map data provision task to selected
sector of HD map data is transmitted proportionately. RSU nodes and makes them co-complete the map
We model the selection of RSU nodes into a NP-hard transmission for target area.
back-pack problem and divide the multi-agent question 3) We model the data provision task designation into a
into multiple one-agent question (Sec. 4). We further partial backpack problem and design a heuristic algorithm to
propose a heuristic algorithm to solve the selection by the solve it in the complexity of O(logn).
complexity in O(logn), which ensures that the decision of 4) We run the simulation experiments and verify the
node selection and their data task allocation can be output in performance of proposed mechanism
real-time.
To show its generality, we discuss how RSUs density,

Fig. 1. The role of HD map in autonomous driving

TABLE I
The HD Map Information
The map content Purpose Data Source Data size (per mile)
2D Orthographic Reflectivity Map Localization Lidar >5 GB
Digital Elevation Model (DEM) Planning&Sensing Sonar/Radar 100 MB
Lane/Road Model Planning&Localization Camera/ Sonar 10 MB
Semantic Layer Planning&Sensing Camera 50 MB

the road curbs, the grade/steepness of a ramp or hilly road etc

II. BACKGROUND OF HD MAP [12].
HD maps are the foundation for autonomous driving 3) Lane/Road Model
implementation. It usually has multiple layers and provides Lane/road model contains the semantics of lane segments
a full stack of information for autonomous vehicles. Layers and road segments. Lane model contains information of lane
of HD Maps have different representations, data structures geometrics (boundaries, width, curvature etc.), lane type (car
and purposes. Usually, HD Maps contain the following 4 lane, bike lane, bus-only lane etc.), lane directionality, lane
layers [9] and we show each size in Table I: marking/divider types (solid vs dashed, single vs double etc.),
1) 2D Orthographic Reflectivity Map restrictions (e.g. left/right turn only), speed limits,
This layer is a 2D planar view of the road surface connectivity between lanes, etc. Lane/road model is critical
extracted from the LiDAR 3D point clouds [10], [11]. for motion planning, vehicle control etc.
2) Digital Elevation Model (DEM) 4) Semantic Map
DEM is a 3D model and contains the height information It is usually a versatile layer that stores the semantics of
of surface of the driving environment, such as the height of static elements in the driving environment, e.g. traffic lights
 3

and their association with lanes, road obstacles etc. downloaded submaps will be erased right after the cars drive
out the area those data maps to. Due to the limited caching
III. MOTIVATIONS room, the used map data will become invalid in a very short
time. As a result, the original “carry-store-forward” data
A. Overview on Data Distribution Solutions
transmission method in V2V cannot work effectively.
C. Shortage of v2v Data Transmission
1) The Analysis for ideal situation

In the V2V transmission, it is assumed that there are two

vehicles Car1 and Car2 running at a constant speed v1 and
v2 in a bidirectional lane of a highway. Here, Car1 is the
source vehicle, Car2 is the target vehicle. If the vehicle uses
DSRC communication and its communication range is r
(generally 100-200 meters), the distance between the two
vehicles in their driving direction is d (d＜r), the vehicle data
transmission rate is R.
According to 3-1, the contact time for them within their
communication range is T, and the amount of data
transmitted during this time is C:
2√𝑟 & − 𝑑&
Fig. 2. V2V and RSU cooperative transmission (If vehicle D would like 𝑇= (3 − 1)
to get data from A: by V2V, the data will be sent by multi-hop through
𝑣* + 𝑣&
A->B->C->D; by V2I, A->RSU1-> RSU2 ->D ) 𝐶 = 𝑇𝑅 (3 − 2)
Here, R is a function of the vehicle distance D, the DSRC
The data distribution methods in vehicle application can channel bandwidth B, the DSRC channel attenuation
be roughly divided into three types: V2V (Vehicle to coefficient h, the channel noise power spectral density N0 and
Vehicle), V2I (Vehicle to Infrastructure), and V2X (Vehicle the power for vehicle data transmission PS. To simplify the
to Vehicle or Infrastructure), that is a fusion of V2V and V2I. &67 8 9: 8
model, it assumes Rmax = B, we can get 𝐶234 ≤ ×
V2I is the communication between the vehicle and the ;< =;8

infrastructure by roadside, like RSU. 𝑅234 .Then we have:

V2V is used for the communication between cars. The 𝐷𝑃F |ℎ|
𝑅 = 𝐵𝑙𝑜𝑔& C1 + K (3 − 3)
DSRC based method is the mainstream in the data 𝑁J
distribution by V2V now. V2V uses the mechanism of “carry When the amount of map data to be transmitted is G, in an
store forward”. The vehicle can receive data from the nearby ideal situation, there is always a vehicle participating in data
vehicles, carry the data in its traveling and make another data transmission and the carry-store-forward of data is always
forwarding when it meets some target vehicle. working, we need to have 𝑁𝑢𝑚 = 𝐺O𝐶 vehicles to travel
In most cases, V2I will cooperate with V2V to form a 234
at lease 𝑁𝑢𝑚 × 𝑉𝑇 mile for map data acquisition.
V2X data distribution solution. RSU can be connected to one
2) The Analysis for Actual situation
or more Traffic Information Centers (TIC) through the
In the real-life situation, if the data is received from
backbone network. It greatly increases the amount of data
vehicles in the opposite direction, the probability that
transmitted in the vehicle network. V2V increase the chance 27
that data can be successfully delivered to dedicated vehicles participate in the transmission is 𝑝 = ;S . When
destination. there has m vehicles in the reverse lane within time duration
t. Thus, the distance used to complete the data transmission
B. In-Vehicle Storage for HD Maps TU2×VW
is .
Autonomous vehicles consume HD Map data in a X
considerably high rate. It can be seen from Table 1 that the If the V2V data is obtained from cars met at different lanes
amount of map data consumed per second during driving is in the same direction, where they travel at varies speeds.
over 300MB. However, that is the data cost from only one of Only in this way, it can have can different data fragments
optional roads in planning. In driving, the vehicle will be fed from different vehicles. The contact time of the two vehicles
&7
with the map data for all possible planning roads in some becomes 𝑇 = |; 9; | , then there requires a distance of
< 8
destination range. Thus, from a storage point of view, such TU2×VW
consumption will digest 4T hard disk data quickly in just half to transmit G data.
X
hour. Therefore, it is impossible for the vehicle to store all
the map data of the target area in advance. HD map data
should be divided into submaps and provided by some online D. V2I
service. Compared with V2I, the data transmission coverage of
From the perspective of temporary data caching, the infrastructure node is much wide. Normally, it is 3.X-5.X
wider compared with the one of V2V. Thus, the contact time B. Modeling the Multi-vehicle map data distribution
of them can be extended correspondingly. However, there Supposing there are in total m car in the lane that require
still needs couples of RSU to participate in the map HD map data, the amount of data that each vehicle needs to
transmission due to its large volume. RSU nodes are always transmit is M, and the remaining energy of the vehicle j is
connected by a backbone network, so that one can get a data 𝑤Z[37 . Assume that all vehicles are electronical, the energy
copy from other nodes within a trivial time slice. Therefore,
the problem in V2I turns to be how some data distribution consumption of a traveling electric vehicles is [13]:
S`
mechanism can reasonably designate transmission tasks to ∫ 𝑃^ 𝑑𝑡
those RSUs. 𝐸= J (4 − 1)
𝑆
E. Conclusion and Observations Where E is the energy consumption per unit mileage, t J is
the driving time, S is the driving mileage and 𝑃^ is the power
From the above analysis, we can get some observations:
requirement of driving a car.
First, in the case of limited in-vehicle storage, vehicles Energy consumed by vehicle j receiving map data from
must use online data provision solutions. f
RSU node i is 𝐸d,Z . As shown below, it is a function of the
Second, the amount of map data is large, it is impractical
for the vehicle to obtain data only through V2V. V2V’s high amount of data, the vehicle transmission rate, and the
energy consumption also cannot support this degree of data distance between RSU node and vehicle lanes.
f f
transmission. 𝐸d,Z = 𝑓(𝑀d , 𝑅d,Z ) (4 − 2)
Third, the downloaded map data will be invalid in a very From the above analysis, the cost that vehicle can
short time, which makes the original “carry-store-forward” autonomously navigate through the area is the sum of the
V2V transmission impractical. energy used for driving through it and the energy for the map
Last, V2I is a comparative feasible solution for HD map data transmission of the area. Assuming that the remaining
data service, however, an effective and efficient data energy of car i is 𝑊d[37 , to enable autonomous driving, it
distribution mechanism is required, and this is the main needs to ensure that:
purpose of this paper. lm
f
j 𝐸d,Z + k 𝐸 𝑑𝑠 < 𝑊d[37 (4 − 3)
IV. MAP DATA DISTRIBUTION MECHANISM J
The transmission rate is related to the bandwidth,
A. V2I Based Solution transmission power, transmission distance, and path loss.
According to the literature [14], it is assumed that the
maximum downlink transmission power of the edge node is
p p p
𝑃234 , and 𝑃d,Z 𝑃d,Z represents the downlink transmission
power allocated by the edge node i for its communication
with vehicle j, here 0 ≤ 𝑃dp，Z ≤ 𝑃234 p
. Path fading can be
-u
expressed as ds , where ds,v and σ respectively represent the
distance between RSU node i and vehicle j, as well as the
path loss coefficient. The receiving channel bandwidth of
vehicle j is 𝐵Zf , the attenuation coefficient of the receiving
channel is ℎ* , and the power spectral density of noise in the
channel is 𝑁J . According to Shannon's formula[15], the rate
at which vehicle j uses to receive data from edge node i is:
Fig. 3. The overview for V2I based HD Map Service
f 𝑑d9x 𝑃dp，Z |ℎ* |
V2I based map data transmission exhibits great potentials 𝑅d,Z = 𝐵Zf 𝑙𝑜𝑔& C1 + K (4 − 4)
𝑁J
in serving HD map data in real time. However, there may
co-exist several vehicles which all require the map data from Under the multi-vehicle model, it is assumed that the
the same RSU node concurrently. When the number of probability of any two vehicles meeting within the coverage
vehicles exceeds some threshold, the data transmission will of the same RSU node is p. Thus, the probability of k vehicles
be much slowed down due to the transmission rate inference. simultaneously in the same RSU follows the Poisson
Therefore, the vehicle must continue to search the next distribution:
possible data source. 𝑒 92| (𝑚𝑝)}
𝑃(𝑋 = 𝑘) = (4 − 5)
Thus, as stated in Fig 3, the V2I based solution should be 𝑘!
able to find a series of αi, 0<αi<1 & ∑αi =1 for each RSU Since multiple vehicles will compete for the bandwidth,
node in the area, which is responsible for providing map data they will be interfered by the surrounding vehicles when
service proportionally to the designated αi. With well transmitting data. Thus, the transmission rate will be:
consideration of constraints like resource sharing and energy 𝑑d9x 𝑃dp，Z |ℎ* |
cost, it aims to have the transmission time minimized without
€
𝑅d,Z (𝑋 = 𝑘) = 𝐵Zf 𝑙𝑜𝑔& •1 + … (4 − 6)
𝑁J + ∑}Zƒ„Z 𝑑d9x 𝑃dp，Z ƒ |ℎ* |
breaking the energy budget.
According to the analysis, the multi-vehicle map service
problem with energy limits can be transformed into the
 5

following optimization problem. That is to minimize total the model of multi-vehicles. Because we should find the
data transmission time with the subject to the energy optimal time for m vehicles, thus the algorithm complexity
limitation.: is O(mnlogn). The pseudo-code of the algorithm is as
T follows.
Z
𝑚𝑖𝑛 𝑚𝑎𝑥 ‹j 𝑡d , 𝑗 = 1,2, … , 𝑚• (4 − 7)
dŒ*
subject to:
T 29*
Z f
j 𝑡d • j 𝑃(𝑋 = 𝑘)𝑅d,Z (𝑋 = 𝑘)… ≥ 𝑀Z (4 − 8)
dŒ* }Œ*
0 ≤ 𝑡dZ ≤ 𝑇d , 𝑖 = 1,2, … , 𝑁 (4 − 9)
2
f
j 𝐸d,Z < 𝑊Z[37 (4 − 10)
ZŒ*
The final result is the tuple <𝑐dZ ，𝑡dZ > for each vehicle. 𝑐dZ
is an expression in <0,1>, representing whether vehicle j will
transmit data transmission with RSU node i when it passes
by. 𝑡dZ represents the time duration vehicle j used to transmit
data with node i.
C. Multi-vehicle Map data distribution solution
To solve this model, we can first decompose the multi-
vehicle problem into a single-vehicle problem. The, we
can find the optimal solution locally, and combine all the
solutions together to get a globally optimal solution for the
multi-car problem. For the solution of the single-vehicle
problem, it can be assumed that there is only one vehicle j
obtaining data from <RSU List>. Each RSU node in <RSU
List> can be selected or not be selected, and the total data
amount transmitted in its limited time is no less than M.
Besides, the remaining energy consumption of the vehicle
must ensure driving safety. Thus, it can be considered as a
variant of a knapsack problem.
The backpack problem can be divided into 0-1 backpack
problems and partial backpack (fractional knapsack)
problem. In our transmission model, even if the RSU node
is selected, only part of the map data can be transmitted, so
our question on map distribution can be regarded as a partial
backpack problem. The partial backpack problem itself is an
NP-complete problem with a complexity of O(2n), so it
cannot find out an optimal solution by force.
We propose a heuristic algorithm named ETDM
(Energy-Time Distribution Mechanism) that uses the greedy
strategy to find the optimal solution. The idea of the ETDM
is to sort the optional RSU nodes in descending order by the
average transmission rate each gives. The optimal
complexity of the sorting algorithm can be O(nlogn). After
sorting, a new bandwidth sequence is obtained
( 𝑅′* , 𝑅′& , … , 𝑅′} ), then the transmission tasks can be
allocated proportionally according to the bandwidth from
high to low. The RSU node finally assigned some the
transmission task may not use up its time contacting with
vehicles. V. EXPERIMENT AND CONCLUSION
To this end, we should find a time-optimal solution for a
A. Experiment setup
fractional backpack problem and combine them into a
solution for multi-car model. The maximum value of the In this paper, we implement the simulation on MATLAB
optimal solution of all vehicles will be taken as a result for platform, and the traffic data is extracted from a 10-minute
traffic segment of a crossroad from the Luxembourg SUMO First, we evaluate the impact form the data volume. We
traffic (LuST) data set [16]. The vehicles drive through the have simulated 251 vehicles, and all of them have
crossroad into three individual areas, areas A, B, and C. A communicates with the same 60 RSU nodes. We gradually
total of 251 vehicles were detected within 10 minutes, where increase the amount of required data in each experiment,
there had 95 vehicles going into area A, 94 vehicles going which increases from 140G to 300G, and the remaining
into area B, and 62 vehicles going into area C. Here, we energy capacity of the vehicle is set to be 5KWh.
assume the vehicle power of data transmission is 10W and From Fig.4, we can see when the data volume increases
the vehicle battery consumption rate is between 0.15kwh and into above 190G, all algorithms have a noticeable cliff-like
0.25kwh. We randomly generate 60 RSU nodes, and all decline in transmitted data amount. That is because the
parameters of each RSU node like bandwidth, coverage energy capacity of the vehicle is not enough to support full-
radius, path attenuation coefficient in transmission, RSU scale data transmission. At that time, the vehicle turns to
transmission channel noise, etc. were all randomly generated transmit just the underlying layers of map, so there comes a
within their reasonable range. For example, the bandwidth is declining cliff in data volume.
set to be 1 ~ 2GBps, the coverage radius is 100 meters, the As stated in Fig. 5, when the vehicle transmits the same
average distance between each RSU node is 100 meters, the amount of data before the cliff, the proposed ETDM
distance between the RSU node and the road lane is 20 ~ 70 algorithm can significantly save transmission time by 34.8%
meters, and the RSU transmission channel noise power on average. The number of RSU nodes the ETDM algorithm
spectral density is set to be 0.3. used is also on average 6-7 nodes less than the other three
algorithms.
B. Data distribution experiment
The baseline of the evaluation includes: The Original
Algorithm (OA), which is to transmit the map data by the
sequence that vehicles arrives the RSU nodes. Every time a
vehicle enters into the transmission range an RSU node
covers, as long as it contains the required map sectors, the
data transmission can be started. In the Probabilistic
Transmission Algorithm (PTA) ，as long as the RSU node
contains the required map data, it starts to transfer data by
some probability. For example, in a 0.7 probability
transmission algorithm, the probability that the vehicle
communicates with RSU node is 0.7. In our experiments, we
configure the probability in 0.3/0.5/0.7 respectively.
The evaluation metrics include the optimal transmission
time, the average transmission time of each vehicles under
different traffic, the transmission time of a single vehicle Fig. 5. Influence of data volume on the number of transmission base stations
under different traffic, the number of RSU nodes involved in
each transmission, and the map data hit rate in RSU nodes, 2) Influence of traffic flow on transmission time
which is calculated by the variance of the hit rate of all nodes.
We also evaluate how the amount of data and the location of The experiment on traffic is implemented by increasing
the RSU nodes impact on the ETDM algorithm the number of vehicles from 10 vehicles to 250 vehicles by
adding 10 vehicles each time. We can find the shortest
1) Influence of the amount of transmitted map data transmission time and the average transmission time for all
vehicles in each configuration. Here, we only count the
vehicles that have finished full map data transmission.
From the experimental results, no matter it is an optimal
transmission time (Fig. 6) or an average transmission
time(Fig. 7), the performance of the ETDM algorithm is
superior to the other three: when taking about the optimal
time, the ETDM algorithm saves time in 24.5%; when
talking about the average transmission time, the ETDM
algorithm offers 21.1% time saving compared with the OA
and 18.2% time saving compared with the PTA. The Fig. 8
shows the effect of different traffic on the transmission time
for a single vehicle. It can be seen that one vehicle can save
an average of 24.1% transmission time if it applies the
ETDM algorithm.

Fig. 4. Influence of data volume on algorithm transmission time

 7

of the hit rate is larger than others. This is because when the
traffic is light, the ETDM tends to select the nodes having a
larger bandwidth. As a comparison, the OA uses the FCFS
thus its hit rate variance is prone to be large; and the PTA
selects each node with the same probability thus its hit rate
variance is comparatively small. However, when the traffic
is heavy, the variance of the hit rate of the ETDM algorithm
turns to be the smallest one, indicating that the ETDM
algorithm has a relatively evenly distributed RSU hit rate

Fig.6. Optimal transmission time under different traffic

Fig. 9. Influence of traffic flow on the number of hit base station

Fig. 7. Average transmission time under different traffic

Fig. 10. Influence of traffic flow on base station hit rate

VI. RELATED WORK

The previous data service in the Internet of Vehicles is
mainly focusing on the data content distribution and content
caching deployed by RSU. In [17], the author considered the
position of the access points as well as the moving path of
Fig.8. The transmission time of single car under different traffic
vehicles, then decided a way how data can be forwarded from
3) Influence of traffic on usage of RSU node a stationary point to a moving vehicle. It also
How many accesses an RSU node gets depends on the comprehensively selected the locations for data transmission.
current traffic condition. In theory, the nodes with faster In [18], authors use network coding techniques to give
transmission speed will be accessed more frequently. Fig. 9 efficient data distribution between multiple RSUs. In [19], it
shows the statistics of the number that edge node is accessed studied the cost-effective planning problem of heterogeneous
under different traffic, as well as the variance of the hit rate vehicle networks, that is composed of traditional macro base
in Fig. 10. It can be seen that when the traffic is light, the stations and RSUs with buffering supports. Literature [20]
number of accessed nodes under the ETDM algorithms is focused on the cloud based VANET architecture and its
much less than that of other three algorithms, but its variance content retrieval. It proposed a cache placement strategy that
considering the on-vehicle caching layer and RSU caching REFERENCES
layer. In [21], the author introduced the problem of content [1] Liu, S., Li, L., Tang, J., Wu, S. and Gaudiot, J.L., 2020. Creating
distribution and caching in RSU, as well as proposed a Autonomous Vehicle Systems. Synthesis Lectures on Computer
Science, 8(2), pp.i-216.
greedy algorithm to traverse the cached data with reasonable
[2] NEDEVSCHI S, POPESCU V, DANESCU R, et al. Accurate Ego-
cost. Literature [22] studied the deployment of RSU in the Vehicle Global Localization at Intersections through Alignment of
highway. It comprehensively considered factors of wireless Visual Data with Digital Map[J]. IEEE Transactions on Intelligent
interference, traffic distribution, and vehicle speed to Transportation Systems, 2013, 14(2): 673-687.
DOI:10.1109/tits.2012.2228191
maximize the data throughput in IOV. Similar to literature [3] Yong He. Research on GPS Navigation Method Based on High-
[22], [23] studies the deployment of RSU in urban scenarios. resolution Map[D].Shanghai: Shanghai Jiao Tong University, 2015
The author comprehensively modeled the traffic in the urban [4] SUGANUMA N, UOZUMI T. Precise Position Estimation of
Autonomous Vehicle Based on Map-Matching[C]//2011 IEEE
as well as the probability of accidents at different locations. Intelligent Vehicles Symposium (IV). Germany: IEEE,2011: 296-301.
By designing a position awareness mechanism, such data DOI:10.1109/IVS.2011.5940510
service can efficiently minimize the accident reporting time [5] LI H, NASHASHIBI F, TOULMINET [Link] for Intelligent
Vehicle by Fusing Mono-Camera, Low-Cost GPS and Map
in the city. Data[C]//13th International IEEE Conference on Intelligent
There is little research on content distribution and caching Transportation Systems. Portugal:IEEE, 2010: 1657-1662.
in V2V communication. For data caching, they mainly talked DOI:10.1109/ITSC.2010.5625240
[6] RESS C, ETEMAD A, KUCK D, et al. Electronic Horizon —
about its application in Content Centric Networks (CCN) or
Providing Digital Map Data for ADAS applications [J]. Madeira, 2008,
Named Data Networking (NDN). In [24], author proposed a (3):40-49
node selection algorithm based on the theory of minimum [7] LEVINSON J, MONTEMERLO M, THRUN S. Map-Based Precision
vertex cover set in a static network and proposed a Vehicle Localization in Urban Environments [M]. Cambridge: MIT
Press, 2007
cooperative cache mechanism. It used social relations to [8] B. Yu, W. Hu, L. Xu, J. Tang, S. Liu, and Y. Zhu, “Building the
select cache points and solve the connection failure caused computing system for autonomous micromobility vehicles: Design
by vehicle movement. Literature [25]-[27]introduced many constraints and architectural optimizations,” in 2020 53rd Annual
IEEE/ACM International Symposium on Microarchitecture (MICRO),
basic caching strategies in CCN, but most of these caching IEEE, 2020.
strategies are more suitable for static network scenarios. [9] Liu, S., 2020. Engineering Autonomous Vehicles and Robots: The
Literature [28]-[30] studied the use of caching technology in DragonFly Modular-based Approach. John Wiley & Sons.
[10] J. Levinson, M. Montemerlo and S. Thrun, "Map-Based Precision
wireless networks to support user mobility and improve Vehicle Localization in Urban Environments", Robotics: Science and
content distribution. Literature [31] reviewed the Systems, vol. 4, pp. 1, June 2007.
opportunities and challenges of edge computing for [11] Jesse Levinson, Automatic laser calibration mapping and localization
for autonomous vehicles, 2011.
autonomous driving from a security perspective, reviewed [12] [online] Available: [Link]
state-of-the-art approaches in these areas, and discussed self-driving-car-723b4d9cd3f4.
potential solutions to these challenges. [13] Lei Yuan ． Analysis of Energy Consumption Economy and
Sensitivity of Pure Electric Buses[J]．Mechanical Engineering and
VII. CONCLUSION. Design：Chinese and English version，2013(1):1-7．
[14] Teng Yang．Research on Collaboration of Computing Nodes and
HD map is the basis for autonomous driving and map Virtual Machine Migration in Ultra-Dense Networks[D] ． Beijing
data service is crucial for real-time and safety for University of Posts and Telecommunications，2018．
autonomous driving. In this paper, we quantitatively [15] Baoyu Tian，Jie Ynag，Zhiqiang He，XiaoXiang
identified that the existing data transmission solutions cannot Wang．Information Theory Foundation[M]．Beijing：People Post
guarantee the quality of HD map data service. To solve this Press，2008．
[16] Aali S N , Shahhosseini H S , Bagherzadeh N . Divisible Load
problem, we propose a HD map data distribution mechanism Scheduling of Image Processing Applications on the Heterogeneous
on top of V2I data transmission. Our proposed method Star Network Using a new Genetic Algorithm[C]// 2018 26th
divides the map provision task to selected RSU nodes and Euromicro International Conference on Parallel, Distributed and
has map data proportionately transmitted. By multiple RSUs Network-based Processing (PDP). IEEE Computer Society, 2018.
[17] Jeong J, Guo S, Gu Y, et al. TSF: Trajectory-Based Statistical
cooperation, the HD map service can be delivered in-time Forwarding for Infrastructure-to-Vehicle Data Delivery in Vehicular
with optimized in-vehicle energy consumption. Here, the Networks[C]. IEEE International Conference on Distributed
problem that map data distributed between nodes is modeled Computing Systems, Genova, Italy. 2010:557-566.
into a partial knapsack problem and we propose a greedy [18] Ali G, Rahman M, Chong P, et al. On Efficient Data Dissemination
Using Network Coding in Multi-RSU Vehicular Ad Hoc
algorithm ETDM as solution. ETDM turns the multi-agent Networks[C]. IEEE Vehicular Technology Conference, Nanjing,
question into a multiple one-agent question and can generate China. 2016:1-5.
the allocation decisions with complexity of O(logn). [19] Zhang S, Zhang N, Fang X, et al. Cost-effective vehicular network
Experimental results show when the remaining energy planning with cache-enabled green roadside units[C]. IEEE
International Conference on Communications, Paris, France. 2017:1-
capacity of the vehicle is limited, the proposed method can 6.
provide shorter map data transmission time, more balanced [20] Ma J, Wang J, Liu G, Fan P. Low Latency Caching Placement Policy
RSU accesses even when the traffic is heavy. for Cloud-Based VANET with Both Vehicle Caches and RSU
Caches[C]. 2017 IEEE Globecom Workshops (GC Wkshps),
Singapore. 2017:1-6.
[21] Ding R, Wang T, Song L, et al. Roadside-unit caching in vehicular ad
hoc networks for efficient popular content delivery[C]. 2015 IEEE
 9

Wireless Communications and Networking Conference, New

Orleans, LA, USA. 2015:1207-1212.
[22] Wu T J, Liao W, Chang C J. A cost-effective strategy for road-side
unit placement in vehicular networks[J]. IEEE Transactions on
Communications, 2012, 60(8):2295-2303.
[23] Aslam B, Amjad F, Zou C. Optimal roadside units placement in urban
areas for vehicular networks[C]. IEEE Symposium on Computers
and Communications, Cappadocia, Turkey. 2012, pp.423-429.
[24] Liu L, Xie D, Wang S, et al. CCN-based cooperative caching in
VANET[C]. IEEE International Conference on Connected Vehicles
and Expo, Shenzhen, China. 2016:198-203.
[25] Ahmed S H, Bouk S H, Yaqub M A, et al. CODIE: Controlled Data
and Interest Evaluation in Vehicular Named Data Networks[J].
IEEE Transactions on Vehicular Technology, 2016, 65(6):3954-3963.
[26] Nakamura N, Niimi Y, Ishihara S. Live VANET CDN: Adaptive
data dissemination scheme for location-dependent data in
VANETs[C]. IEEE Vehicular NETWORKING Conference, Boston,
MA, USA. 2013:95-102.
[27] Pang B, Li R, Yue P. Research of the interest packet popple broadcast
diffusion attack in NDN VANET[R]. the Symposium. 2017:617-620.
[28] Song J, Song H, Wan C. Optimal Content Placement for Wireless
Femto-caching Network[J]. IEEE Transactions on Wireless
Communications, 2016, PP(99):1-1.
[29] Zhan C, Wen Z. Content Cache Placement for Scalable Video in
Heterogeneous Wireless Network[J]. IEEE Communications Letters,
2017, PP(99):1-1.
[30] Zhao Z, Peng M, Ding Z, et al. Cluster Content Caching: An Energy-
Efficient Approach to Improve Quality of Service in Cloud Radio
Access Networks[J]. IEEE Journal on Selected Areas in
Communications, 2016, 34(5):1207-1221.
[31] Liu, S., Liu, L., Tang, J., Yu, B., Wang, Y. and Shi, W., 2019. Edge
computing for autonomous driving: Opportunities and
challenges. Proceedings of the IEEE, 107(8), pp.1697-1716.