Ari Y' Benbasat

1
Inertial Gesture Recognition

• Ari Y. Benbasat
• Responsive Environments Group
• MIT Media Laboratory

2
Compact Inertial Measurement Unit

• Full sensor set for 3D motion detection in
  compact wireless package.
• Implementation
• 3(1) Accelerometers
• 3 Gyroscopes
• 12-bit ADC/Microcontroller
• 900 MHz wireless link
• Low-power (75mW)

3
S.M. Thesis Work
Project Organization

• Create analysis and interpretation framework for
  such devices
• Analysis Activity detection
• Gesture Recognition Parameterized atomic
  gestures
• Output Scripting Links gestures to outputs
• Applications
• Current Re-implementation of (void)
• Future Gesture-based control and learning

4
Activity Detection

• Simple scheme based on windowed variance
• Piecewise model of model used to analytically
  find threshold
• Finds areas of interest in data streams to be
  analyzed by the gesture recognition system
• Err on side of false positives
• Stuttering gestures OK

5
Gesture Recognition

• Parameterized
• Magnitude and duration are properties of the
  detected gestures, not fundamental to the process
• Atomic
• Considered on axis at a time
• Considered only in units of a number of peaks
• Algorithm
• Expects net zero sum (accelerometers)
• Non-trivial size (gyroscopes)
• Pieces together stuttering gestures by combining
  failed gestures
• Breaks gestures if polarity of adjacent peaks is
  identical

6
Output / Applications

• Simple JPython script allows temporal and logical
  combinations of gestures to be linked to output
  routines
• Value in Applications
• Allows direct, in situ sensing of quantities of
  interest
• Compact / low-power?useable in a wide variety of
  situations
• Low complexity of algorithms allows for
  stand-alone devices ? combined perception and
  expertise in single device
• Not limited to human gesture

7
Sample Analysis
Find Areas of Activity
Perform Gestures and Collect Data
y
x
(no rotation)
Run Recognition
Recombine Atomic Output
2 Peaks




3 Peaks
8
Sample Analysis (2)

• Perform gesture
• Sweeping twist
• Find gestures in stream
• One axis at a time
• Note baseline subtraction
• Recombine atoms
• Can be tied to output

1 Peak (gyro)


2 Peaks (acc)
sound light etc.

causes