Using Mobile Phones to Determine
Transportation Modes
Sasank Reddy et al.,
ACM Transactions on Sensor Networks, Vol. 6, No. 2, Article 13, Feb 2010
2011.04.11
Hyeong-il Ko
Contents
• Introduction
• Related Work
• Approach
• Experimental Setup
• Results
• Conclusion
1
S
FT COMPUTING @ YONSEI UNIV . KOREA
Introduction
• Mobile phones
– Computation, sensing, and communication capabilities
– Carried by people throughout the day
• Target applications for transportation mode inference and
location information
– Physical Activity Monitoring
– Personal Impact and/or Exposure Monitoring
– Transportation and Mobility-Based Recruitment
2
S
FT COMPUTING @ YONSEI UNIV . KOREA
Related Work
3
S
FT COMPUTING @ YONSEI UNIV . KOREA
Approach
• Design Goal
– User convenience
• 4 properties of suggested system to be convenient
– Contained in one sensing unit
– Flexible in terms of the position and orientation
– Able to work for a variety of users without additional training
– Not reliant on external spatial or user history based indexes
• Contribution
– Suggested classifier that uses information from an accelerometer
and a GPS
– Able to run on a commodity mobile device
4
S
FT COMPUTING @ YONSEI UNIV . KOREA
Approach
• Sensor Selection
– Bluetooth
• Not ubiquitous in outdoor settings
– Static Bluetooth beacons mainly exists indoor settings
• Difficult to distinguish if an individual is moving
• Difficult to distinguish if an environment is changing
– Other people carrying devices are moving
5
S
FT COMPUTING @ YONSEI UNIV . KOREA
Sensor Selection
– WiFi and GSM
• Not discriminative when speed profiles are similar
– Slow moving traffic, biking, and walking
• Depends on the density network and end points
6
S
FT COMPUTING @ YONSEI UNIV . KOREA
Sensor Selection
– Accelerometer and GPS
• 10~20% accuracy dropped if only 1 sensor of them is used
• Negligible 0.6% improved when 4 sensing modalities are used
7
S
FT COMPUTING @ YONSEI UNIV . KOREA
Feature Selection
• Feature Selection
– Window Size
• A window of 1 sec
– Type of Features
• Accelerometer
– Magnitude of the force vector from 3 axises
– Mean, variance, energy, and DFT energy coefficient between 1-10Hz
• Speed
– Value directly used from GPS receiver
• Noise filtering step
– Discarding GPS points deemed invalid
– Excluding accelerometer data if too few samples are received for classification
8
S
FT COMPUTING @ YONSEI UNIV . KOREA
Feature Selection
• Selection Method
– CFS(Correlation based Feature Selection)
• Feature subset selector that eliminates irrelevant and redundant attributes
• Feature subset
– Variance along with DFT energy coefficients between 1-3Hz
– Speed from the GPS receiver
9
S
FT COMPUTING @ YONSEI UNIV . KOREA
Classifier Selection
• Classifiers
– Instance Classifiers
• E.g. C4.5 DT, KMC, NB, NN, and SVM
– CHMM(Continuous HMM)
• Output symbols : independent multi-variate Gaussian distributions
• Hidden states : classification classes
• Transition probability
– Instance based classifier + DHMM(Discrete Hidden Markov Model)
• DHMM output symbols : instance-based classifications
• Hidden states : classification classes
• State transition probabilities
10
S
FT COMPUTING @ YONSEI UNIV . KOREA
Experimental Setup
• Hardware Platform
– Nokia n95
•
•
•
•
•
•
•
•
•
CPU : 332 MHz ARM processor
RAM : 128MB
3 axis accelerometer that can sample at 32 Hz
Built-in GPS receiver that can sample at 1 Hz
WiFi radio that can scan at 0.33 Hz
GSM cell radio that can sample at 1 Hz
Bluetooth radio that can scan at 0.08 Hz
Battery : 950 mAh
OS : Symbian S60 3rd Edition
• Software Setup
– Weka Machine Learning Toolkit
– Generalized Hidden Markov Model library
– Python
11
S
FT COMPUTING @ YONSEI UNIV . KOREA
Experimental Setup
• Data Collection
– Volunteers to obtain data set
• 16 individuals
– 8 male + 8 female
– The ages of 20-45
– Accelerometer, GPS, WiFi, and GSM information obtained
– How to collect data
• 1.25 hrs of data per position per individual
– Positions : Still, Walk, Run, Bike, Motor, All
• Total 120 hrs
12
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Classification Accuracy
13
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Structure of Overall Classifier
14
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Device Placement Variation
– Mobile phones are often carried at different positions
– Classifier is trained on data from all 6 positions
• Arm, bag, chest, hand, pocket, and waist
15
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Extended Transportation Mode Traces
– Would the DT+DHMM classifier perform in “everyday” use?
– 1 of the volunteers
• carried the mobile phone over 4 weeks
• documented instances of each of the transportation modes
16
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Extended Transportation Mode Traces(Cont’d)
– Would the DT+DHMM classifier perform in challenging urban
environment?
– The volunteer
• Carried the mobile phone 3.5 hrs in urban canyons
• At least 30 mins for each transportation mode
– Result
• Average accuracy : 92.6%
• Still, walking, and motorized transport : 95%
• Biking and running state : around 88%
17
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Memory and CPU Benchmarks
– Using Nokia Energy Profiler
• 20 mins trials were performed
18
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Energy Consumption
19
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
• Energy-Aware Detection
– Authors’ objective
• To create a transportation classifier that captures the behavior of individuals when
they are outside
– The classifier should be energy efficient
– The most effective method to determine when the user is outdoors
again
• GPS
– Sampling GPS for the purpose is power hungry
– Trigger approach proposed
• Attempting to sample the GPS when only changes occur to the primary GSM cell
tower would be more efficient in terms of energy usage
• GSM cell towers are used to determine the start of outdoor trips
• Filter to eliminate the “ping pong” effect
20
S
FT COMPUTING @ YONSEI UNIV . KOREA
Results
– Trigger approach proposed(Cont’d)
• To test the performance of the GSM triggered approach,
– 16 individuals labeled indoor/outdoor status and collected GSM cell tower every
1 sec for a day
• Total time of the day trace data collection
– Average : 23.2 hrs
– Mininum time : 20.7 hrs
– Maximum time : 26.8 hrs
• Corresponding outdoor time
– Average : 3.09 hrs
– Maximum : 12.0 hrs
– Minimum : 0.93 hrs
• The average percentage of outdoor time identified
– 91.5%
• 12.4% energy save compared to GPS
21
S
FT COMPUTING @ YONSEI UNIV . KOREA
Conclusion
• Transportation mode classification system
– Distinguishing between being stationary, walking, running, biking
and in motorized travel
– Employing a DT followed by a DHMM
– Using a mobile phone equipped with a GPS receiver and an
accelerometer
– Convenient for a user
• By not having strict position and orientation requirements
– Achieving a high accuracy level(93.6%)
• Based on a dataset of 120 hrs of data from 16 users
– Not relying on external spatial indexes
– Working well without user-specific training information
22
S
FT COMPUTING @ YONSEI UNIV . KOREA