framework for hand grasps F. Stival M. Moro E. Pagello Learning - - PowerPoint PPT Presentation

INTELLIGENT AUTONOMOUS SYSTEMS LAB

A first approach to a taxonomy-based classification framework for hand grasps

Learning Applications for Intelligent Autonomous Robots – LAIAR 2018 Full-day IAS-15 workshop June 11, 2018

• F. Stival
• M. Moro
• E. Pagello

IAS-LAB

Introduction

Emulation of the human body behaviour

Interaction with robotic devices controlled by physiological human signals (sEMG)

Human intention of movement Coherent robot action

Alison Gibson, Mark Ison and Panagiotis Artemiadis, "User-independent hand motion classification with electromyography," ASME Dynamic Systems and Control Conference (DSCC), 2013

IAS-LAB

Strengths of the work

Binary classification based

• n a quantitative

taxonomy of hand grasps

Taxonomy-based classification

General model built on data from many different subjects

Subject independence

Continuous online estimation of the performed movement

Online elaboration

N O V E L T Y

IAS-LAB

Subject-independent framework

Goal Ready-to-use model: Good performances since first trials of a new user

Common underlying behavior in the task performed by different subjects Extraction of common constraints by looking to different interpretation of the task

IAS-LAB

Quantitative taxonomy of hand grasps

Hyerarchical binary

• rganization of 8 significative

hand grasps Considered both physical and physiological signals 40 subjects involved in the study – from NinaPro dataset

IAS-LAB

Taxonomy-based classification Decomposition of the classification phase between couples of movement groups The closest to the roots the more challenging the classification but less problematic miss-classification error (the two movements will be close to each other)

IAS-LAB

Actions sequence

Training set (offline)

Gaussian Mixture Model Multiple Subjects Processing

Test set (online)

New Subject Processing Taxonomy- based Gaussian Classification

IAS-LAB

Processing phase

EMG signals are non-stationary Analysis in both time and frequency Mother wavelet (db2) Synthesis values MAV =

1 𝑂

𝑦

𝑂 =1

Smoothing and Normalization Wavelet Transform

GMC New subject Processing Multiple subjects GMM Processing

IAS-LAB

Gaussian Mixture Gaussian Mixture Classification (GMC) Gaussian Mixture Model (GMM) 𝜂

ℎ = {𝜊ℎ 𝑢 , 𝛽ℎ𝑢} ∈ ℝ𝐸

Weighted sum of K Gaussian components

GMC New subject Processing Multiple subjects GMM Processing

𝑄 𝜂

ℎ = 𝜌𝒪 𝜂 ℎ; 𝜈, Σ 𝐿 =1

EMG values assumed by the channels at the instant t Angles assumed by the joints at the instant t Bayesian information criterion (BIC)

Class 1 Class 2 2 k 1

1.

𝝍𝟐

𝒎 first l samples

• f the signal

2 considered class k total samples

Selected the class corresponding to the largest computed value 𝑄𝐸𝐺 𝜓1

𝑑 , = 1. . , 𝑗 = 1. . 𝑜

• Columns normalization
• Rows averaging

IAS-LAB

Preliminary results Leave-One-Out approach on N-1 subjects and testing on new, unseen subject Tested on levels of increasing complexity

C O M P L E X I T Y

Movements of Level N are grouped together in Level N+1

IAS-LAB

Preliminary results Results averaged fixed the level

Good accuracy achieved: mean accuracy of 76.5%, reaching 97.29% in one of the higher levels

Low levels: harder to classify since the movements are strictly related

IAS-LAB

Conclusions

Preliminar study on a subject- independent Taxonomy-based classification solution computed

• nline using EMG signals

Exploitation of a quantitative taxonomy of hand grasps Based on Gaussian Mixture Model and Classification

IAS-LAB

Conclusions

Good accuracy achieved: mean accuracy of 76.5%, reaching 97.29% in one of the higher levels New classification algorithm

Future works:

Comparison of the accuracies: taxonomy vs. traditional approach

Low levels: harder to classify since they are more related

Thank you for the attention Any question?