Gait Recognition using Empirical Mode Decomposition Based Feature Extraction

1

• Gait Recognition using Empirical Mode
  Decomposition Based Feature Extraction
• Prem Kuchi
• Research Associate
• Research Center for Ubiquitous Computing (CUbiC)
• Arizona State University
• Tempe, AZ, USA

2
Introduction

• Gait n, (gat) Manner of walking or stepping
  bearing or carriage while moving
• (Courtesy Websters Revised Unabridged
  Dictionary)

3
Introduction

• Determined by
• Weight
• Limb length
• Habitual posture
• Bone structure
• Age
• Health Status

4
Introduction

• Human Recognition (Biometric)
• The above dependencies make it unique to each
  person
• Recognize known persons and classify unknown
  persons
• Can be used at resolutions where other biometrics
  (face, iris, etc.) fail
• Unintrusive (A major advantage)

5
Problem Statement

• Given the trajectories defined by the time series
  of a set of markers on the human body, the
  problem is to identify the person based on the
  information in a database.

6
Thesis Contributions

• Proposed methodologies for
• the extraction of feature vectors.
• combining trajectory information from multiple
  markers.
• creating a database for recognition and analysis
  of gait.

7
Organization

• Background
• Theory
• Methodology
• Results
• Discussion
• Conclusions and Future Work

8
Existing Approaches - Structural Methods

• Body structure is recovered, and a set of points
  on the body are tracked.

9
Existing Approaches - State-Space Methods

• Represent human movement as a sequence of static
  configurations

10
Existing Approaches - Spatiotemporal Methods

• Action is characterized by the entire 3D
  spatio-temporal (XYT) data volume spanned by the
  moving person.

11
Background

• Traditionally Fourier analysis was used for
  analyzing gait
• Recent research shows that gait is Nonlinear and
  Non-stationary
• Because of nonlinearity, traditional basis
  functions do not work
• Need non-linear basis functions
• Usually empirically derived

12
Background
13
Guiding Principle

• Intrinsic Mode Functions (IMF) capture the
  nonlinear and non-stationary behavior of gait
• Hence, IMFs are used for our analysis
• IMFs are derived using a process called
  Empirical Mode Decomposition
• EMD is effective on short sequences of data

14
Empirical Mode Decomposition

• Proposed by Huang et. al (1998)
• Basis functions
• Empirically derived
• Complete
• Nearly Orthogonal
• Decomposes a signal into a set of Intrinsic Mode
  Functions

15
Intrinsic Mode Functions (IMF)

• A function is an IMF if
• The number of extrema and the number of zero
  crossings are either equal or differ at most by
  1.
• At any point, the mean value of the envelope
  defined by the local maxima and the envelope
  defined by the local minima is zero
• Instantaneous frequency can be calculated from
  IMFs

16
Sifting

• Method for decomposing any general signal X(t)
  into its constituent IMFs
• Two steps
• Two smooth splines are constructed connecting all
  the maxima and the minima of X(t) respectively to
  get its upper envelope Xmax(t) and Xmin(t).

17
Sifting

• The extrema are found my determining the change
  of sign of the derivative of the signal. All the
  data points should be covered by the upper and
  lower envelopes.
• The mean of the two envelopes is subtracted from
  the data to get a difference signal X1(t)

18
Stopping Criterion

• If sifting is carried to an extreme, then we get
  a pure frequency modulated signal of a constant
  amplitude
• So, SD is used as a threshold (set to about 0.2
  to 0.3)

19
Sifting

• This process results in C1(t)
• Now, C1(t) is subtracted from the original signal
  to get the residue R1(t) and the whole process is
  repeated

20
Dynamic Time Warping

• Different people have different stride lengths
• Also, same person might have a different stride
  lengths for different gait cycles
• The length differences may not be uniform in
  nature
• For E.g. A person usually does not slow down or
  speed up during the entire gait cycle.

21
Dynamic Time Warping

• Signals have to be normalized taking into account
  the non-uniformity in the length differences

22
Example
23
Classification

• Multi-Layer Perceptron
• Back-propagation algorithm

24
Combining Multiple Markers

• Bayes Risk Criterion (BRC) is used
• For a single marker (source) BRC can be written
  as follows

25
Combining Multiple Markers

• With Multiple markers, the equation becomes

• Putting P(m1) P(m2) 0.5 (events m1 and m2 are
  equally probable)

? can be tuned for different FAR and FRR
26
Hypothesis I

• IMF decomposition provides a better means of
  extracting features from gait signals than a
  method that employs fixed basis functions

27
Hypothesis II

• Trajectories from multiple markers provide more
  information relevant to gait recognition than
  that from a single marker

28
Dataset Used

• Gait trajectories of 5 subjects in 5 different
  trials were used
• Trajectories of 15 markers placed on the person
  were used

29
Feature Vector Extraction
The mid-frequency IMFs highlight differences
between different peoples gaits
30
Feature Vector Extraction
The Similarity between the residues
Randomness of the First IMFs
31
Feature Vector Extraction

• The first IMF and the residue were discarded
• The rest of the IMFs were added together
• The Fourier Transform of the resulting composite
  signal was the feature vector

32
Single Marker System
Trajectories
Marker Tracking (Nonlinear Filter)
Cycle Extraction And Normalization
Vertical Displacement Of Foot Marker
Empirical Mode Decomposition
Single Marker Processor
Feature Vector Generation
Output Identity
Feature Comparison Engine (MLP)
Database
33
Multiple Marker System
34
Results Single Marker System

• Neural Network Parameters
• number_hidden 250, decay_rate 0.03

CCR for the Single Marker System
35
Results Single Marker System
Output confidence for the toe trajectory
36
Results Single Marker System
Algorithm Performance in presence of Noise
and Output confidence for Case 3
37
Discussion

• Recognition confidence for correct classification
  is high
• Confidence for misclassifications is low
• Performance does not significantly deteriorate
  with noise
• However, with noise, confidence for correct
  classification decreases

38
Results Multiple Marker System
Output confidence for Knee trajectory
Bayes Risk Output
39
Discussion

• The values of Bayes risk are significantly
  greater than 1
• Implies very little ambiguity in deciding the
  identity
• If a particular confidence measure for the toe
  trajectory is low, its corresponding confidence
  measure for knee trajectory is high
• Combining information from multiple trajectories
  is beneficial

40
Conclusions

• Feature vector based on EMD provides satisfactory
  CCR
• Multiple markers increase the CCR and decrease
  the ambiguity during recognition
• Both methods result in high confidence
• So, performance might not drastically decrease
  with increase in the size of the database

41
Future Work

• Tracking specific points on the human body
• Unscented particle filter (UPF), Dual Estimation
• Addition of multimodal information
• E.g. Force Data
• Testing on a larger database

42
DRAG

• DRAG Database for Recognition Analysis of Gait
• Currently, there is no standard database for
  systematic study of gait recognition algorithms
• DRAG provides
• 3D motion-capture data (using external markers)
• Computed joint angles
• Ground reaction loading (using plantar pressure
  insoles)

43
DRAG
LEDs for Marker Data
Plantar pressure insoles for Force data
44
Thank You Questions??