Transcript Park sense

Group #6
Evan Davidson
Afsaan Kermani
Viker Lamardo
Scott Moriarty
What is Park Sense?
Park sense is a system which maps
out a parking lot and detects and
displays which parking spaces are
vacant and which are occupied.
Motivation
• We wanted to solve an everyday problem that is
relevant to UCF students
• Can take up to 30 minutes to find parking during
peak traffic hours
• 55,000 students at UCF and only 15,500 parking
spaces available, not all of which are available to
students
• Help students and faculty get to class on time
Objectives
• Accurate car detection
• Detect the car during all times of the day
• Capable of accommodating a network of
cameras so that the project is scalable and can be
applied to most of the parking lots in the real world
Requirements
• Cost effective and easy to install
• Weather resistant
• Fixed position
• Wide range of view
• Scheduled day/night mode
Specifications
• Range: minimum of 2 spaces
• Cost: under $10 per parking space
• Power source: standard 120VAC (wall outlet) or
12V battery powered
• Accuracy: 95% and a system refresh time of less
than 20 seconds
• Operating environments Temperature: 0 to 115 °F
Humidity: 80% RH
Features
• LCD display that greets the user as they drive in
to the parking lot
• Website showing live updates on the status of
the parking spaces in the parking lot
• It doubles as a security system because we are
using live video feeds
• Parking stats
Research
Ultrasonic Sensors
Advantage
Disadvantage
• Wide range allowed us to
possibly use one sensor per two
parking spots
• Would require a large number of
sensors to cover a full parking lot
Emitter/Receiver
1.64’’
Ultrasonic
Infrared Sensors
Advantage
• Less
expensive
ultrasonic sensor
Disadvantage
than
the
• Range was not as wide as the
ultrasonic sensor
Infrared
Camera
•We decided to use the Vivotek IP7330 for the Park Sense System
•Description: bullet-style network camera designed for outdoor applications
•Advantages:
•shields from harsh conditions such as rain and dust
•supports tamper detection (i.e. blockage, redirection, & spray-painting)
•Functions as a security camera for students in the parking lot
•Has both a long and wide range
•Has both day and night vision capabilities
•Easy connection router and server to perform image processing
•Supports PoE (power over Ethernet)
5.91’’
2.34’’
System Block Diagram
Camera
Captures an Image
Image processing
Data analysis
Server
Generates statistics
Outputs data
Display
Image capture
• The network camera captures an jpg image from
the mjpg stream
Image processing
• The captured image is then processed through
the code to accurately object detect cars in the
parking lot
Data output
• The program then outputs to a .csv file with
information based on the processed image.
Dynamic website
• The website is able to update the number of
rows, columns, and statistical values by reading
the .csv file that our initial analysis produces.
Display
Because we are already generating a webpage, all
we have to do is keep that site displayed on an
LCD screen mounted at the entrance to the
parking lot
System refresh & update display
• The website is then updated with the proper
graphical view of the vacant parking lots.
Networking
• Originally we planned on using multiple sensors
in a mesh network design, but due to budget
constraints and design modifications we
decided to use a basic Ethernet network to
allow for scalability in practical applications
• The camera operates on a local area network
(LAN) with the computer and display.
OpenCV
Haar-Like Features

The value of a Haar-like feature is the difference
between the sum of the pixel gray values within
the black and white rectangular regions

The image is then scanned by a sub window
containing a Haar-like feature trying to detect on
the classifiers it was trained on
Cascade of Classifiers
•
AdaBoost uses a “weak” learning
algorithm and a training set to create
strong classifiers
– Adaptive because later classifiers are geared to
be in favor of sub-windows misclassified by
previous classifiers
•
Series of classifiers are applied to every
sub-window, the first eliminates a large
number of negative sub-windows and
passes all positives (lots of false positives)
– Followed by a set amount of stages which do
the same to reduce the number of false
positives in the final product
Data Preparation
Collect positive images.
 Obtain negative sample set.
 Create info files.
 Generate vec files.
 Train cascade.

Object Marking

For the positive sample set, object
marking was required. The positive info
file looked like the following:
Testing a Cascade
Software
C++
 OpenCV Library
 Single Run, repeated using a batch file

 Can be continual
Low CPU usage
 Scalable

Software Flow
Capture image
2. Gray scale
3. Object detection
4. Region of interest rectangle comparing
5. Output to .csv file
6. Sleep for set time interval
repeat
1.
Determining a Hit
•
•
•
•
For each rectangular object detected, the X,Y
coordinates and the height and width are
stored in a cvRect() object within another
object which contains the total number of
objects found, CvSeq
These are passed to the function RectDetect()
which also takes in the same set of values for
a single ROI
The boundaries of each are compared to
determine there’s an overlap
Once a hit is announced for a given ROI no
other detected objects will be compared to that
ROI to save on processing time
Testing
Testing
• Haar training:
• Created a variety of cascades including cars,
headlights and windshields. Also tested
rectangles to possibly detect empty spots.
• Testing the cascades
• Analyzing hits, misses, and false hits
•Website:
• Loading the .csv properly
• Aligning .csv cells with parking spots
The Test Setup
• Determine best test subjects among
model cars or pictures of cars
• Determine the best camera
positioning for accuracy.
• Control the lighting
• Testing IR
• Calibrate the regions of interest
Testing Results
Trying to detect full cars and headlights,
resulted in ~ 0%
 We tried to detect windshields on a
cascade using ~500 positives and ~500
negatives.

 Results varied between 40-80%
 Shadows generate false positives.
 Camera distance from the cars affected
accuracy
 Nuanced lighting led to diminished accuracy
 IR detection worked toward center of image.
Testing Results Continued
The windshields likely succeeded
because it resembles a Haar-like
feature. Line features and center
surround features.
 The windshield is somewhat consistent
between all cars.
 Ideally, multiple cascades would be used
to detect all cars.

Problems
• We are all hardware oriented in a mostly
software project
• Lighting conditions in the parking lot were a
problem as stated.
• Cars that are the same color as the asphalt.
• Working with open source library was difficult.
• Recommended cascades call for between 2,000
and 4,000 positives and 4,000 to 10,000
negatives.
• Training of such a robust cascade would
require at least a week of processing on a
home computer
Budget
Part
Cost
Vivotek IP 7330 Camera
$270
Server
$0 – already owned
LCD Display
$0 – already owned
Router
$0 – already owned
Miscellaneous
$50
Total
$320
Milestones
Demo
and
Q&A