coherence in one-shot gesture recognition for human-robot ...shapiro/ee562/notes/robotics.pdfgesture...

26
Coherence in One - Shot Gesture Recognition for Human - Robot Interaction Maru Cabrera December 3 rd 2018 12/3/2018 AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI - MARU CABRERA 1

Upload: others

Post on 22-Sep-2020

8 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Coherence in One-Shot Gesture Recognition for Human-Robot Interaction

Maru Cabrera

December 3rd 2018

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA1

Page 2: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Relevance of Gestures in HRI

vs

โ– Human have the unique ability to quickly adjust their context and learn from very few examples.

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA2

Page 3: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Motivation for One-Shot Gesture Recognition

โ– Lack of a comprehensive method that generalizes gesture recognition from few observations.

โ–We focus on the process used to generate a given gesture:โ–Cognitionโ–Learning and Generalizationโ–Physical execution

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA3

Page 4: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Motivation for Coherence in Gesture Recognition

โ– Explore gesture recognition and understanding when the roles between performer and listener are exchanged.

โ– Including the human aspect within the framework to artificially generate examples.

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA4

Page 5: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Overview of One-Shot Learning Framework

Skeleton Info Gesture exampleโ€œGist of

Gestureโ€Generated Dataset

Train Classifier

HMM

SVM

DTW

CRF

Artificial Generation Methods

Kinect Sensor

Performance Metrics

- Recognition accuracy- Efficiency- Coherency

Forward Method

Backward Method

1

2

3

4

5 6

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA5

Page 6: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Extracting the Gist of the Gesture

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA6

เตŸ๐’™๐’’ โˆˆ {๐’™1 โˆช ๐ผ๐‘ƒ๐ท โˆฉ ๐ผ๐‘ƒ๐ถ โˆช ๐’™๐‘ฏ, ๐‘ž = 1,โ€ฆ , ๐‘„

2 โ‰ค ๐‘„ โ‰ค ๐ป

Given ๐‘”1๐‘– = {(๐‘ฅ1, ๐‘ฆ1, ๐‘ง1 ), . . . , (๐‘ฅ๐ป, ๐‘ฆ๐ป , ๐‘ง๐ป)}

Set of inflection points

๐’™๐’’ = (๐‘ฅ๐‘ž , ๐‘ฆ๐‘ž , ๐‘ง๐‘ž)

Working Hypothesis:

Compact amount of information stored during cognitive processes of

gesture perception

Page 7: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Validating Extracted Gist of the Gesture

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA7

17 Participants

EEG Power Dynamic wavelets

Motion EEG

Correlation

Page 8: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Validating Extracted Gist of the Gesture

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA8

Page 9: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Validating Extracted Gist of the Gesture

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA9

Page 10: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Overview of One-Shot Learning Framework

Skeleton Info Gesture exampleโ€œGist of

Gestureโ€Generated Dataset

Train Classifier

HMM

SVM

DTW

CRF

Artificial Generation Methods

Kinect Sensor

Performance Metrics

- Recognition accuracy- Efficiency- Coherency

Forward Method

Backward Method

1

2

3

4

5 6

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA

Page 11: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Artificial Gesture Generation โ€“Forward Approach

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA11

Input:๐’ˆ๐Ÿ๐’” โ€“ 3D hand trajectory of a gesture of class s๐’™๐’” = (๐’™๐’”,๐’š๐’”,๐’›๐’”) โ€“ 3D position of the shoulder

K โ€“ Number of artificial trajectories to generate

Variance Estimation based on ๐’™๐’” and Gaussian Mixture Model (GMM)

Output:

เตŸ๐‘ฎ๐’” = { เทœ๐‘”1๐’”, เทœ๐‘”2

๐’”, โ€ฆ เทœ๐‘”๐พ๐’” โ€“ Set of artificial trajectories

for gesture class s

เทœ๐‘”๐‘˜๐‘– = ๐’œ๐‘ก

เทจ๐บ๐‘– ๐‘˜ = 1,โ€ฆ , ๐พ ; ๐‘– = 1,โ€ฆ , ๐‘ ; ๐‘ก = 1

เตŸ๐‘ฎ๐‘– = { เทœ๐‘”1๐‘– , เทœ๐‘”2

๐‘– , โ€ฆ เทœ๐‘”๐‘˜๐‘– , โ€ฆ , เทœ๐‘”๐‘

๐‘–

Page 12: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Artificial Gesture Generation โ€“Backward Approach

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA12

Page 13: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Artificial Gesture Generation โ€“Backward Approach

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA13

Jerk Minimization

Smooth changes in joint space usingthe third derivative of the joint angle

Least Energy Expenditure

Changes in joint space using torqueto calculate economic trajectories

๐‘†๐ฝ๐‘’๐‘Ÿ๐‘˜ = min

๐‘–=1

๐พ

เถฑ๐‘ก1

๐‘ก2แˆธ๐œƒ๐‘– +โ‹ฏ+min

๐‘–=1

๐พ

เถฑ๐‘ก๐‘„โˆ’1

๐‘ก๐‘„แˆธ๐œƒ๐‘– ๐‘†๐ธ๐‘›๐‘’๐‘Ÿ๐‘”๐‘ฆ = min

๐‘–=1

๐พ

เถฑ๐‘ก1

๐‘ก2

๐œ๐‘– ร— แˆถ๐œƒ๐‘– +โ‹ฏ+

๐‘–=1

๐พ

เถฑ๐‘ก๐‘„โˆ’1

๐‘ก๐‘„

๐œ๐‘– ร— แˆถ๐œƒ๐‘–

Begins in one IK solution for human arm (k= 7) for one IP ๐“ˆ๐‘ž๐‘ฃ, ends in

solution for different IP ๐“ˆ๐‘ž+1๐‘ค

๐“ˆ๐‘ž๐‘ฃ = ๐œƒ1

๐‘ฃ, ๐œƒ2๐‘ฃ, โ€ฆ , ๐œƒ๐พ

๐‘ฃ

๐“ˆ๐‘ž+1๐‘ค = ๐œƒ1

๐‘ค, ๐œƒ2๐‘ค, โ€ฆ , ๐œƒ๐พ

๐‘ค

Initial

Final

๐‘‰๐‘ž solutions in IP ๐‘ž of gesture trajectory

๐‘Š๐‘ž+1 solutions in IP ๐‘ž + 1

Page 14: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Artificial Gesture Generation โ€“Combined F+B Approach

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA14

Combination at gesture instance level.

Artificially generated observations from both approaches used as training data

Page 15: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Overview of One-Shot Learning Framework

Skeleton Info Gesture exampleโ€œGist of

Gestureโ€Generated Dataset

Train Classifier

HMM

SVM

DTW

CRF

Artificial Generation Methods

Kinect Sensor

Performance Metrics

- Recognition accuracy- Efficiency- Coherency

Forward Method

Backward Method

1

2

3

4

5 6

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA

Page 16: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Gesture LexiconMSRC-12

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA16

Contains sequences of human movements (Kinect skeleton)

12 iconic and metaphoric gestures

โ—ฆ Gaming commands and media player

Lexicon reduced to 8 gesturesโ—ฆ Excluded gestures with leg

motions or whole upper body.

Fothergill, S., Mentis, H., Kohli, P., & Nowozin, S. (2012). Instructing people for training gestural interactive systems. In Proceedings of the SIGCHI Conference (pp. 1737-1746). ACM.

Page 17: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Classification Algorithms

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA17

Hidden Markov Models (HMM):

โœ“ One-vs-all scheme

โœ“ 5 states left-to-right

โœ“ Baum-Welch algorithm

Support Vector Machines (SVM):

โœ“ One-vs-all scheme

โœ“ Radial Basis Function (RBF) Kernel

โœ“ MATLABยฎ library

Conditional Random Fields (CRF):

โœ“ Multi-class scheme

โœ“ Samples encoded using BIO: Beginning, Inside, Outside

โœ“ CRF++ toolkit

Dynamic Time Warping (DTW):

โœ“ Multi-class scheme

โœ“ Gesture Recognition Toolkit (GRT)

Page 18: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Training and Features

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA18

1 gesture instance per class was

used to extract inflection points

200 gesture examples per class

generated for training

100 gesture instances per class

from each data set for testing

Classifiers trained and compared

in terms of accuracy (๐ด๐‘๐‘%), and

recognition coherence (๐›พ)

Feature vector* ๐‘”๐‘– for training:

๐‘”๐‘– = {๐’™๐Ÿ = ๐‘ฅ1, ๐‘ฆ1, ๐‘ง1 , โ€ฆ , ๐’™๐‘ฏ}

๐‘”๐‘– = แˆถ๐’™๐Ÿ ๐œท๐Ÿโ€ฆ แˆถ๐’™๐’‹ ๐œท๐’‹โ€ฆ แˆถ๐’™๐‘ฏโˆ’๐Ÿ ๐œท๐‘ฏโˆ’๐Ÿ

แˆถ๐’™๐’‹ = ๐’™๐’‹+๐Ÿ โˆ’ ๐’™๐’‹ ๐œท๐’‹ = tanโˆ’1๐‘ฆ๐‘—

๐‘ฅ๐‘—, tanโˆ’1

๐‘ง๐‘—

๐‘ฆ๐‘—, tanโˆ’1

๐‘ง๐‘—

๐‘ฅ๐‘—

* M. G. Jacob and J. P. Wachs, โ€œContext-based hand gesture recognition for the

operating room,โ€ Pattern Recognition Letters, vol. 36, pp. 196โ€“203, Jan. 2014.

Page 19: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Overview of One-Shot Learning Framework

Skeleton Info Gesture exampleโ€œGist of

Gestureโ€Generated Dataset

Train Classifier

HMM

SVM

DTW

CRF

Artificial Generation Methods

Kinect Sensor

Performance Metrics

- Recognition accuracy- Efficiency- Coherency

Forward Method

Backward Method

1

2

3

4

5 6

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA

Page 20: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Performance Metrics

Coherence (g) is defined asthe intersection betweenthe agreement indices (AIx)for humans and machines,whether each agentcorrectly recognized eachgesture or not. ๐›พ =

๐ด๐ผ๐‘ฅ๐‘€ โˆฉ ๐ด๐ผ๐‘ฅ๐ป

๐ด๐ผ๐‘ฅ๐ป

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA20

๐ด% =๐‘ก๐‘œ๐‘ก๐‘Ž๐‘™๐‘ก๐‘Ÿ๐‘ข๐‘’โˆ’โ„Ž๐‘–๐‘ก๐‘ ๐‘ก๐‘œ๐‘ก๐‘Ž๐‘™๐‘ ๐‘Ž๐‘š๐‘๐‘™๐‘’๐‘ 

Page 21: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Accuracy Results

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA21

Forward and Backward approach for One-Shot Gesture Learning compared.

K-fold cross validation scheme with k = 10.

10 ๐ด๐‘๐‘% values for m and s for each approach.

Page 22: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Coherence Experiment Setting

โ– Two scenarios were explored with Baxter performing artificially generated gestures:โ– Scenario 1 (MH): Gestures are

recognized by 10 human participants.

โ– Scenario 2 (MM): Gestures are recognized by 4 classification algorithms.

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA22

Page 23: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Coherence Experiment Setting

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA23

Page 24: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Experimental Results

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA24

Page 25: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Takeaway Messagesโ– In its application to one-shot learning, the proposed methodhighlights the use of context for gesture recognition from the wayhumans use their bodies.

โ– The obtained results show the performance of the method,demonstrating independence from the selected classification strategy.

โ– The robotic implementation opens a different route towardscoherence in humanโ€“robot interaction.

โ– Coherence can be related to gesture classification when humans andmachines interchange the roles of performing and recognizing agesture.

โ– The calculated coherence metric is our main indicator that thegenerated gestures capture human-like variations for all the gestureclasses.

12/3/2018AI FOR ENGINEERS: COHERENCE IN GESTURE RECOGNITION FOR HRI -

MARU CABRERA25

Page 26: Coherence in One-Shot Gesture Recognition for Human-Robot ...shapiro/EE562/notes/Robotics.pdfGesture Recognition Explore gesture recognition and understanding when the roles between

Maru Cabrera

[email protected]

Coherence in One-Shot Gesture Recognition for Human-Robot Interaction