kinect virtual-learning (sim u kin graduation project)
TRANSCRIPT
Kinect Virtual Learning
Team members: Mohamed Hesham Mahmoud Soliman Ahmed Amr Yousef Ahmed Ahmed Nasser Date: 5/7/2015
Agenda• Problem & Solution• Presentation Application • Interviewing Application • Pattern Recognition techniques• Speech Recognition • Demo• Questions ?
Problem• Why Soft Skills ?• Business & Education• Chance !!• Software !!
Existing Solutions• Expensive Courses
• Online Materials
Idea & Solution• Simulate the presentation & interview Processes • train the presenter & interviewee on the correct behavior
Body movements Speech recognition Facial Expression
• Feedback of weakness points
What is Kinect !
Message Delivery • Body Language• Speech • Facial Expression
Project Overview
Presentation Application
Detected Body MistakesReferences
Body Joints
Mistakes Categories
Dynamic Mistakes
Body Language
Body Language
Body Language
Body Language
Body Language
Body Language
Static Mistakes
Body Language
Body Language
Body Language Hand in Pocket
Leaning Body Language
Body Language Hand on waist
Body LanguageUp leg
Presentation Application Flow
Interview Application
References
Detected Face Mistakes
Facial Techniques Face Joints Face Orientation
Face Gestures Touching face and Yawning
Facial Techniques Face Joints Face Orientation
Face Gestures No Eye Contact
Face Gestures
Feeling Shy
Face Gestures Smiling
Face Gestures Cleaning Glasses
Face Gestures Aggressive
Face Gestures Bad Posture
Interview Application Flow
Pattern Recognition Techniques
Classifiers Categories
Rule-Based Classifiers • Conditions
Threshold
• 12 Joints with XYZ Coordinates
Hidden Markov Model Classifiers
• Dataset 50 sample for each mistake
• Models 9 states 200 Observations
• Phases Learning Decoding Evaluate
Rule-Based HMM80
82
84
86
88
90
92
94
96
ArmLegBody center
Classifier Arm Leg Center Type
Rule-Based 92% 88% 90% Static
HMM 95% 85% 93% Static – Dynamic
Results and Statistics
Classifiers Conclusion
• HMM is more accurate than Rule-based and support Dynamic states
• Rule-based is complex to detect specific threshold for different bodies
Gesture Detection Flow
• Not all features are relevant to all gestures.
Each gestures has its own feature vector.
• More than one Gesture can happen at the same time. We group related gestures together under respective limbs.
• Gestures can be related to more than one limb We divide gestures into parts called “states”.
Problems:
Gesture Detection Flow
Right
Arm
Hand Over Hand
Hand On Waist
Hand In PocketLeft Arm
Hand Over Hand
Hand On Waist
Hand In Pocket
Right Leg
Cross Leg
Up Leg
Left Leg
Cross Leg
Up Leg
body
Leaning left
Leaning right
Gesture Detection Flow
Flow :
• Each Gesture has a certain condition on the detected states.• Body consists of 5 Limbs.• Each Limb has a most probable state.• Each state has a classifier object that receives the feature vector.• Each state has its own feature vector.
Gesture Detection Flow
Gesture types :
• Static Gestures : No movement involved Happens when any state from a group of states happen. Happens when all states from a group of states happen.
• Dynamic Gestures : Requires the body to move Happens when a sequence of states happen in a short period of time.
Speech Recognition
Speech Recognition• Speech recognition process
Speech Recognition• Presentation application (Filler Words)
Fillers words and phrases people use to cover verbal gaps—are word crutches. Presenters often use them out of fear.
• The most common fillers are: So, And, All right, Okay, Like, Now, Well, You know,
Right, Um and Uh.
English Test phases
First Phase "Put Question"
Second Phase "Paragraph Test "
Third Phase “Knowledge "
Conclusion• HMM is more accurate than Rule-based Classifiers.• Kinect is the best device to use due to infrared feature.• Kinect V2 is better than Kinect V1 in Joint detection.• Kinect V2 has face Joints property over Kinect V1.
Future Work• Interviewing Enhancement• Try other classifiers seeking better accuracy• Provide the Grammar Builder with more alternatives
Sponsorship
Presentation Demo
Interview Demo