biometric gait recognition cmpe 58z introduction to biometrics term project

40
BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT MUSTAFA OZAN ÖZEN PINAR SAĞLAM LEVENT ÜNVER MEHMET YILMAZ

Upload: abdalla

Post on 22-Feb-2016

64 views

Category:

Documents


0 download

DESCRIPTION

BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT. MUSTAFA OZAN ÖZEN PINAR SAĞLAM LEVENT ÜNVER MEHMET YILMAZ. MOTIVATION. Gait : Particular way or manner of moving on foot . Gait Recognition : Identifying people with respect to their gait features . - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

BIOMETRIC GAIT RECOGNITIONCMPE 58Z INTRODUCTION TO BIOMETRICS

TERM PROJECT

MUSTAFA OZAN ÖZENPINAR SAĞLAMLEVENT ÜNVER

MEHMET YILMAZ

Page 2: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

MOTIVATION Gait: Particular way or manner of moving on

foot. Gait Recognition: Identifying people with

respect to their gait features. Advantages:1. Can be used at distance2. Can be used at low resolution3. Acceptable by people

Page 3: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

General Gait Recognition Approaches CASIA Database The approaches we currently used:1. “Averaged Sillhouettes” Approach.2. “Absolute Joint Positions” Approach.3. “Abdelkader’s Eigengait” Approach.4. “What if it happens?” Approach.

OUTLINE

Page 4: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

General Gait Recognition Approaches

Gait Recognition Approaches

MV-Based FS-Based WS-Based

Silhouette-Based Model-Based

Page 5: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

CASIA Gait Database In this project, CASIA GaitDataBaseA is

used CASIA GaitDataBase:i. Has 20 different persons data. Each

person has 12 different sillhouette gait data set. But we only used 2 or 4 dataset (from right to left gait data).

ii. In other words, there were one test and one training data set for each person. Each data set consists of max. 75, min. 37 frames

Page 6: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

CASIA Gait DataBase – Sample Sillhouettes

Page 7: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

“Averaged Silhouettes” Appraoch

Silhouette Extraction Gait Cycle Calculation Averaged Silhouette Respresentation Similarity Computation Results and Discusion

Page 8: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Silhouette Extraction GMM to extract silhouettes

Unable to download the database

Sample silhouettes from CASIA Database

Page 9: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Gait Cycle Calculation

0 10 20 30 40 50 60 70 802200

2400

2600

2800

3000

3200

3400

3600

0 10 20 30 40 50 60 70 802000

2200

2400

2600

2800

3000

3200

3400

Page 10: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Gait Cycle Calculation Problem in Gait Cycle Calculation

Can not estimate gait cycle

What to do?????

Page 11: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Gait Cycle Calculation

0 10 20 30 40 50 60 70 8025

30

35

40

45

50

55

60

65

70

75

0 10 20 30 40 50 60 70 8025

30

35

40

45

50

55

60

65

70

75

Page 12: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Averaged Silhouette Representation

Page 13: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Averaged Silhouette Representation (Direction

Correction)

Page 14: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Averaged Silhouette Representation (Height

Normalization)

Page 15: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Similarity Computation Calculate Euclidean Distance

Form the Similarity Matrix

Page 16: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Results And Discussion EER = 58.9% Closed Set Identification

Rate = 73.68% Individual Silhouette

Frames = ~73% Averaged Silhouette (From

paper) = ~79% Low EER

=> Low quality silhouettes Not so bad Closed Set

Identification Rate0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Page 17: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Absolute Joint Positions In the case of this project, the feature points

are the position of the joints.

PCA is applied to these feature points and the feature size is reduced.

Then, spatio temporal correlation is used for classifying.

Page 18: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Absolute Joint Positions Absolute joint

positions – the physical positions of each joint in each frame can be used as a basis for gait signature.

8 absolute joint positions of each frame are used as feature points.

Page 19: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Extracting Absolute Joint Positions To extract absolute joint positions, the

corresponding positions are clicked in each frame.

Page 20: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Extracting Absolute Joint Positions

Page 21: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Layout of joint position feature matrix & feature vector

Feature Matrix

Feature Vector

Page 22: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Principle Component Analysis A person is identified by one feature vector. After PCA, we projected feature vector into

a feature space which gives the best level of recognition.

Page 23: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Spatio Temporal Correlation

The next step is to perform the recognition by pattern classification.

Algorithm:1. Each element of the class cluster one is compared

with the other class, and the distance is calculated. 2. The total distance between all corresponding class

elements are summed and a measure of the distance of the two classes is calculated.

3. The training class which has the smallest distance from the query cluster is chosen to be the class (i.e. person) which the query belongs to.

Page 24: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Spatio Temporal Correlation

Page 25: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Discussion This project recognise 7 person of 20

people. Restrictions: 1.The dataset that we have worked on is not

qualified.

Page 26: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Discussion Restrictions:2. We don’t have enough data for training

and test set.3. Any other advanced classification methods

can be applied rather than spatio temporal correlation

Page 27: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Abdelkader’s Eigengait Approach

Abdelkader’s eigengait approach of gait recognition is also a silhouette – based technique.

This technique creates self similarity matrices from the image sequences.

After creating self similarity matrices, the rows of these matrices are appended to form a single feature vector.

All the feature vectors are gathered together and PCA is applied to project the data into a new feature space which is called Eigengait.

Finally k-NN is applied to the Eigengait data for classification.

Page 28: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Abdelkader’s Eigengait Approach

Self Similarity Matrices are created by comparing the similarity of pixel intensities over N frames.

Ot1 and Ot2 are extracted images at time t1 and t2 respectively.

x and y values are representing the pixels of the image.

Bt1 is the minimum bounding box surrounding the extracted object.

Page 29: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Abdelkader’s Eigengait Approach

Self Similarity Plot

Page 30: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Abdelkader’s Eigengait Approach

Self Similarity Matrice Characteristics

Page 31: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Abdelkader’s Eigengait Approach

Calculate the k – nearest neighbor to the unclassified feature vector in the training set.

Determine the class which has the most points in the k selected points.

Page 32: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Discussion

SOTON Database will be used for the next experiments.(normalized, not noisyabout 10 instances for each class)

Page 33: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Discussion Abdelkader’s Eigengait Approach has % 25

identification rate on CASIA Database. The rate is very low because the dataset is not

sufficient for Eigengait approach. We used 1-NN classifier because we can create

only one self similarity matrix for each class. Data is not normalized according to the phases

and cycles which is very essential for sel similarity matrices.

Page 34: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Ozan’s “What if it happens?” approach 2 ideas coming together

◦ Using skeletons◦ Using Motion history images

Page 35: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

A picture is worth a thousand words

IF...

Page 36: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

What about a video?

Page 37: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

A little bit of results?

Pure Skeleton Skeleton + time

Pure Full Image Full image + time

Page 38: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

Why didn’t it happen?

Page 39: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

ComparisonAveraged Silhouette(Paper)

Averaged Silhouette(Impl.)

Absolute Joint Positions(Paper)

Absolute Joint Positions(Impl.)

Eigengait Approach(Paper)

Eigengait Approach(Impl.)

Identification Rate 79% 73% 55% 35% 93% 25%

Page 40: BIOMETRIC GAIT RECOGNITION CMPE 58Z INTRODUCTION TO BIOMETRICS TERM PROJECT

“Average Sillhouettes” Approach:1. “Simplest Representation Yet for Gait Recognition:

Averaged Silhouette” Zongyi Liu and Sudeep Sarkar “Absolute Joint Positions” Approach:1. “Gait Recognition using Absolute Joint Positions” Mark

Ruane Dawson “Abdelkader’s Eigengait” Approach1. “Motion-Based Recognition of People in EigenGait Space”

Chiraz Ben Abdelkader

REFERENCES: