MARS: A Muscle Activity Recognition System Enabling Self-configuring Musculoskeletal Sensor...

21
MARS: A Muscle Activity Recognition System Enabling Self-con guring Musculoskeletal Sensor Networks IPSN 2013 NSLab study group 2013/06/17 Presented by: Yu-Ting 1

Transcript of MARS: A Muscle Activity Recognition System Enabling Self-configuring Musculoskeletal Sensor...

1

MARS: A Muscle Activity Recognition System Enabling Self-configuring Musculoskeletal Sensor

Networks

IPSN 2013

NSLab study group 2013/06/17Presented by: Yu-Ting

2

Outline

• Introduction• System Architecture• Evaluation• Conclusion

3

Motivation

• Correct motion & prevent injury– Non-intrusive– Scalable (autonomous setup)– Accurate

4

Disadvantage of Related Works

• Vision-based: LOS, clothing & skin cover• Needles: painful, low level activity• Larger sensors with contact gels:

low level activity

5

Sensing of Muscles

• Accelerometer– Tremors & oscillations: 3.85 Hz ~ 8.8 Hz– Internal vibration: 10 Hz ~ 40 Hz

6

System Overview

7

Outline

• Introduction• System Architecture• Evaluation• Conclusion

8

Sensor Node Network

• Provide error detection checksum• Anti-alias filter for the accelerometer• Wired to mobile data aggregator– SPI interface, 1Mbps– 10 Hr for 2200mAh battery

9

Mobile Data Aggregator

• On Yellow Jacket board– Support 6 sensors & 2.5 meters

• Receive data from all nodes by TDMA• Decode checksum• Reasons of errors– Damaged sensors– Out of sync nodes

• Postpone data sampling until the next cycle

• Wi-Fi to backend server

10

Backend Server – Muscle Activity Recognition

• 10Hz high pass filter: avoid signal from tremors• Feature extraction in Matlab using algorithms from WEKA

– 6 time domain features• RMS:

related to the intensity of an action• Cosine correlation:

relation of vibrations at different axes

– 15 frequency domain features• Apply DFT (Discrete Fourier Transform)• 3 information entropy of DFT magnitude• 3*4 bands PSD (Power Spectral Density)

– N sensors, M=21

• J48 decision tree classifier

11

Backend Server – Motion Tracking & Visualization

• Complimentary filter fusion of sensor data– Obtain accurate orientations of the sensors– By quaternion-based complimentary filter [19,25]

• Range of motion limitation• Visualization and rendering– Java & Unity Gaming Engine

12

Outline

• Introduction• System Architecture• Evaluation• Conclusion

13

Vibration Signature Feature Ranking

• Muscle vibrations are directional• Current MARS assume the orientation of

sensors doesn't change• Future MARS will try to use polar coordinates

14

Detection of Muscle Vibration

• PSD of accelerometer– Large difference in PSD– PSD is unique for different person

15

User Study

• 4 females & 6 males from different background• Isolated and compound muscles• Compare three classfiers

16

Precision & Recall

• Precision: positive predictive value• Recall: as sensitivity

17

Result of User Study – Isolation Type

18

Result of User Study – Compound Type

19

Outline

• Introduction• System Architecture• Evaluation• Conclusion

20

Conclusion

• Pros– Fine-grained muscle activity monitoring– Fast personalized system setup

• Sensors can be moved/changed afterwards

– Real time processing with visualization• Cons– Not convenient enough to wear the system– Need to be trained individually– The accuracy of the system may still vary with

placement

21

Q&A