Head Gesture based Control of an Intelligent Wheelchair

1
Head Gesture based Control of an Intelligent
Wheelchair

• Pei JIA and Huosheng HU
• Department of Computer Science
• University of Essex

2
Outline

• Introduction
• Overview of Adaboost
• Overview of Camshift
• Integrated Framework
• Experimental Results
• Conclusion

3
IW Basic Functions

• Interactions between the users and the IWs
• Hand based control joystick, keyboard, mouse,
  touch screen
• Voice based control audio
• Vision based control camera
• Other sensor based control pressure sensors
• The services IWs afford
• navigation
• wandering

4
Exceptions for HGBII

• Head out of the image view
• Only profile face captured
• Varying illumination
• Different face colours
• Different face appearance
• Background noise
• Occlusion

5
Overview of Adaboost

• Classical PR Framework

1) Weak study 2) Independent
Classifiers Design (Supervised Learning Adaboost
Feature Selection, A Cascade of Classifiers)
Feature Extraction (Rectangle features)
Preprocessing (Filtering)
Data Acquisition (Image Capturing)
Classification
6
Overview of Camshift

• Classical Optimization Control

search window size centre
Re-centre the search window, rescale its size
Current search window size is the face size its
centre is the face centre
Both search window size and centre are stable
Target
End Condition Satisfied?
Initialization
Control
Resolution Found
Yes
No
7
Integrated Framework
1) too big or too small? 2) out of the image? 3)
face ellipse upright? 4) ratio of height to its
width is acceptable for a possible face?
1) if (frontal face is detected) nose template
matching to obtain IWActionState 2) else if (only
profile left/right face is detected) IWActionStat
e TURNLEFT/TURNRIGHT 3) else if (both profile
left and right faces are detected) IWActionState
is dominated by the profile face with bigger
detection window
8
Experimental Results (1)

• Camshift Face Tracking is not robust enough

9
Experimental Results (2)

• Adaboost Face Detection is not fast enough
• If
• Captured frame 640480
• Minimum face size 2020.
• Adaboost Refinement 320240.
• The face is lost once every ten frames.
• Thus, the average time cost per frame of our
  proposed method is 0.350.0890.0110/10
  0.117, which is much less than 0.35 seconds.

10
Experimental Results (3)

• Face Direction Recognition by the proposed method

11
Experimental Results (4)

• Video Simulation
• http//privatewww.essex.ac.uk/pjia/Research/Proje
  ct.html

12
Q A
Thank you very much!
Welcome 2 http//www.longervision.com/phpBB