Presentation is loading. Please wait.

Presentation is loading. Please wait.

Behavior Classification by Eigen-decomposition of Periodic Motions Michael Rudzsky Joint work with Roman Goldenberg, Ron Kimmel, Ehud Rivlin Computer Science.

Similar presentations


Presentation on theme: "Behavior Classification by Eigen-decomposition of Periodic Motions Michael Rudzsky Joint work with Roman Goldenberg, Ron Kimmel, Ehud Rivlin Computer Science."— Presentation transcript:

1 Behavior Classification by Eigen-decomposition of Periodic Motions Michael Rudzsky Joint work with Roman Goldenberg, Ron Kimmel, Ehud Rivlin Computer Science Department Technion-Israel Institute of Technology Geometric Image Processing Lab

2 Dynamism of a Dog on a Leash Giacomo Balla, 1912

3 The Red Horseman Carlo Carra, 1914

4 Muybridge Horse Eadweard Muybridge, Animals in Motion, 1887

5 Horse - decomposition

6 Segmentation and Tracking qActive Contour qFast Geodesic Active Contours uAOS uLevel Sets uFast Marching Goldenberg, Kimmel, Rivlin, Rudzsky, IEEE T-IP 2001

7 Tracking in color movies Goldenberg, Kimmel, Rivlin, Rudzsky, IEEE T-IP 2001

8 Background qBackground subtraction Chan, Vese, Active Contours without Edges, IEEE T-IP 2001 Paragios, Deriche, Geodesic Active Regions for Motion Estimation and Tracking, ICCV-99

9 Tracking

10

11

12 Information extraction

13 Walking man - periodicity

14 Walking cat -periodicity

15 Periodicity Analysis tt+3t+6t+9  +3  +6  +9

16 Inter-frame correlation

17 Spatial Alignment 50x50

18 Temporal Scaling Original period - 11 frames Resampled period - 10 frames

19 Temporal Alignment

20 Parameterization qn - number of frames in the training set q50x50 - normalized images qM 2500 x n - training samples matrix  MM T = U  V T, the principle basis {U i, i=1..k}

21 Distinguishing by static appearance qImage I written as a vector v I qParameterized representation in basis B, p = B T v I qDTFS ||p - v I ||

22 Back-projection Original 11 frame one period subsequence Projection to the `dogs & cats’ basis and the DTFS values

23 Recognizing motions q{I f, f=1..T} - one period, temporally aligned set of normalized object images qp f, f=1..T - projection of the image I f onto the principal basis B of size k qOne-period subsequence representation Vector P of size kT - (p f, f=1..T) qIf k = 20 and normalized duration of one-period is T=10, then P is of size 200.

24 Classification -dogs & cats walkrungallopcat...

25 Classification -dogs & cats

26 Classification -people walkrunrun45

27 Classification -people

28 Learning curves Dogs and catsPeople

29 Parameterized modeling Ju, Black, Yacoob, Cardboard People, ICFG-96 Black, Yacoob, Parameterized Modeling and Recognition of Activities, ICCV-98

30 Optical Flow Polaba, Nelson, Recognizing Activities, ICPR-94

31 Motion History Images (MHI) James W. Davis, OSU Motion Recognition Lab

32 Star Skeleton Distance from the center of mass After low pass filter

33 Phase portrays Shavit, Jepson, Motion Understanding from Qualitative Visual Dynamics, 1993

34 Summary qSegmentation uactive contours qPeriodicity analysis uglobal contour characteristics qAlignment uspatial utemporal qParameterization uPrincipal basis projection qClassification uNearest neighbors


Download ppt "Behavior Classification by Eigen-decomposition of Periodic Motions Michael Rudzsky Joint work with Roman Goldenberg, Ron Kimmel, Ehud Rivlin Computer Science."

Similar presentations


Ads by Google