This document describes a study that used simulation to evaluate different parking guidance policies for a connected vehicle intelligent parking system. The simulator modeled a parking lot with sensor-equipped and non-sensor vehicles. Four policies were tested: random assignment, nearest parking, maximum satisfaction guidance, and near-optimal guidance. The near-optimal policy aimed to maximize the quality of parking space information gathered by routing sensor vehicles to the most informative spaces. Simulation results found the near-optimal policy produced the most stable and accurate estimates of parking space occupancy over time, even with low numbers of sensor vehicles, making it the best policy for initial deployment of connected vehicle parking systems.

1. Background
Objectives
Methods
Results
Conclusion
References
Funding Source:
Mobility Transformation Center,
University of Michigan
Parking Policies Tested
Connected mobility - a recent development in
transportation is possible as connected vehicles with
sensors like short range radars (probe cars) are able to
perceive the environment and broadcast the information
to enable a system of information sharing. An intelligent
parking system enables collecting parking occupancy
information with minimum infrastructure upgrades and
thus facilitate with optimizing the parking guidance. In
our study, we address the challenge of collecting real-
time parking space information by probe cars in a smart
and efficient way in order to make parking search easier
and cost-effective for a driver.
Parking Simulator for Parking Policy Evaluation:
• Four realistic scenario-based parking policies were
tested on a developed simulator to arrive at the most
efficient parking guidance policy. The simulation of the
intelligent parking system was also visualized.
• The proposed near-optimal guidance policy was found
to be most stable and least sensitive to traffic counts
among all policies. The near-optimal policy is also
suitable for low penetration rates of probe cars.
[1] Luo, Q., R. Saigal and R. Hampshire, Searching for
Parking Spaces via Range-based Sensors.
Preprint, 2016. arXiv:1607.06708 [cs.RO]
[2] Singh, A., Krause, A., Guestrin, C., Kaiser, W.J. and
Batalin, Maxim.A., Efficient Planning of Informative Paths
for Multiple Robots. In IJCAI, vol. 7, pp. 2204-2211. 2007.
An Evaluation of Information Sharing Parking Guidance Policies
Using a Bayesian Approach
• Propose physical-level characteristics for an intelligent
parking system – parking lot infrastructure, sensor
equipped (probe) and non-sensor cars and interaction
behavior.
• Use a simulation-based Monte-Carlo approach to test
realistic scenario-based parking policies.
• To arrive at a near-optimal parking policy that
maximizes quality of parking space information
gathered and shared by sensor-equipped probe cars.
a) Event module for Car Arrivals - Cars arrive as
Mt/M/N/C queue system. N is the number of parking
spaces, C is the queue capacity of cars waiting for the
next available parking space.
b) Routing Module - Cars arrive on the shortest path. On
the way out, cars are routed in the same route as they
arrive (2-way routing) or on a different route (1-way
routing).
c) Scanning Module - Applies Recursive Bayesian
Updating to compute the posterior parking space
occupancy probabilities based on prior occupancy
probabilities to get the system state. Further the system
state is discounted with a factor β to account for
diminishing information at each time step.
Bayesian Updating Equation,
P Xi = P Xi Xi = 0 P Xi = 0 + P Xi Xi ≠ 0 P Xi ≠ 0
Routing module
Parking space and
route assignment
Event Module
Car arrival generation
Scanning module
Generate system state based
on occupancy probabilities
Figure 3. Simulator Visualization: the real parking lot is replicated
In the visualization. The parking lot layout is visualized. Red blips
represent probe cars and blue blips represent non-probe cars.
Policy 1: Random assignment: both types of cars are
assigned to available parking spaces randomly to represent
the average performance of the system.
Policy 2: Nearest parking: both types of cars park to the
available parking space closest to the entrance to simulate a
destination-oriented parking policy assuming the entrance is
the final destination of all the drivers.
Policy 3: Maximum satisfaction guidance: normal cars park
to the space closest to the entrance while probe cars park to
the space estimated to be most likely empty (Maximum
exploitation policy).
Policy 4: Near-optimal guidance: normal cars park to the
space closest to the entrance while probe cars park to the
space which will maximize information gain from scanning
(Maximum exploration policy).
The information gain function to be maximized is:
MI(χN(t),a(t)) = H(χN(t))−H(χN(t)|χa(t))
Figure 2. Simulator Modules
Figure 5. Effect of probe car penetration (Two-way)
Policy Evaluation Criteria:
Calculated as Relative Error of Occupancy Estimation
(REOE):
e t =
wrongly estimated parking space
total parking spaces
Figure 4. Two-way parking simulation results
Observations:
• Relative error of occupancy estimation is high initially
across all policies, but the error decreases as more
information is gained on the state of parking lot by
probe cars. Beyond this oscillation of relative error
occurs with variation in information.
• As the penetration level of probe cars increases,
relative error decreases across policies as more cars
scan for information.
• Information oscillation in 1-way case is less than
compared to the 2-way case. This is because 1-way
path allows probe cars to scan parking spaces in a
wider area.
• Near-optimal guidance policy has stable estimation
errors over time compared to the fluctuating errors in
other policies.
• Near optimal guidance policy is less sensitive to
number of vehicles parked in the system. For other
policies, relative errors increases with a drop in the
number of parked vehicles.
• Near-optimal guidance policy works well even with
low-penetration rates of probe cars, making it suitable
in initial phases of connected mobility implementation.
Simulation Parameters:
• Car Arrival process (non-homogenous Poisson) with
intensity λ(t) is a piece-wise function of time of day.
• Parking time modeled as exponential process with mean
𝛍 = 𝟔𝟎 seconds.
• Initial estimation of parking spaces 𝐩 𝛘 𝟐 = 𝟎. 𝟓
Discount factor of estimation 𝛃 = 𝟎. 𝟓
Scanning range of probe cars = 6 parking spaces
Proportion of probe cars, 𝛄 = 𝟎. 𝟏 𝐭𝐨 𝟎. 𝟗
• Average performance of each policy is evaluated by
applying a Monte Carlo method to generate event list
repeatedly.
• Simulation is repeated for 1000 times for each 𝛄 .
Discussion
Figure 1. Project Overview