Professional Documents
Culture Documents
Real Time Pakistani Sign Language Recognition Using Kinect
Real Time Pakistani Sign Language Recognition Using Kinect
Flow Diagram
Real time Pakistani Sign Language
Input Recognition using Kinect
Dataset
Model
Results
UML Diagram
National University of Computer and Emerging Sciences
FAST-NUCES Peshawar Campus
2 May 2018
Aziz Ullah (P14-6111), Najeeb Khan (P14-6112), Hafiz Saud Arshad (P14-6097)
1 / 29
Outline
1 Introduction
Introduction
Flow Diagram
2 Flow Diagram
Input
Dataset
3 Input
Model
Results
4 Dataset
UML Diagram
5 Model
6 Results
7 UML Diagram
2 / 29
Introduction
Introduction
Flow Diagram
3 / 29
Introduction
Dataset
4 / 29
Introduction Scope 1/2
Flow Diagram
Input
Dataset • The system can predict signs using only one hand.
Model • Only signing hand should be at the minimum distance
Results
from the Kinect.
UML Diagram
• Currently the system can predict four sentences of
Pakistan Sign Language.
• Can be easily extended by simply extending the dataset
and then train again.
5 / 29
Introduction
Scope 2/2
Flow Diagram
Input
Dataset
• Can be installed and reconfigured by using different
Model
dataset in different environments.
Results
UML Diagram
• In a hospital.
• In a training school.
• Social centers.
• In social centers for deaf and mute.
6 / 29
Innovation
Introduction
Flow Diagram
Input
• Firstly, there is very less work done in PSL recognition.
Dataset
Model
• The work done is mostly based on gestures(static signs);
Results
not words(dynamic signs) Recognition.
UML Diagram • We are working on words or sentences in PSL for the first
time.
• Many of the features used are thought by the team.
• Achieved got some significant results comparable to those
of simpler static signs.
7 / 29
Flow Diagram
Introduction
Time series
Flow Diagram Input Sign Segmentation Train
of features
Input
Dataset
Model
8 / 29
Input
UML Diagram
9 / 29
Input Transformation
Introduction
UML Diagram
• Finger tips and Center of the palm is identified using the
Skeleton.
• Following Features are extracted for training purpose
• Distance between two consecutive finger tips
• Angle of each convexity defect
• Minimum normal distance from convex hull to deepest
point of the defect.
• Area of the hand pixels.
10 / 29
Transformed Input Visualization
Introduction
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
Figur: Hand Segmented Figur: Hand Skeleton
11 / 29
Alternative Approaches 1/2
Introduction
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
12 / 29
Alternative Approaches 2/2
Introduction
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
13 / 29
Dataset
Introduction
Flow Diagram
Input
• We append the features extracted from all the frames until
Dataset the video ends in an Excel file.
Model • Dataset contains 30 instances of each sentence.
Results
• The dataset consists of four unique sentences.
UML Diagram
• All and only the videos of same sentence are kept in the
same directory.
14 / 29
Analysis of the Dataset
Introduction
Flow Diagram
• For an instance of the dataset
Input • The feature length of each frame is variable in the video
Dataset instance.
Model
• Depends upon the shape in the frame and the number of
Results
defects identified by the algorithm.
UML Diagram • For all instances in the dataset
• The Each instances of the same sign may differ in number
of frames.
• Depends upon variability of sign duration.
• For all instances of the same sign
• The frames of two instances can also differ in feature
length
• Depends upon the shape in the frame.
15 / 29
Dataset Sample 1/2
Introduction
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
16 / 29
Dataset Sample 2/2
Introduction
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
Flow Diagram
Dataset Padding
Input
Dataset
18 / 29
Model
Introduction
Flow Diagram • Now the model is trained using Random Forest Classifier.
Input
• Random Forest is an ensemble classifier that uses multiple
Dataset
Model
decision tree classifiers.
Results • 15 decision trees are used.
UML Diagram
• Each of Max-depth 8.
• Split criteria for each node of the tree is Entropy. Which
implies the maximum information gain.
19 / 29
Prediction using Model
Introduction
Flow Diagram • Take the sign from user, evaluate it against the model and
Input then predict the label.
Dataset
• Prediction label is selected based on the majority votes of
Model
Results
the decision trees.
UML Diagram • As depicted by the diagram.
20 / 29
Introduction
Flow Diagram
Results of Random Forest
Input
Dataset
• Cumulative Accuracy achieved is 82%.
Model
Results
• Cumulative F1-Score achieved is 0.83.
UML Diagram • The results of 10-cross folds is shown below.
• F1-Score
0.9142 0.9142 0.9142 0.8035 0.5761 0.8309 0.7416 0.8309 0.7321 1.0
• Accuracy
91.667 91.667 91.667 83.333 58.333 83.333 75.0 83.333 75.0 100.0
21 / 29
Results
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
22 / 29
Introduction Alternative Results using SVM Classifier
Flow Diagram
Input
• F1-Score
1.0 0.833 0.675 0.7428 0.5 0.5863 0.3653 0.675 0.7321 1.0
• Accuracy
100 83.333 66.667 75.00 50.0 58.333 41.666 66.333 75.0 100.0
23 / 29
Results of SVM Classifier
Introduction
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
24 / 29
Introduction Alternative Results using Decision Tree
Flow Diagram
Input
• F1-Score
1.0 0.833 0.675 0.7428 0.5 0.5863 0.3653 0.675 0.7321 1.0
• Accuracy
100 83.333 66.667 75.00 50.0 58.333 41.666 66.333 75.0 100.0
25 / 29
Introduction Alternative Results using Decision Tree
Flow Diagram
Input
Dataset
Model
Results
UML Diagram
26 / 29
Introduction Alternative Results using Naive Bayes
Flow Diagram
Input
• F1-Score
0.502 0.675 0.521 0.475 0.182 0.171 0.214 0.375 0.555 0.497
• Accuracy
50. 66.666 50. 50. 25. 25. 25. 41.666 75.0 58.333
27 / 29
UML Diagram
Input
Dataset
Model
Results
UML Diagram
28 / 29
Thank You
Introduction
Flow Diagram
References
• H. Cooper, E. Ong, N. Pugeault, R. Bowden, ”Sign Language Recognition using Sub-Units”,GU2
Input 9PY UK Journal of Machine Learning Research 13 (2012) 2205-2231
Dataset • N. Pugeault, R. Bowden, Centre for Vision, ”Spelling It Out: Real Time ASL Fingerspelling
Recognition ”, Centre for Vision, 6-13 Nov. 2011
Model • A. Kuznetsova, L.Leal-Taixe, B. Rosenhahn, ”Real-time sign language recognition using a consumer
Results depth camera”, Institute fuer Informationsverarbeitung, Leibniz University Hannover Appelstr. 9A,
Hannover, 30167, Germany
UML Diagram • N.Tanibata, N. Shimada, Y. Shirai, ”Extraction of Hand Features for Recognition of Sign Language
Words ”, 2002
• Z. Halim PhD & G. Abbas PhD, ”A Kinect-Based Sign Language Hand Gesture, Recognition System
for Hearing- and SpeechImpaired: A Pilot Study of Pakistani Sign Language ”, RESNA , 06 Oct
2014.
29 / 29