Upload
ifitt
View
121
Download
1
Embed Size (px)
DESCRIPTION
Predicting from GPS and Accelerometer DataWhen and Where Tourists Have Viewed Exhibitions
Citation preview
ENTER 2014 Research Track Slide Number 1
Predicting fromGPS and Accelerometer Data
When and Where Tourists Have Viewed Exhibitions
aJunya KawaseaYohei KuratabNaoto Yabe
aDepartment of Tourism ScienceGraduate School of
Urban Environmental SciencesTokyo Metropolitan University, Tokyo,
Japan
bGraduate School of EducationJoetsu University of Education, Niigata,
Japan
ENTER 2014 Research Track Slide Number 2
They measure:• three-dimensional position• velocity
GPS devices
ENTER 2014 Research Track Slide Number 3
We can tell:
• Where the subjects have passed
• Where the subjects have been staying
Like this point.
ENTER 2014 Research Track Slide Number 4
Problem
However, are these techniques sufficient for tourist activity surveys?
An intelligent technique for inferring tourist activities from their GPS logs will be highly desirable.
ENTER 2014 Research Track Slide Number 5
Our Previous SurveyReported at ENTER2012
(Kawase et al.2012)
We found – an significant relationship between tourists’
walking speed and their viewing behaviours.– Staying is not equivalent to viewing.
ENTER 2014 Research Track Slide Number 6
Outline of Our Previous Experiment
• Purpose : To investigate the relationship between people’s activity (viewing or not) and GPS data
• Subjects : 5 students (3 males & 2 females).• The subjects went around the zoo freely,
carrying a GPS logger• They were video-recorded
ENTER 2014 Research Track Slide Number 7
Result: Walking speedand Probability of Viewing
ENTER 2014 Research Track Slide Number 8
tt
tt v
P
PP 3237.06194.0
1loglogit
Logistic Regression Model
EstimateStandard
deviation
Chi-squared
p-valueExp
(estimate)
0.6194 0.0226 752.1810 0.0000
-0.3237 0.0105 948.5306 0.0000 0.7235
Probability of viewing at time t
Walking speedat time t
ENTER 2014 Research Track Slide Number 9
• Remaining problems– Only one parameter to presict tourists‘
behaviours– Limited subjects
• Additional experiment– More parameters– With larger number of subjects,
including families accompanying kindergartner
ENTER 2014 Research Track Slide Number 10
Outline of Additional Experiment
• Tama Zoological park in Tokyo• Subjects
– 5 student pairs in their 20s.– 9 families
parent(s) with kindergartners• They were video-recorded
ENTER 2014 Research Track Slide Number 11
Data Example
The subject was viewing an exhibition
The subject was not viewing an exhibition
ENTER 2014 Research Track Slide Number 12
Candidates for Parameters
• Visitors’ walking speed: vt
• Increase of speed from previous log: at
• Distance to the nearest exhibition: dt
• Decrease of distance from previous log:r-
t= -(dt-dt-1)/dt
• Decrease of distance to the next log:r+
t= -(dt+1-dt)/dt
ENTER 2014 Research Track Slide Number 13
Probability of viewing at time t
Walking speedat time t
Estimate Standarddeviation
Chi-squared p-value
Exp(estimat
e)β0 1.6713 0.0231 5243.7826 0.0000
β1 -0.6531 0.0099 4312.5704 0.0000 0.5204
β2 0.2217 0.0197 127.1331 0.0000 1.2481
β3 -0.0747 0.0015 2472.2617 0.0000 0.9280
β4 -0.0107 0.0039 7.4203 0.0064 0.9893
Increase of speed from previous log
at time t
Distance to the nearest exhibition
at time t
Decrease of distance to the next log at time t
ttttt rβdβaβvββp 43210 +++logit
Logistic Regression Model for 20s
The subject is viewing an exhibition more likely when he/she is located close to the exhibition, and walking
slowly or pausing.
ENTER 2014 Research Track Slide Number 14
Table 1. Number of activity logs of the students in their twenties
Logistic Regression Modelfor 20s
Predictionsuccess
rateViewingNot
Viewing
ActualViewing 18790 6438 74.48%
Not Viewing 6369 13670 68.22%
This model Model usingvt alone
NagelkerkeR2 0.3106 0.2063
Predictive value 71.71% 70.46%
NagelkerkeR2 and Predictive value
By using additional parameters,we successfully improved our modelfor estimating the visitors’ viewing
activities.
ENTER 2014 Research Track Slide Number 15
Logistic Regression Modelfor families
Probability of viewing at time t
Walking speedat time t
Estimate Standarddeviation
Chi-squared p-value
Exp(estimat
e)β0 0.5449 0.0211 669.750 0.0000
β1 -0.3137 0.0098 1014.3037 0.0000 0.7308
β2 0.0712 0.0200 12.7142 0.0004 1.0738
β3 -0.0902 0.0020 2060.2309 0.0000 0.0321
β4 0.0068 0.0068 0.0231 0.1564 1.0068
Increase of speed from previous log
at time t
Distance to the nearest exhibition
at time t
Decrease of distance to the previous log at time t
ttttt rβdβaβvββp 43210 +++logit
ENTER 2014 Research Track Slide Number 16
Table 2. Number of activity logs of the kindergartners
Logistic Regression Modelfor families
Predictionsuccess
rateViewingNot
Viewing
ActualViewing 4548 8064 36.06%
Not Viewing 3127 21337 87.22%
This model Model usingvt alone
NagelkerkeR2 0.2085 0.2063
Predictive value 69.77% 70.46%
NagelkerkeR2 and Predictive value
ENTER 2014 Research Track Slide Number 17
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.0~
0.2~
0.4~
0.6~
0.8~
1.0~
1.2~
1.4~
1.6~
1.8~
2.0~
2.2~
2.4~
2.6~
2.8~
3.0~
3.2~
3.4~
3.6~
3.8~
4.0~
probabilityof viewing
walking speed level (km/h)Kindergartners 20s
In Sum• For the students: Successful• For the family(kindergartner): Insufficient
The kindergartners showed interests in many things and switched their focus quickly from one things to another.
Adults also have possible activities other than viewing an exhibition
even in front of an animal exhibition
ENTER 2014 Research Track Slide Number 18
・ Accelerometer・ Gyrocompass
etc
ENTER 2014 Research Track Slide Number 19
Considering AdditionalHuman Behaviours’ Data
• Accelerometers
• The studies of human behaviours using accelerometers are already seen in many fields
ENTER 2014 Research Track Slide Number 20
Our Pilot Experiment
• We using: GPS and Accelerometer loggers
ENTER 2014 Research Track Slide Number 21
Our Pilot Experiment
• We calculatedthe average of absolute value of each output for every second, considering that these parameters indicate the intensity of the subject’s movement.
ENTER 2014 Research Track Slide Number 22
Logistic Regression Model
Estimate
Standard
deviation
Chi-squared p-value
Exp(estimat
e)
β0 4.0288 0.1806 497.783 0.0000
β1 -0.3413 0.0372 84.2470 0.0123 0.7109
β2 -0.0001 0.0000 6.2691 0.0231 0.9999
β3 -0.0002 0.0001 5.1617 0.0000 0.9998
β4 -0.0012 0.0001 270.7085 0.0000 0.9988
β5 -0.0006 0.0001 102.5665 0.0000 0.9994
ytxtytxttt aβaβvββp 543210 ++logit
Probability of viewing at time t
Acceleration x-axisat time t
Angular velocityaround x-axis
at time t
Angular velocityaround y-axis
at time tAcceleration y-axis
at time t
The subject is viewing an exhibition more likely
when he/she is taking slower action.
ENTER 2014 Research Track Slide Number 23
Logistic Regression Model
This model Model of Section3
NagelkerkeR2 0.4635 0.3106
Predictive value 76.67% 71.71%
NagelrkerkeR2 and Predictive value
• the output of accelerometers seems highly useful for improving the reliability of the prediction model.
ENTER 2014 Research Track Slide Number 24
Conclusions• We revealed
– we can make more reliable model for youth by considering additional parameters available from GPS logs
– On the other hand, it is difficult to predict kindergarteners’ viewing state from the same parameters
• We demonstrated that the combined use of a GPS sensor and an accelerometer is promising