Paper Summary By Jeff Ploetner

1
Paper SummaryBy Jeff Ploetner

• Vehicle Reidentification Using Multidetector
  FusionSun, C.C. Arr, G.S. Ramachandran, R.P.
  Ritchie, S.G.Intelligent Transportation
  Systems, IEEE Transactions onVolume 5,  Issue
  3,  Sept. 2004 Page(s)155 - 164

2
Problem

• Vehicle Reidentification
• How to identify a vehicle at one location and
  reidentify it at a future location(s)
• Useful for Traffic analysis
• Tracking
• Travel Times, as opposed to point speeds
• Origin/Destination estimation

3
Potential Solutions

• License plate matching
• In vehicle solutions
• GPS, toll tags, tracking beacons
• Other sensors
• Laser profiles
• Weigh-in-motion axle profiles
• Ultrasonic sensors
• Magnetometers

4
Test Setup

• Sensors used
• Video
• Inductive loops

5
Assumptions and Limitations

• Very simplistic setup
• Detector stations are only 130 m apart!
• Assume vehicles stay in order
• No passing
• No changing lanesmakes reidentification trivial?
• Assume that vehicles stay in a group (platoon),
  and there is no jumping from one group to the
  next
• Video taken during just 1.5 hours in the morning
• No lighting transitions, not much variation
• 581 vehicles, 200 for training set, rest for test
  set
• Poor feature selection
• Well see why

6
Features

• They use 6 features
• Magnetic Vehicle Signature
• Vehicle Speed
• Platoon Traversal Time
• Maximum magnetic amplitude
• Magnetic length
• Color

7
Feature Analysis

• Magnetic Vehicle Signature
• Normalized magnitude as well as length
• Vehicle speed
• Probably not good to use
• Platoon Traversal Time
• Time from first to last car in the platoon of X
  number of vehicles
• Given downstream platoon, match to possible
  upstream platoons
• Time bound it for reasonable travel times (10-68
  mph)

8
Feature Analysis

• Color
• Quantize/Group RGB values into 125 subsets
• Have a 125 dimension feature vectorconsisting of
  the number foregroundpixels in each cube

9
Data Fusion
Distance for the feature
Total Distance

• Linear Opinion Pool
• Nearest Neighbor Approach, using a Weighted
  Distance Measure
• Weights selected by finding optimum on training
  data
• Easily allows for adding more features
• Many other methods available
• Majority vote, etc

Feature Weight
10
Results
11
Sensitivities
12
Conclusions

• Feature selection is critical
• Think things through!
• Have reasons for picking features
• Use uncorrelated features (and sensors) for
  maximum robustness
• Try to make realistic assumptions and get
  realistic test data
• Easy to get good results on small data sets
• Hard to get robustness under all conditions