Recognizing Human Activities

1

• Recognizing Human Activities
• In
• Time Sequential Images
• CS 365 project
• Presented By
• Abhishek Gaurav
• Patala Arun Kumar

2
Approaches

• Top down approach
• In this approach 3-d model of human is
  constructed

3
Approaches

• Bottom up approaches
• These approaches are more robust.
• Based on feature extraction.
• View point affects the recognition capability.
• This is what we are going to implement
• Two methods
• The Representation and Recognition of Action
  Using Temporal Templates
• Recognizing human action in time sequential
  images using Hidden Markov Model

4
Temporal Template Method

• First describing where there is motion and then
  describing how the motion is moving.
• This is done by construction of
• Binary Motion Energy Image (MEI) which represents
  where motion has occurred.
• Motion History Image (MHI) in which intensity
  represents the recency of motion.
• In the training section requires videos from
  different view points

5
MEI and MHI

• Motion Energy Images Motion
  History Images

6
Technical Details

• Steps
• I (x, y, t) image sequence
• D (x, y, t) binary image sequence indicating
  motion
• ET (x, y, t) union (i 0 to T-1) D (x, y, t
  i) binary image which indicates area of
  motion from t T to t only one ET is
  calculated for one activity.
• HT (x, y, t) T if D (x, y, t) 1
• max (0, HT (x, y, t 1)
  1) otherwise
• Hu Moments are calculated for ET (x, y , T) and
  HT (x ,y ,T)
• For the same action these Hu Moments are stored
  with different view points as different sets.
• They are calculated for different persons several
  times and stored as covariance matrix to
  accommodate the variations.
• To recognize any action its Mahalanobis distance
  is calculated from stored actions and the closest
  match is reported using different Ts.

7
Recognition Using Hidden Markov Model (HMM)

• HMM models
• Markov Chain In this chain there are of set of
  states in a sequence and the probability of
  occurring a state depends on previous states i .
  e P (xn1 q xn)
• Hidden Markov Model
• In Hidden Markov Model states are not observable
  but output symbols are observable whose
  probability depends on state.
• A first-order MM example

8
HMM

• T length of the observation sequence.
• Q q1, q2, , qN set of states.
• N number of states in the model.
• V v1, v2, , vM set of possible output
  symbols.
• M number of observation symbols.
• A aij, aij is the probability of transiting
  from state qi to state qj.
• B bj(k), bj(k) is the probability of output
  symbol vk at state qj.
• Z zi, zi is probability of initial state
  being qi.
• L A, B, Z complete parameter set of an HMM.
• Recognizing a output symbol sequence O O1, O2,
  , OT means getting i such that prob (Li O) is
  maximum which is done using Forward Algorithm.
• Training means building HMM for each category by
  optimizing the model parameters A, B, Z such
  that probability of observed sequence prob (O
  L) is maximized. This is done using Baum Welch
  Algorithm.

9
Using HMM

• Extraction of feature using
  Feature vector to output symbol
• Mesh feature analysis. and
  processing to HMM.
• Feature vector is got.

10
Results
11
CITATIONS

• The Representation and Recognition of Action
  Using Temporal Templates
• James W. Davis and Aaron F. Bobick
• http//alumni.media.mit.edu/jdavis/Publicatio
  ns/publications_402.pdf
• Recognizing Human Action in Time-Sequential
  Images using Hidden Markov Model
• Junji YA MAT0 Jun OHYA Kenichiro ISHII
  http//ieeexplore.ieee.org/iel2/418/5817/00223161.
  pdf?isnumberarnumber223161