Microelectronics

1
Crash Scene Re-Generation Using Structural
Hidden Markov Models
Professor Djamel Bouchaffra
(Advisor)
Raef Aidibi (Ph.D. Student)
Computer Science Engineering 131 Dodge
Hall
Phone [phone], email bouchaff_at_oakland.edu
4 Our Approach Implement a learning system
possessing the capability to search for
structural information of the sequence of events
preceding the collision i.e., search for the
structural information representing a component
behavior.
The National Transportation Safety Board has
recommended that Automobile Manufacturers and the
National Highway Traffic Safety Administration
work cooperatively to gather information on
automotive crashes, using collision sensing and
recording devices. National Transportation
Safety Board Symposium On Recorders 1 - Problem
Statement Current crash scene re-generating
methods lack the capabilities of searching for
component structural information of the sequence
of events preceding the collision, where design
problems are not revealed efficiently early
enough.

• 5 The Structured Hidden Markov Models Approach
• The sequence that describes the entire pattern
  (S) is explained by a single Hidden Markov Model
  (Oi) which represents the Component Status
  Sequence.
• This hidden Markov model is extended to contain
  structural information that are embedded within
  the pattern.
• The contribution of each component to the entire
  pattern is evaluated. These components are merged
  together to describe the structure of this
  pattern.

e.g. dynamic stability, injectors status, yaw
angle, tire torque
Component Structure Status
Natural Sensor Network
e.g. vehicle speed, engine rpm, brake pressure,
injectors durations. status, torque,
Center of mass ends up outside of the vehicle
area of contact
Design Defects Scene (Roll Over)
State of the Vehicle
e.g. accelerating, braking, decelerating,
turning
Sequence describing the entire pattern
Component Observation Sequence
Structured HMM containing structural Information
State Sequence representing State of the Vehicle,
Monitored by Sensor Network
Usually occur at intersections
Side Impact Crash Scene (Road Design Defect)

• 5 Re-Generating the Crash Sequence using SHMM
• Probability Evaluation
• Predicting the scene represented by the
  classified sequence, captured by the sensor
  network
• The evaluation problem in SHMM consists of
  evaluating the probability for the model
• ? (pA B SD), to produce the sequence O. is
  expressed as

State Decoding The state decoding problem
consists of determining the optimal state
sequence Find the best structural sequence that
led into this scene class, that best esplains the
sequence
Current Methodologies in Generating/Analyzing
crash scenes (Puzzle Construction)
They think I m a learning Machine
Structure Decoding The structural decoding
problem consists of determining the optimal
structure sequence
Parameter Re-estimation In the structural HMM
paradigm, a forward-backward maximization
algorithm to re-estimate the parameters contained
in the model is utilized.

• Who Would benefit from the Learning Machine?
• Design Engineers
• Police Officers
• City Planners
• Insurance Companies
• Public

• 6 References
• D. Bouchaffra and J. Tan, "Structural Hidden
  Markov Models using a Relation of Equivalence
  Application to Automotive Designs",  in Data
  Mining and Knowledge Discovery Journal,
  Springer-V, 2005
• D. Bouchaffra and J. Tan, "Introduction to
  Structural Hidden Markov Models Application to
  Handwritten Numeral Recognition",  Intelligent
  Data Analysis Journal, IDA, Editor-in-Chief A.
  Famili, Vol., 101, IOS Press, 2006
• D. Bouchaffra and J. Tan, "Structural Hidden
  Markov Model and Its Application in Automotive
  Industry", Enterprise Information Systems V,
  Camp, O. Filipe, J.B. Hammoudi, S. Piattini,
  M.G. (Eds.), XIV, 332 p., Hardcover, ISBN
  1-4020-1726-X, Published by Springer, 2004
• Djamel Bouchaffra and Jun Tan, "Introduction to
  Structural HMM and it's Application in Pattern
  Classification", ANNIE'2004, Smart Engineering
  System Design-Neural Network, Fuzzy Logic,
  Evolutionary Programming, Complex Systems and
  Artificial Life, Nov. 7-10, 2004, University of
  Missouri-Rolla.
• Djamel Bouchaffra and Jun Tan, "Introduction to
  the Concept of Structural HMM Application to
  Mining Customers' Preferences for Automotive
  Designs", The 17th International Conference on
  Pattern Recognition (ICPR) Cambridge, United
  Kingdom, 23-26 August, 2004 (Proceedings
  Published by IEEE Computer Society).
• Raef Aidibi, "Introduction to Hidden Markov
  Models Decision Processes (HMMDP)", International
  Computer System and Information Technology,
  ICSIT'05 (IEEE/ CDTA) July 19-22, 2005 Algiers

2 - Drawbacks 1 Heavily based on assumptions 2
Inaccurate time-consuming 3 Might lead to
wrong conclusions 4 Does not reveal design
defects 5 Diverts focus from core problem