a framework of energy efficient mobile sensing for automatic human state recognition, at mobisys...
TRANSCRIPT
A Framework for Energy Efficient Mobile Sensing for Automatic User State Recognition
Yi Wang, Jialiu Lin, Murali Annavaram, Quinn A. Jacobson,
Jason Hong, Bhaskar Krishnamachari and Norman Sadeh
OUTLINE
Motivation
EEMSS introduction
System modules
Case implementation
Performance evaluation
Conclusion
PEOPLE-CENTRIC MOBILE SENSING
Health monitoring
Keeping track of patients/elders
Social networking
Automated “Twittering”
Automated profile updates
Ringtone adjustment
Location based services
Mobile advertising
CHALLENGE: LIMITED BATTERY
Battery capacity of mobile device is low
Sensors are main source of energy usage
Blind sensing drains battery soon
Need intelligent sensor management
Energy Efficient Mobile Sensing System (EEMSS) is important
Possibly trade some detection accuracy for much longer device lifetime
SENSOR MANAGEMENT METHODOLOGY
Only utilize a minimum set of sensors
Recognize state & detect state transition
E.g.: No need for GPS when indoor
Hierarchical management
Sensors are activated when necessary
E.g.: Accelerometer -> WiFi scan -> GPS
If multiple sensors achieve the same task: Use energy efficient ones
How to determine sensor energy efficiency?
DETERMINE ENERGY EFFICIENCY
Energy = Power drain × Operating duration
Sensor power consumptions on N95 devices:
Keep an eye on sensor operating duration
USER STATE DESCRIPTION
Task 1: Identify the states to be detected Working/meeting/walking/driving outdoor …
Task 2: For each state define entry criteria Sensing thresholds from one or more sensors
E.g.: “Outdoor” + “High travel speed” = “Vehicle”
A state is entered when the criteria are satisfied
Task 3: For each state also specify the necessary sensors to be monitored Detect state transition
DESIGN BENIFITS
Scalable
Add or remove a state upon user’s interest
Different criteria for different individuals
Real-time update
User’s habits may change
Sensing criteria can to be refined in real-time
CLASSIFICATION MODULE
Classification algorithms are the key to high state recognition accuracy
Mobile phone is the only sensing resource
Capability limitations
Computing issues
Real-time
Computing power
REAL-TIME AUDIO RECOGNITION
Based on microphone sensing
Focus on the detection of speech
Audio features Energy Silence ratio SSCH peak1
3 outputs: Speech Loud/Noisy Silent
1 B. Gajic and K.K. Paliwal. “Robust speech recognition in noisy environments based on subband spectral centroid histograms”
SPEECH DETECTION PERFORMANCE
Algorithm is tested on 1085 speech clips
91.14 % are classified as speech
Complexity of algorithm is O(N2)
N: Number of frequency samples
Overall processing time of a 4 seconds sound clip on N95
~10 seconds
EEMSS CASE IMPLEMENTATION
Implemented and tested on Nokia N95
Sensors operated include:
Accelerometer, Microphone, WiFi detector, and GPS
User states are featured by:
Location, background sound, and user motion information
E.g.: “office + quiet + still” => User working in office
STATES INTERESTED
State Name State Features Sensors Monitored
Location Motion Background Sound
Working Office Still Quiet Accelerometer, Microphone
Meeting Office Still Speech Accelerometer, Microphone
Office_loud Office Still Loud Accelerometer, Microphone
Resting Home Still Quiet Accelerometer, Microphone
Home_talking Home Still Speech Accelerometer, Microphone
Home_entertaining Home Still Loud Accelerometer, Microphone
Place_quiet Some Place Still Quiet Accelerometer, Microphone
Place_speech Some Place Still Speech Accelerometer, Microphone
Place_loud Some Place Still Loud Accelerometer, Microphone
Walking Keep on changing Moving Slowly N/A GPS
Vehicle Keep on changing Moving Fast N/A GPS
SENSOR DUTY CYCLES
When activated, sensors are turned on and off periodically
Trade-offs in sampling:
Frequent sampling: Wastes energy and provides redundant information
Infrequent sampling: Saves energy but low state detection accuracy
Current implementation provides small event detection delay
EEMSS PERFORMANCE EVALUATION
We evaluate EEMSS in terms of: Ability to record one’s state in real time State recognition accuracy Energy efficiency
User study at CMU & USC with 10 users Each user carried Nokia N95 phone as daily
used cell phone with EEMSS running Ground truth was manually recorded by
each userFine-grained entries with time and state
records
STATE RECOGNITION ACCURACY
Confusion matrix of recognizing “Walking”, “Vehicle”, and all other states*
All other states
Walking Vehicle
All other states 99.17% 0.78% 0.05%
Walking 12.64% 84.29% 3.07%
Vehicle 10.59% 15.29% 74.12%
* “All other states” include “Working”, “Meeting”, “Office_loud”, “Resting”, “Home_talking”, “Home_entertaining”, “Place_quiet”, “Place_speech”, and “Place_loud”.
EEMSS ENERGY USAGE AT A GLANCE
User worked in office, then walked to library and stayed (20 min empirical interval)
CONCLUSION
User state recognition based on mobile sensing is popular
Energy efficiency is required due to low device battery capacity
Our sensor management methodology: Utilizing minimum number of sensors to
accomplish sensing tasks
Manage sensors hierarchically
EEMSS achieves good state recognition accuracy and energy efficiency