Building the NINAPRO Database: A Resource for the Biorobotics Community
2 likes914 views
The document outlines the creation of the Ninapro database, a resource designed for the biorobotics community, focusing on semg data for controlled prosthetics. It details the acquisition setup, methods used for data collection from intact and amputated subjects, and analysis techniques including principal component analysis and classification performance metrics. Future work aims to establish benchmarks and expand the database with more subjects and movements.
Building the NINAPRO Database: A Resource for the Biorobotics Community
- 1. Building the NINAPRO Database: A Resource for the Biorobotics Community 1Manfredo Atzori, 2Arjan Gijsberts, 3Simone Heynen, 3Anne-Gabrielle Mittaz Hager, 4Olivier Deriaz, 5Patrick van der Smagt, 5Claudio Castellini, 2Barbara Caputo, and 1Henning Müller 1Dept. Business Information Systems, HES-SO Valais, Switzerland 2 Institute de Recherche Idiap, Switzerland 3 Department of Physical Therapy, HES-SO Valais, Switzerland 4 Institut de recherche en réadaptation, Suvacare, Switzerland 5 Institute of Robotics and Mechatronics, DLR (German Aerospace Centre), Germany
- 2. 1. Introduction: what is electromyography Electromyography (EMG) is the measurement of electrical activity that creates muscle contractions The signal path: • Originates in a motor neuron • Travels to the target muscle(s) • Starts a series of electrochemical changes that leads to an action potential • Is detected by one or more electrodes (Jessica Zarndt, The Muscle Physiology of Electromyography, UNLV) 2
- 3. 1. Introduction: electromyography controlled prosthetics • 2-3 degrees of freedom • Few programmed movements • Very coarse force control • No dexterous control • No natural Control • Long training times In contrast to recent advances in mechatronics 3
- 4. 1. Introduction: sEMG Data Bases • NO large scale public sEMG databases, only private ones (Fukuda, 2003; Tsuji 1993; Ferguson, 2002; Zecca, 2002; Chan, 2005; Sebelius, 2005; Castellini, 2008; Jiang, 2009; Tenore, 2009; Castellini, 2009) • NO common sEMG acquisition protocol • NO common sEMG storage protocol 4
- 5. 1. Introduction: project motivations & goals • Creation and refinement of the acquisition protocol • Acquisition of the database • Public release of the database • Worldwide test of classification algorithms • Augment dexterity of sEMG prostheses • Reduce training time 5
- 6. 2. Database: acquisition setup (1) Laptop: Dell Latitude E5520 ! Digital Acquisition Card: National Instruments 6023E sEMG Electrodes: 10 double-differential Otto Bock 13E200 ! Printed Circuit Board, Cables & Connectors ! Data Glove 22 sensors Cyberglove II (Cyberglove Systems) Inclinometer: Kübler 8.IS40.2341 6
- 7. 2. Database: acquisition setup (2) 1. 8 equally spaced electrodes 2. 2 electrodes on finger flexor and extensor muscles 3. Two axes inclinometer 4. Data glove 7
- 8. 3. Methods: acquisition procedure Intact subjects: • The subject is asked to repeat what is shown on the screen with the right hand. Amputated subjects: • The subject is asked to think to repeat what is shown on the screen with both hands. • In the meanwhile the subject needs to do the same movement with remaining hand. 8
- 9. 2. Database: movements Exercise 1 Hato, 2004 12 movements ! ! ! ! ! ! Sebelius, 2005 Farrel, 2008 ! ! ! ! ! ! Crawford, 2005 Feix, 2008 Exercise 2 17 movements ! ! ! ! ! ! ! ! DASH Score ! ! ! ! ! ! ! ! ! Exercise 3 23 movements ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! 9
- 10. 2. Database: data Data stored for each subject: • One XML file with clinical and experimental information • Unprocessed data (sEMG, Cyberglove, Inclinometer, Movie) • One preview picture for each exercise • One picture of the arm without the acquisition setup • One picture of the arm with the acquisition setup on Subjects: • Currently stored: 27 intact subjects • To be acquired: ~100 intact subjects ~40 amputated subjects 10
- 11. 2. Database: public, with web interface url: http://ninapro.hevs.ch 11
- 12. 3. Analysis: evaluation of the acquisition protocol • Principal Component Analysis data that is easily separable visually will often also be easy to classify • Classification idea of how discriminative the sEMG signals are for movements and subjects • Groups of subjects: 1, 8, 27 subject • Sets of movements: 3, 11, 52 movements 12
- 13. 3. Analysis: preprocessing 1. Synchronization: linear interpolation of all data at 100Hz 2. Filtering of sEMG signals: Butterworth, zero-phase, 1Hz, second order 3. Segmenting: each movement (including rest) is divided into three equal parts 4. The data contained in the central segment is averaged for each electrode 1 2 3 4 13
- 14. 3. Analysis: Principal Component Analysis Two principal components for each of the nine cases considered • Movements are easy to distinguish in cases with few subjects and few movements. • Overlap increases combining data from multiple subjects • Overlap increases increasing the number of movements. 14
- 15. 3. Analysis: Quantitative classification performance Intra-subject classification: • Multi-class LS-SVM with RBF kernel is trained for each subject • Training: 5 movement repetitions • Test: 5 movement repetitions • Experiment repeated 25 times with different random splits Inter-subject classification: • Multi-class LS-SVM with RBF kernel is trained for each subject • Training: 5 movement repetitions of one subject • Test: 5 movement repetitions of each of all the other subjects • Experiment repeated 25 times with different random splits 15
- 16. 3. Analysis: LS-SVM Results Intra-subject classification: • Errors from 7.5% to 20% • High standard deviation (performance variability among different subjects) Inter-subject classification: • Only marginally above chance level 16
- 17. 5. Conclusions: Database • Acquisition setup: portable, based on scientific research and industrial application needs • Acquisition protocol: complete and easy to be reproduced • Movements: 52, selected from the scientific literature • Data: currently 27 intact subjects are stored Data Analysis & Evaluation • PCA: movements are easy to distinguish in cases with few movements and few subjects • Intra-subject classification: results comparable to those found in the literature with the same number of movements • Inter-subject classification: classification slightly above chance level 17
- 18. 5. Future Work: • Establishing a standard benchmark • Collecting data from a large number of movements Add a custom-built force-sensing device to acquire dynamic finger/hand/wrist data. • Collecting data from a large number subjects Further releases of the database will contain data recorded from a larger number of subjects. 18
- 19. THANKS FOR THE ATTENTION Please, cite: Manfredo Atzori, Arjan Gijsberts, Simone Heynen, Anne-Gabrielle Mittaz Hager, Olivier Deriaz, Patrick Vand der Smagt, Claudio Castellini, Barbara Caputo and Henning Müller, Building the NINAPRO Database: A Resource for the Biorobotics Community, in: Proceedings of the IEEE International Conference on Biomedical Robotics and Biomechatronics, Rome, 2012 Full publication: http://publications.hevs.ch/index.php/publications/show/1172 For more information: http://www.idiap.ch/project/ninapro/ http://ninapro.hevs.ch Contacts: manfredo.atzori@hevs.ch