Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Autonomous Vehicle as an Intelligent

Transportation Service in a Smart City


I-Cheng Lin*, Che-Yu Lin*, Hsuan-Man Hung*, Qimei Cui†, Kwang-Cheng Chen¶, Fellow, IEEE
*Graduate Institute of Communication Engineering, National Taiwan University, Taipei, Taiwan

Beijing University of Posts and Telecommunications, Beijing, China

University of South Florida, Tampa, United States
E-Mail: *{r04942067, r04942097, r04942113} @ntu.edu.tw, ¶kwangcheng@usf.edu
Abstract—Autonomous vehicles (AVs) and crowdsourcing big traveling time to reach the destination. Given the QoS, the
data analytics may dramatically change future intelligent required number of AVs should be minimized in order to
transportation in smart cities. AVs may evolve more like a service maximize the utilization of each AV.
than a product. To provide best user experience of such service,
To investigate the optimal operation of such brand new
three primary factors including waiting time, travel time, and
supply of AVs are taken into consideration for multi-objective service, we can construct a queuing network model based on
optimization facilitated in three steps. With the queueing network the road topology, to jointly optimize the waiting time, traveling
model and traffic flow analysis, optimal operation of service could time, and operational cost. There are further technological
be achieved with minimum average travel time. The optimal challenges in this study and we employ techniques to resolve
number of available AVs could be identified while guaranteeing these challenges:
the waiting time of customers. To manage the supply and demand
- The optimization is actually a multi-objective
of the service in each geographical area, bipartite graph matching
is adopted to accomplish optimal resource allocation. It is further optimization, and thus we partition the optimization into
shown that the optimization of operation of autonomous vehicle several steps.
fleet can be successfully achieved, to outperform what human- - The collective efficiency involves a complex queuing
driving vehicles can possibly do. A new intelligent transportation network depending on complicated geographical
paradigm of great energy saving emerges. locations and routes, which is not possible to obtain
analytical results. We have to rely on approximation
I. INTRODUCTION methodology to reach numerical results.

S MART city has been one of the major application scenarios


for Internet of Things (IoT). Intelligent transportation
playing a key role in smart city involves serious concerns in
- To transform the real road topology into the queueing
network model, state-dependent service and interfering
queues must be taken into queueing network model,
travel time and in energy efficiency due to transportation in city which further complicate the optimization.
being a major energy consumption in human society [10]-[12]. - Since the empirical operational data for massive number
Recent success in autonomous vehicles (AVs), as a disruptive of AVs are not available to reach numerical results, we
technology elevation, creates tremendous potential to change have to leverage big data analytics from existing taxi
our daily life in the future [9]. However the application operation to proceed the initial investigation.
scenarios, required enabling technology, and issues associated
with practical engineering deployment, remain open by
incorporating AVs into intelligent transportation in smart cities.
In this paper we study the operation of massive AVs in a
smart city to reach the ultimate goal of time and energy
efficiency. Under appropriate management, the introduction of
AVs become more like a service (Car as a Service, CaaS) than
a product [13]. Due to disruptive views as: (i) Effective self-
driving of AVs being more effective than idle in a garage or
parking lot. (ii) The same level (or even better) of convenience
for sharing AVs as owning an AV for any individual.
Figure 1 illustrates the concept of such service similar to taxi
or Uber operation, but with further considerations. Benefitted
Fig. 1. Service for future AVs of a smart city. Once a user has a demand of
from the autonomous technology and centralized management, transportation from origin to destination, the management system of
global efficiency instead of localized issues can be achieved: autonomous vehicular transportation assigns a nearby available autonomous
Such efficiency includes mainly three aspects, that is, to vehicle taking this user to destination. The time required for the user to get a
vehicle is the waiting time. The time for the user to reach the destination is
guarantee the quality of service (QoS) and provide the best user the travel time. After reaching the destination, the vehicle can standby at the
experience, the waiting time for a user to get an AV, and the destination or move to certain area depends on the results of demand and
supply matching.

978-1-5090-5935-5/17/$31.00 ©2017 IEEE


As intelligent vehicular transportation relies on effective a server with a queue, and an intersection of roads is viewed as
routing, we luckily note the analogue between an AV trip and a a switch, while Figure 2 illustrates an example. When each road
taxi route that has been studied, such as the strategy of a taxi segment is represented by a server and a queue, there are two
driver to pick up passengers [3][4], to weight road segments to kinds of vital information to be taken cares, its length and
recommend the most economically efficient route [1], and number of lanes . The length together with velocity directly
customer hunting strategy [5]. However, the successful massive influences the distribution of service time, and the number of
operation of AVs as a means of intelligent transportation must lanes characterize the number of parallel servers on the road
consider the collective efficiency rather than the efficiency of segment. AVs are assigned to appropriate lane to optimize road
any single vehicle, to avoid the dilemma as tragedy of commons usage. Consequently, the road may be further approximated as
in the game theory. Consequently, we consider AV fleet a single server queue with the number of lanes influences the
management to identify the optimal operation to benefit the service rate of the server.
smart city from all aspects, beyond the scope of existing fleet An interaction functions as a switch in the
management [21][22]. telecommunication network, which routes the coming vehicles
The rest of this paper is organized as follows. In section II we to designated directions as outputs. For example of a crossroad,
will state our engineering problem and our network model. The each vehicle typically has four choices, going straight, turning
optimization procedure is then explained. In section III, we will right, turning left, and rarely happening U-turn. It is equivalent
analyze the cost of the queueing network and state the to a 4 by 4 switch [17] as Figure 3. Today’s traffic signaling of
difference between AVs and human-driving vehicles (HVs) interaction typically controls two directional flows at the same
from a macroscopic view. In section IV, we will show how to time to avoid dangerous situations. For future networked AVs,
balance the source and demand of autonomous CaaS via with the aid of cloud computing, edge computing, and on-board
prediction and bipartite graph matching. Section V presents our computing, the vehicles can behave more like packets in
optimization results numerically with comparison to the communication switch for better efficiency with safety.
efficiency of AVs without optimization and HVs. The
B. Optimization Problem
balancing of supply and demand and optimal number of
available vehicles are also shown. Finally, we will conclude our We first aim at achieving global optimality of average travel
work in section VI. time for AVs in a smart city. Given the road topology and
demand of source-destination, to reach the minimal average
II. ENGINEERING PROBLEM AND MODEL travel time in a collective manner, we have to find the
appropriate routes for AVs and to effectively manage the traffic
The ultimate goal of this paper is to identify the optimal
operation of AVs’ CaaS, which involves three steps:
1. Given current traffic flows in the city, find the optimal
operation that minimize the average travel time of all
vehicles in the city.
2. Under certain average waiting time constraint, find the
optimal number of available AVs in each area with the
traffic information from step 1.
3. To maintain the optimal number of available AVs in
each area, bipartite matching is applied to satisfy the
demand and the supply in each geographical area, by re-
assigning portion of available AVs to other areas.
Remark: Since re-assigning available AVs may generate extra
though minor traffic flow that not serving users but reducing Fig. 2. Corresponding network model of certain road topology. The orange
circle indicates the switches. The number of input and output of each switch
waiting time for users, such a procedure shall execute in a depends on the original road topology.
recursive manner to reach multi-objective optimization.
In this section, we will first explain our queueing network
model, and then demonstrate optimization under traffic flows,
while the cost function in optimization will be derived in next
section.
A. Network Model
To comprehend the moving behavior of AVs, we construct a
network model of road topology to analyze the traffic [20]. The
Fig. 3. A crossroad with four inputs , , , from different directions
network model generally has two kinds of components: server and four different outputs , , , to different directions can be viewed
with a queue, and switch. Each road segment is represented as as a 4x4 switch which routes the incoming vehicles from different directions
to the output that the vehicle is heading to.
and utilization of roads. Mathematically, we have the following time for a user to get an AV for his/her use. It is also good to
optimization: estimate the energy consumption from basic information of the
Optimization1: Given demand , the goal is to find the optimal route and traffic conditions. The analysis of cost relies on two
route flow that the cost function ( ) reaches its minimum, factors in the following, while we may tell the difference for
with the intuitive constraints AVs and HVs.
ℎ = , ∀( , ) ∈ ℳ A. Road Segments
∈ℛ (1)
The key to relate velocity, traffic intensity, vehicle density
ℎ ≥ 0, ∀ ∈ ℛ , ∀( , ) ∈ ℳ
and other basic parameters of traffic is safe distance. Under the
The meaning of the notations can be found in Table 1. Where scenario of future city with autonomous vehicle, the car should
the constraints suggest the overall assignment of all source-
automatically keep safe distance to the front car[18], this
destination pairs should sum up to the corresponding demand
distance will only consists of stop distance, no reaction time is
and each assignment should be non-negative. While the
optimization is straightforward, the numerical solution is considered, where the safe distance can be represented by
complex to obtain, particularly in near real-time in practical its speed and the maximum acceleration :
needs, since the possibilities of routing for each ( , ) pair (4)
=
grows exponentially. Therefore, we transform the route flow 2
optimization problem into the link flow optimization as follows: If we ignore the length of car, the traffic density can be
Optimization2: Given all ’s, the goal is to find the optimal written as:
that the cost function ( ) reaches its minimum, subject to 1 2
constraints = = (5)

( , ) − ( , ) = ,( , ) , On a road segment with length and number of lane ,


:( , )∈ℰ :( , )∈ℰ (2) assume the maximum average speed is . Before the traffic
∀ ∈ , ∀( , ) ∈ ℳ intensity reach the corresponding , the vehicle can keep its
( , ) ≥ 0, ∀( , ) ∈ ℰ, ∀( , ) ∈ ℳ average speed at . Once the traffic intensity is greater than
with , if = this , the vehicles should lower their average speed to
,( , ) = − , if = (3) decrease the safe distance, thus the road can accommodate more
0, otherwise vehicles. If we approximate each road as a single server queue,
The purpose of the constraint is to balance the flows in the the corresponding service rate can be written as:
intersections, i.e., flows for ( , ) demand that enters ≤
( )= (6)
intersection should equal to the one that exit . Though the 2 ℎ
feasible sets for the two problems are not equal, under non- The corresponding travel time on this road segment will be:
negative cost assumption, the results are of no difference [16].
Luckily, we can employ the Frank-Wolfe algorithm to ≤
accomplish the numerical results in the later part of this paper. ( )= (7)
III. ECONOMIC ANALYSIS ℎ
2
It is difficult to proceed the analysis of management and This is only for the case of AVs. A HV cannot always follow
economy in this research due to the fact so many aspects to the vehicle in front of it and keep the safe distance precisely.
consider the cost in our concerns. To reach a grand vision, we Thus, the road space cannot be utilized efficiently, also the
focus on quality of services and user experience in this paper. average speed will decrease due to the increase of traffic
That is, to provide services that users can access any time, any intensity. In the single-regime traffic model [14], the average
location, with the shortest possible waiting. Therefore, the velocity will always be lower than the maximum speed, relate
travel time is selected as the cost function, which includes the to the factor , which indicates the ratio of current traffic
TABLE I. Notations.
ℳ the set of all ( , ) pairs of demand intensity and the traffic intensity that the traffic jam occurred.
ℰ the set of all links If we adopt the relation given by Greenshields, the average
the demand from to , ∀( , ) ∈ ℳ velocity can be written as:
the collection of ′s
ℎ the demand from to through route , ∀( , ) ∈ ℳ, ∀ ∈ ℛ = 1− (8)
the collection of ℎ ′s
1, if ∈ ℛ passes link Thus it’ll have longer travel time ( ) compare to the
the indicator function which equals
0, otherwise one of AVs in (7). That is:
∑( , )∈ℳ ℎ =
the total flow from to that passes link , ∀( , ) ∈ ℳ, ∀ ∈ ℰ
∑( , )∈ℳ = the total flow that passes link , ∀ ∈ ℰ
the collection of ′s
( )= ∙ (9) =1+
− , (14)

B. Intersections where indicates the average waiting time due to the back-
,
As stated in earlier Section, the future intersections with AVs pressure from the th output queue, which means the average
can be viewed as switches. To analyze the delay caused by such waiting time due to the th output doesn’t have any buffer for
“vehicle switches”, we can view a vehicle switch of input queue the head vehicle that’s routing to the th output. We have:
and output queue. If all the input are destined to different
, ( )
output, then nothing will happen. If there are more than one , = (15)
“input” vehicles are routed to the same output, we face so-called
where , ( ) indicates the difference between average
contention as communication networks.
Thanks to the existing results of packet switching in number of waiting customers in the equivalent queue and the
computer networks or discrete-event systems, we can modify average occupied number of buffer.
the results in [15] to solve our problem. If we adopt Kleinrock The engineering meaning of service time is equal to the time
approximation, the inter-arrival time of the input of the switch for the vehicle to move to the output queue plus the sum of
can be approximated as independent exponential distribution waiting time due to the back-pressure. For the detail, please
under exponential distributed input traffic. [23] The input queue refer to [15].
can be modeled as an / /1 queueing system, where its Above result is for the delay in the intersection for AVs. For
serving pattern is that it can forward vehicles to the output if it HVs in the intersections, the concept of communication
has enough space, otherwise it’ll wait until the space is clear. switching is inappropriate and thus replies on traffic signal
The output queue here can be approximated as a / / queue system. Applying the model by Webster [16], the delay for
with slots of buffer since indicates the number of lanes in HVs in the intersection can be written as:
the output, which serves like a server system with slots of 1−
buffer. The service time can be viewed as the time required for = 0.9 + (16)
2 (1 − ) 2 1−
the vehicle to travel through the distance that equal to the safe
distance, thus there’re enough space for another vehicle to where indicates the ratio of current flow over the saturation
insert. Under the speed around 50km/h, the service time is flow, and indicate the time of green light and the entire
around 1s. We will take average service time as 1s in the cycle.
following analysis. Intuitively, current traffic signal system for HVs can not fully
Consider the × switch, indicates the arrival rate of utilize the road space due to lacking real-time information
the th input. The arrival rate of the th output queue can be networking and computing/decision, while switching AVs can
written as: always achieve the purpose by adjusting target AVs and
neighboring AVs.
= (10)
IV. RESOURCE MANAGEMENT VIA BIPARTITE GRAPH
The average delay that arrives at the th input and MATCHING
departs at the th output for a vehicle to experience can be To realize the scope of CaaS, guaranteeing the supply of AVs
written as: to meet the demand from users’ service requirements is a must.
= + (11) We therefore face a new application scenario of supply-
where indicates the delay in the th input queue and demand matching problem. To guarantee the QoS, we should
indicates the delay in the th output. keep the number of supply at the optimal number of available
From the formula of / / queue, can be written as: AVs in any designated area. Three primary factors in our
queuing model are identified to match supply-demand
=1+ (12) relationship, a kind of resource allocation or management: the
( − 1)! − departure rate indicating the rate that users obtain requested
where is the steady-state probability that there’s no vehicle AVs, the arrival rate indicating the rate that users reach
in the output queue. destinations and thus AVs are available again, these two rates
Modify the results from [15], and follow the results for influence the supply of available vehicles. The rate that new
/ /1 queue, can be written as: requests generated directly indicates the amount of demand.
For an equivalent queueing system, the rate that new requests
= (13) generated is the arrival rate of customers, while the number of
2(1 − ) available vehicles indicates the number of servers of this queue
where is the equivalent service time of the th input / /1 and influences the rate of service. The average service time can
queue, and is the second moment of it. We have: be obtained numerically based on the traffic flow and the
corresponding travel time in previous step. Thus, optimal
number of available vehicles can be obtained with certain
waiting time constraint.
To achieve above purpose, we have to develop two
technologies: forecasting the demand and supply such that near
real-time management is possible, socially enabled resource
balancing to match the supply-demand [19].
A. Forecasting via Big Data Analytics
In a fully autonomous transportation system, it still takes
time to drive an AV to a designated user, and such waiting time
is heavily cared by any user of our proposed intelligent
transportation system. Therefore, we must establish a
forecasting mechanism for the demand with respect to
departure and destination statistics, to accomplish the desirable
intelligent AV system. Since such forecast shall be near real-
time and update online, the regression with learning capability Fig. 4. STL Decomposition
in the form of time series appears attractive to implement. of available vehicles moving in/out the area with its sign
As explained earlier, it is impossible to obtain the operational indicating moving in or out. To keep the supply and demand
data of massive AVs at this time. However, the big data from balanced, we have:
large-scale taxi service or Uber appears to meet out initial need. − − + =0 (19)
We are lucky to obtain recent big data of Beijing taxi data to Now, we face a typical resource matching problem with an
develop the forecasting mechanism. Such taxi data consists of extra objective to minimize optional cost, and [19] suggests a
12,000 taxis operating over a month with 24-hour GPS trace for method of multi-objective bipartite graph matching to assign
each taxi. After big efforts in data cleaning, we examine the AVs to appropriate locations with minimal operational cost.
statistics of such taxi data. Without surprising, our statistical Figure 5 delineates this scenario, where represents the th
analysis as Figure 4 shows obvious periodic behavior in days destination and there’re total destinations. And represents
and weeks. Consequently, the forecasting (of time series) the th starting point, after adding the change of number of
should include so-called seasonal decomposition method to
optimal available vehicles, there’re total ′ starting points.
extract the periodic behavior. The consequent demand can be
means the cost moving from the th destination to the th
represented as:
starting point. We reach the following optimization:
= + + (17)
Optimization3: Finding pairs ( , ), where 1 ≤ ≤ , 1 ≤
where representing the regression (trend), representing
the seasonal effect, and representing the random nature of ≤ , = ( , ) that minimize:
user behavior. To estimate, the STL (Seasonal and Trend
(20)
decomposition using Loess) is adopted [8] to obtain Figure 4.
Once obtaining the estimate of , the subsequent
Subject to:
forecasting turns out to be forecasting = − , the
common auto-regressive-moving-average (ARMA) in time = =1∀ , (21)
series analysis could satisfy our needs [24]:
1, ℎ
+ = (18) where =
0, ℎ
The solution of this optimization can be obtained by
where is the linear combination of its past values and the employing the Hungarian Algorithm proposed by Khun and
random noises, with ~ ( , ). The orders ( , ) can be Munkres [6], with polynomial complexity. In other words,
determined by applying the information criterion. In the work, based on this optimization, the intelligent transportation system
we use Akike Information criterion (AIC)[25], for it can find
the optimal approximation under the formulation of (18).
Combining STL and ARMA, we can achieve the forecast on the
demands in terms of geographical regions and times.
B. Multi-Objective Bipartite Graph Matching
In geographical area , we denote AVs taking new
passengers, AVs reaching destinations, and the optimal
number of available vehicles changes which is due to the
change of demand. Another parameter indicates the number Fig. 5. Matching Problem
for AVs determines the assignment and consequent routing of time reactions and seamless traversal through intersections,
each available AV to a specific area to keep smooth operation resulting in shorter travelling time with no doubt. Compared to
of the system. the centralized AV system, lacking of information about other
AVs can only route the shortest path in best efforts under a
V. SIMULATION OF OPERATION FOR AV TRANSPORTATION distributed manner, to result in imbalanced utilization of roads
SYSTEM and thus longer travel time. The huge gain is impressive!
In this section, leveraging taxi big data and road topology, B. Optimal Number of Vehicles & Bipartite Graph Matching
we demonstrate the optimal operation of AV transportation Figure 8 depicts the optimal number of available AVs in each
system. We first look into the average travel time to show the area. To offer appropriate number of AVs in each area, effective
difference between HVs and AVs, and to demonstrate the moving available AVs among areas is required. Such matching
optimization to make AVs more efficient. Then, the optimal of supply-demand to guarantee the waiting for users consumes
number of available AVs and the cost to match the demand and energy. Figure 9 shows the average gasoline required for a
supply is shown. single AV, estimated from the average speed and distance to
move. We can see that the average cost required for moving
A. Average Travel Time during the daytime is much smaller comparing to the one during
Here we analyze the average travel time under three different the middle of the night. This is due to the sparsity of both the
situations: HVs in a smart city, AVs in a smart city with/without demand and the supply, thus it requires much more cost to make
optimization. Without surprising, HVs suggest the longest the demand and supply meet in average. However, please note
travel time. Figure 6 shows the average cost (average travel that the number of required AVs in late night is far less but the
time required to pass the road segment) of each road segment total energy to guarantee the waiting time is still much less than
for three different situations and Figure 7 compare the average day time.
cost of each O-D pair under 3 different situations. Under the By comparing with the data of real taxi, our simulation result
same routing strategy, the AVs take advantage of almost real- shows that the average energy that required to balance the

Fig. 6. Average travel time between each area under different scenarios: (a) City with HVs. (b) City with AVs (c) City with AVs after optimization. Links between
areas are bidirectional. The color indicates the average time (in hour) for a vehicle to go from one area to another. Vehicles in (a) and (b) decide its routes with
certain strategy that the vehicles starting from the same area can reach the minimal average cost. It can be viewed as kind of local optimization. While vehicles in
(c) decide its routes by the results of global optimization. We can see under the same strategy, AVs have lower cost compare to HVs. Under the case without
global optimization, the traffic will gather at some road segments while the traffic after global optimization disperse to the routes that’s originally low usage.
Rate of Average Travel Time (%)

100

80

60

40

20

0
HVs AVs AVs (opt.)
Fig. 7. Average travel time for each O-D pair under different scenarios. The
average travel time for HVs is about 0.26hr. AVs without optimization is
about 76% of the time spent for HVs, AVs after optimization spent only about Fig. 8. The optimal number of available AVs in different areas given a selected
59% of the time of the case for HVs. time.
[13] Lucas Mearian, Get ready for car-as-a-service (CaaS), Computerworld
Report, Available: http://www.computerworld.com/article/3004807/
emerging-technology/get-ready-for-car-as-a-service-caas.html
[14] Fred. L. Hall, Traffic Stream Characteristics. 1996
[15] Ilias Iliadis, Wolfgang E. Denzel, “Analysis of Packet Switches with Input
and Output Queueing,“ IEEE Transactions on Communications, vol. 41,
no. 5, pp. 731-740, 1993.
[16] M. Patriksson, The Traffic Assignment Problem – Models and Methods.
Linköping Institute of Technology, Linköping, Sweden, Oct., 1994.
[17] Joseph Y. Hui, Switching and Traffic Theory for Integrated Broadband
Networks. Springer US, 1990.
[18] Albert Rizaldi, Fabian Immler, and Matthias Althoff “A Formally
Verified Checker of the Safe Distance Traffic Rules for Autonomous
Vehicles.” 2016.
Fig. 9. Gasoline Consumption by Moving Vehicles [19] L. Wang, H. Wu, W. Wang, K.C. Chen., “Socially Enabled Wireless
Networks: Resource Allocation via Bipartite Graph Matching,” IEEE
supply and demand is only about 4.3% of the average energy Communication Magazine, vol. 53, no. 10, pp. 128-135, Oct. 2015.
required for close-by taxi to a passenger. If adding the energy [20] Tom Van Woensel and Nico Vandaele, “Modelling Traffic Flows with
consumed for the AVs to reach the user that sent a request, the Queueing Models: A Review,” Asia-Pacific Journal of Operational
average energy required for AVs to reach next user is only Research 2007 24:04, pp.435-461
[21] W. B. Powell and T. A. Carvalho, “Dynamic control of logistics queueing
around 15% of current taxi service. Such a dramatic energy networks for large-scale fleet management,” Transp. Sci., vol. 32, no. 2,
efficiency gain comes from balancing supply-demand and no pp. 90–109, May 1998
need for searching passengers. [22] K. T. Seow, N. H. Dang and D. H. Lee, "A Collaborative Multiagent Taxi-
Dispatch System," in IEEE Transactions on Automation Science and
Engineering, vol. 7, no. 3, pp. 607-616, July 2010.
VI. CONCLUSION [23] Kleinrock, L., Communication Nets: Stochastic Message Flow and Delay.
In this paper, we present a new technology paradigm by New York: McGraw-Hill, 1964.
[24] Whittle, P., “Hypothesis testing in times series analysis,” Uppsala:
treating AVs as a service of intelligent transportation in a smart Almqvist & Wiksells Boktryckeri AB.
city. Via our queuing network model and leveraging big data [25] Akaike, H., "Information theory and an extension of the maximum
analytics on existing taxi fleet, we successfully demonstrate the likelihood principle," 2nd International Symposium on Information
efficiency of massive AV operation to greatly satisfy users and Theory, Tsahkadsor, Armenia, USSR, September 2-8, 1971, Budapest:
Akadémiai Kiadó, pp. 267–281.
to provide amazing energy saving.

REFERENCES
[1] H. Dong, X. Zhang, Y. Dong, C. Chen and F. Rao, "Recommend a
Profitable Cruising Route for Taxi Drivers," IEEE 17th International
Conference on Intelligent Transportation Systems (ITSC), pp. 2003-2008,
2014.
[2] G. Clare and A. Richards, "Optimization of taxiway routing and runway
scheduling," IEEE Transactions on Intelligent Transportation Systems,
vol. 12, no. 4, pp. 1000-1013, 2011.
[3] Nicholas Jing Yuan, Yu Zheng, Liuhang Zhang, and Xing Xie. “Tfinder:
A recommender system for finding passengers and vacant taxis,” IEEE
Transactions on Knowledge & Data Engineering, 25(10):2390–2403,
2013.
[4] Yong Ge, Hui Xiong, Alexander Tuzhilin, Keli Xiao, Marco Gruteser, and
Michael Pazzani, “An energy-efficient mobile recommender system,” In
Proceedings of the 16th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 899–908. ACM, 2010.
[5] B. Li, D. Zhang, L. Sun, C. Chen, S. Li, G. Qi, and Q. Yang, “Hunting or
waiting? discovering passenger-finding strategies from a large-scale real-
world taxi dataset,” In PERCOM Workshops, 2011 IEEE Int. Conf. on,
pp. 63 –68, Mar. 2011.
[6] Harold W. Kuhn, "The Hungarian Method for the assignment problem,"
Naval Research Logistics Quarterly, 2: 83–97, 1955
[7] Z. Deng, M. Ji, “Spatiotemporal Structure of Taxi Services in Shanghai:
Using Exploratory Spatial Data Analysis,” In IEEE Geoinform 2011 19th
Int. Conf. on, pp. 1-5, Jun. 2011.
[8] R. B. Cleveland et al., ”STL: A Seasonal-Trend Decomposition Procedure
Based on Loess,” Journal of Official Statistics, vol. 6, No. 1, pp.3-73,
1990.
[9] IEEE News Releases, Sept. 2012, Available: http://www.ieee.org/about/
news/2012/5september_2_2012.html
[10] Smart Cities project, Available: http://www.smartcities.info/
[11] IEEE Smart Cities, Available: http://smartcities.ieee.org/
[12] Building a Smart + Equitable City, Available: http://www1.nyc.gov/site/
forward/innovations/smartnyc.page

You might also like