ViewInvariant Representation and Recognition of actions

1
View-Invariant Representation and Recognition of
actions

• Paper By Rao,Yilmaz and Shah)
• International Joural of Computer Vision
• 50(2),203-226,2002
• Presented ByXiangdong Wen
• Advisor L.J. Latecki

2

• Introduction
• Related work
• Perception of Motion
• Representation
• Learning
• Experiments
• Conclusion

3
Natural Actions

• Events
• 1.Low level description
• change of direction,stop, pause,
• 2.High level description
• opening a door,starting a car
• Temporal textures
• ripples on the water,
• a cloth waving in the wind
• Activities
• walking, running, jumping

4
Recognition of human actions

• Extract relevant information.
• Represent it in a suitable form.
• an abstraction of the data
• view-invariant,compact,reliable
• Interpret visual information.
• recognition
• learning

5
Related work

• Izumi and Kojiama(2000) head model
• Siskind and Moris(1996) HMM system
• Davis et al.(2000) a sinusoidal model
• Polana(1994) normal flow
• Madabushi and Aggarwal(2000) head
• Seitz and Dyer(1997) cyclic motion
• Tsai et al.(1994) FFT find the period
• Bobick and Davis(1997) aerobic exercise

6
Perception of Motion

• Human Perception
• Spatio-Temporal Curvature
• How it Captures motion boundaries
• Previous Approaches
• Generate and Smooth of Trajectories

7
Human Perception

• Dynamic instant an instantaneous entity that
  occurs for only one frame.
• Intervalthe time period between two dynamic
  instants.

8
Sample movies(1)
9
Sample Movies(2)
10
Spatio-Temporal Curvature

• rx(t),y(t),t
• vx(t),y(t),1
• ax(t),y(t),0
• r(t) X r(t)
• K(t)------------------------
• r(t)3

11
Previous Approaches

• Rubin and Richards(1985)
• Polar coordinates, they using s(t) and
  Angle(t) separately.
• Gould and Shah(1989)
• Velocity vector v(t)v_x,v_y
• Trajectory Primal Sketch(TPS)
• Both have alignment problem.

12
Generating and Smoothingof Trajectories

• Skin detection by Kjeldean and Kender(1996).
• Mean-shift tracking by Comaniciu et al.(2000).
• Anisotropic diffusion smoothing by Perona and
  malik(1990).

13
Representation

• Instants
• Timethe frame number
• Position the position of the hand
• Sign of the turning angle
• intervals

14
View-Invariance

• The number of instants
• The signs of instants

15
Learning

• Starting with no modal
• Matching two actions
• 1.same number of instants with same sign.
• 2.Use hand positions to form a 4n matrix M, If
  Rank(M)lt4 then match
• Match error ?_4.

16
Experiments

• Using 47 different action clips performed by 7
  individuals.
• At most one hand in each frame.
• Compare the speed to find the hand.
• Results are amazing.

17
Conclusion

• View-invariant
• Dynamic instants and intervals
• Spatio-temporal curvature
• The system Learns without training
• Experiments results are good

18
THANK YOU!