Transcript Detecting Driver Distraction Using a Data Mining Approach

The Detection of Driver Cognitive Distraction
Using Data Mining Methods
Presenter: Yulan Liang
Department of Mechanical and Industrial Engineering
The University of Iowa
1
Driver distraction
• Driver distraction and inattention
has become a leading cause of
motor-vehicle crashes
o Nearly 80% of crashes and 65% of
near-crashes (the 100-car study)
o Increasing use of In-Vehicle
Information Systems (IVISs), such as,
navigation systems, MP3 players, and
internet services.
• Driver distraction represent a big
challenge for developing IVISs
o Benefits of the IVIS functions
o Safety
o One solution: driver distraction
mitigation systems
People use In-Vehicle Information Systems (IVISs)
during driving
2
Driver distraction mitigation systems
• Distraction detection is a crucial function
o Cognitive distraction
o Visual/manual distraction
o Simultaneous(dual) distraction
Sensor
Techology
Model-based Driver
Distraction Detection
Indicators of distraction
Detection techniques
Mitigation
System
Driver state----------------· Physiological responses
· eye glances
· fixations, saccades, and
smooth pursuits
...
Visual/Manual
distraction
Strategy n
Driver input----------------· Steer
· Throttle
· Brake
...
Vehicle state--------------· Lane position
· Acceleration
· Speed
...
Strategy 2
Cognitive
distraction
.
.
.
Mitigation
strategy
Strategy 1
Focus of dissertation
An overview of driver distraction mitigation systems
3
Indicators of driver distraction
• Cognitive distraction (subtle, no direct measures of “mind off road”)
o
o
o
o
Concentrate gaze distribution
Impair information consolidation
Degrade driving performance (less serious and consistent)
Impair driver adaptation in tactical driving
Suitable for realtime detection
Performance indicators:
--Eye gaze
Duration and location of fixations
Distance of saccades
Duration, location, distance, and speed of smooth
pursuits
--Driving performance (less serious and consistent)
Abrupt steering control
Large lane-position variability
Miss safety-critical events
Not suitable for realtime detection
4
Detection algorithm for driver distraction
• Driving is complex and continuous human behavior
• Data mining approaches are suitable to detect driver distraction
o Insufficient knowledge impedes using theories to detect distraction precisely
o Data mining techniques can detect non-linear and time-dependent relationships
o Linear regression, decision tree, Support Vector Machines (SVMs), and Bayesian
Networks (BNs) have been used to identify various distractions
Support Vector Machines (SVMs)
Bayesian Networks (BNs)
5
Bayesian Networks (BNs)
• To model probabilistic
relationship among
variables
H
– wide applications, especially
modeling human behavior
• Three kinds of variables
– Hypothesis, evidence, hidden
• Conditional dependency
Cognitive
distraction
Eye
movement
pattern
S
E1
E2
E3
Bayesian Networks (BNs)
Eye movements
Driving performance
Static and Dynamic BNs
• Static BNs (SBNs)
– in single time point
• Dynamic BNs (DBNs)
– across time (Markov process)
t-2
Ht-1
Ht
t-2
St-1
St
E1t-1
E2t-1
E3t-1
T=t-1
E1t
E2t
E3t
T=t
A dynamic BN
• Comparison btw SVM and BNs
– Both can model complex relationships
– Results of BNs can quantify relationships using
information theory measures (such as mutual information)
– DBNs can model time-dependent relationship
– SVMs are more computational efficient than BNs.
Methods
• Data source
– two cognitive conditions
• auditory stock ticker: tracking the change and overall
trends of two stock prices
» without visual distractors
• 4 IVIS drives and 2 baseline drives (15 minutes each)
• to define distraction for models
– data collection (60Hz)
• eye movements
» gaze screen intersection
coordinates
• Driving performance
» lane and steering position
Driving scenario
Data reduction
Plot of eye data
• Eye movements
– eye data eye movements
– 7 eye movement measures
• 3 driving performance
measures
– lane position
– steer wheel position
– steering error
fixation
-duration
-position
smooth pursuit
-duration
-distance
-speed
-direction
blink frequency
Training Data
measures
…...
summarized
instances
(19 measures)
training
data
• Summarization
SBNs, SVMs
– window size
(5, 10, 15, or 30 s)
…...
• Training data
DBNs
…...
…...
…...
Summarization
across window
random
selection
– SBNs SVMs
– DBNs
– 2/3 of total data
SVM and BN training parameters
• SVMs
–
–
–
–


 xi  x j
2
Radial Basis Function (RBF) K xi , x j  e
10-fold-cross-validation to obtain C and γ in the range of 2-5 to 25
Continuous predictors (performance measures)
“LIBSVM” Matlab toolbox
• BNs
–
–
–
–
No hidden node and constrained network structure
Training sequences for DBN –120 seconds long
Discrete predictors
a Matlab toolbox
(Murphy) and an
accompanying
structural learning
package (LeRay)
H1
E1
1
E2
1
T=1
H2
E3
1
E1
2
E2
2
T=2
E3
É ...
Ht
É
É ...
...
É ...
2
E1
É ...
t
E2
t
E3
T=t
11
t
Using SVMs and DBNs to detect cognitive
distraction
SVM prediction for a participant
d'
Comparison between BNs and SVMs
d '  1 ( HIT)  1 ( FA)
12
• Changes in drivers’ eye movements and driving
performance over time are important predictors of
cognitive distraction.
• SVMs have some advantages over SBNs
– Parameter selection: 10-fold across-validation
– Computational ease: training time
• Improving algorithm
– Consider time-dependent relationship in behavior
– Reduce computational load
13
A layered algorithm to detect cognitive
distraction
• Off-line supervised clustering identifies multiple feature behavior
based on subset of behavioral measures based on the training data
o Temporal eye movement measures
o Spatial eye movement measures
o Driving performance measures
Different from clustering, supervised
clustering more likely produce meaningful
clusters in terms of driver cognitive state.
• The higher layer: DBNs
identify cognitive state
from the feature behavior
(cluster labels) with
consideration of time
dependency
14
Supervised clustering
• categorize classified data
The fitness function of supervised clustering (Zeidat et al., 2006) X is a clustering solution,
β is the parameter to balance the ratio of impurity and penalty in the fitness function, k is
the number of clusters in X, n is the total number of data, and c is the number of classes in
the data.
1'
1'’
1
3
2
3'
2
4
A. Traditional clustering
B. Supervised clustering
15
Supervised clustering algorithm
• Single Representative Insertion/Deletion Steepest Decent
Hill Climbing with Randomized Restart – repeat
something similar to SPAM r times and chose the best
• REPEAT r TIMES
– curr = a randomly created set of representatives (with size between c+1
and c)
– WHILE not done DO
• Create new solution S by adding a non-representative or removing a
representative in curr (if size(curr) = k’, new possible solutions are in size of
k’+1 and k’-1 )
• Determine the element s and S for which the objective function in SPAM q(s)
is minimal (if there is more than one minimal element, randomly pick one)
• IF q(s)<q(curr) THEN curr:=s
ELSE IF q(s)=q(curr) AND |s|>|curr| THEN curr:=s
ELSE terminate and return curr as the solution for this run
• Report the best out of the r solutions found
16
Thank you !!
Questions ??
17