Emulating Human Recognition of Driving Context

1
Emulating Human Recognition ofDriving Context

• Dr. Kevin R. Dixon
• Sandia National Laboratories
• [phone]
• krdixon_at_sandia.gov

2
Automotive Cognition Project

• Humans multitask while performing many jobs
• Increasing workload may result in performance
  decline on all tasks
• Talking on phone while driving increases accident
  risk similar to intoxicated driving
• Present information and tasks in a manner
  appropriate for operators current capacity

3
Human Recognition ofVehicle-Driving Situations

• Create a cognitive model of a human back-seat
  driver
• Identify potentially dangerous driving
  situations
• Approaching Slow-Moving Vehicle
• Entering On-Ramp
• Preparing to Change Lanes
• ...
• Use this information to minimize impact of
  untimely distractions during high-difficulty
  situations
• Delay unimportant messages
• Present material visually or auditorily

4
Human Recognition ofVehicle-Driving Situations

• Collected 24 hours of real-world driving data
• Vehicle-based sensors
• Driver-posture sensors
• Asked humans to label video according to their
  perception of current situations

5
Emulating Human Recognition of Driving Situations

• Trained cognitive model to emulate and predict
  the human recognition of driving situations
• The goal is to automatically derive a cognitive
  model that explicitly identifies implicit
  relationships that humans used when labeling the
  videos

6
Supervised Learning

• Sandia Cognitive Frameworks pattern recognizer
  is a type of nonlinear dynamical system
• Tune system parameters to minimize error between
  the humans labels and the estimated labels from
  the cognitive model

7
Optimizing the Model

• The vast majority of the time (75), nothing
  interesting happens, according to human labelers
• theyre watching German drivers, not Italian
  drivers
• This low ratio of interesting to boring
  situations does not force the system to learn
  what causes those rare, but important, events

8
Optimizing the Model

• Penalize the system for incorrectly predicting
  events inversely proportional to how frequently
  the event occurs
• Forces the system to pay attention to the
  interesting situations as much as the boring ones
• Iteratively tune the parameters of the
  approximated model to minimize the weighted
  percentage wrong
• Typically converges in less than one minute

9
Learning in Action
each row is an input
each column is output
10
Learning in Action

• Algorithm changes the weights to minimize the
  error between its labels and the humans labels

11
Human Recognition ofVehicle-Driving Situations

• Results for predicting human recognition of
  driving situations

12
Human Recognition ofVehicle-Driving Situations
13
Automated Context Extraction

• How do we know that we identified appropriate
  contexts?
• Better to let data tell us what contexts are
  important

14
Automated Context Extraction

• Identify for statistically meaningful
  regularities in the data
• These clusters may or may not correspond to an
  obvious real-world analogue
• Meaningless An artifact from sensor
  instrumentation
• Meaningful Underlying physical process induces
  clusters
• We transform temporal signals into vectors using
  regression over a pre-defined time window

15
Unsupervised Learning
16
Adapting Models to New Users

• Suppose we obtain a data set from a group as

• But we want to adapt the model to a new user

17
Adapting Models to New Users
18
Difficulty Scoring

• Instead of identifying specific contexts, human
  merely indicates perceived difficulty using an
  analog device, where zero is easy, and 100 is
  sudden death
• Train cognitive model to emulate and predict
  human perception of driving difficulty

19
Difficulty Scoring

• Test-Set Results
• Average error (L1) 4.99
• Mitigation Correct 93.17

20
Whats Next?