Using Mobile Phones to Determine Transportation Modes

Using Mobile Phones to
Determine Transportation Modes
Sasank Reddy, Min Mun, Jeff Burke,
D. Estrin, M. Hansen, M. Srivastava
TOSN 2010
Determining Transportation Modes
• Possible transportation modes:
– Stationary, walking, running, biking, motorized
• Applications:
– Physical activity monitoring
– Personal Impact and/or Exposure Monitoring
– Transportation and Mobility-based Recruitment
(distributed data gathering)
Related Work
Pedometer (e.g., Omron)
Mobile phones based pedometers
FitBit/Philips Tracmor (calorie burnt), BodyMedia (GSR…)
SenseWear (accel, audio, barometer)
Cell-tower assisted detection
– Signal strength of multiple cell-towers nearby (but it’s not available under
current smartphones)
– Wi-Fi/cell-tower : coarse grained classification (stationary, walking, motorized)
• This work: GPS + accel in smartphones: still, walk, run, bike, vehicle
Design Principles
• Design principles:
– Using only a mobile phone (less cpu/memory footprint,
energy efficient)
• Outdoor movement is monitored w/ GMS cell tower (which
triggers GPS outdoors)
– Not susceptible to position/orientation
• Feature selection must be good enough
– Able to work for a variety of users w/o additional training
• Classifier should be generic enough
– Not using external spatial data (or user’s movement
history data)
• Even without such data it should be accurate enough
• Design a new classifier w/ mobile phone (GPS, Accel)
Design Space
• Sensor selection
– Bluetooth
– WiFi and GSM
– Accelerometer and GPS
Feature Selection
Window size: 1 second
– Too small: may not effective (esp: accel sampling freq)
– Too large: introduces noise (spanning multiple activity)
Types of features:
– Mobility: accelerometer
Magnitude of force vector
Mean, variance, energy
DFT (discrete Fourier transform): energy cofficients between 1-10 hz (Bao and Intille 2004)
Appropriate for detecting pedestrian based motion
– Speed: GPS
Invalid points? Analyzing accuracy, dilution of precision, change of speed, etc..
Correlation-based feature selection
– (CFS vs. PCA..); CPS eliminates irrelevant/redundant attributes..
– Accel: variance along with DFT energy coefficients between 1-3Hz
– GPS: speed
Classifier consideration:
– Instance classifiers: DT, KMC, Naïve Bayes (NB), Nearest Neighbor (NN), SVM, etc.
– HMM-based classifier: continuous vs. discrete
Phone Position and Accuracy
User Variation

similar documents