Presentation is loading. Please wait.

Presentation is loading. Please wait.

Estimation and Classification of Human Movement Using 3 Axis Accelerometers Eric Cope Advisors: Dr. Antonia Papandreou-Suppappola Dr. Bahar Jalali-Farahani.

Similar presentations

Presentation on theme: "Estimation and Classification of Human Movement Using 3 Axis Accelerometers Eric Cope Advisors: Dr. Antonia Papandreou-Suppappola Dr. Bahar Jalali-Farahani."— Presentation transcript:

1 Estimation and Classification of Human Movement Using 3 Axis Accelerometers Eric Cope Advisors: Dr. Antonia Papandreou-Suppappola Dr. Bahar Jalali-Farahani March 30, 2009

2 2 Motivation

3 3 Qualifier's Summary Brief Background Human Physiology Sensor Technology - Accelerometers Formulation of Human Movement using Accelerometer (gravity, movement, noise) Solutions for two models using Kalman Filtering Simulations Future Work

4 4 About Me: Eric Cope Education BSE – Electrical Engineering – ASU – 2004 Focus: Analog Circuits, DSP, RF MSE – Electrical Engineering – ASU – 2006 Focus: Analog Circuits, DSP PhD – Electrical Engineering – ASU Focus: DSP and VLSI Implementation

5 5 About Me: Eric Cope Profession Medtronic – Sensors Manufacturing Intern – Product Development IC Design Intern – IC Design Engineer (PD) – Senior IC Design Engineer (PD) 2008 – Senior IC Design Engineer – Digital Technologies

6 6 The Physiology of Human Movement States Walking / Running Standing / leaning Sitting Slouching, leaning forward) Lying Down Propped Up Stomach, side, back Transitory States Standing to Sitting Sitting to Lying Down Standing to Lying Down (falling)

7 7 Medical Implications of Human Movement Quality of Life Measurement Disease Detection Heart Failure Fall Detection – AMI, Syncope Activity Detection / Estimation Objective Measurement of Activity Obesity Impact

8 8 Sensor Technology and Their Benefits and Costs

9 9 Why Accelerometers in Implantable Medical Devices? Low Power - <200nA Cheap MEMS technology enables mass production CMOS technology allows calibration of low reproducibility processing -> easy to manufacture Low Processing Needs Piggybacking other medical device needs

10 10 Types of Accelerations Linear Acceleration w.r.t. to direction vector Ex: a runner accelerating in a straight line Angular Acceleration w.r.t. to direction vector As an object rotates around a point, it is experiencing an acceleration always pointing to the point about which it is rotating Ex: Planetary motion Theta is the time-varying angle of the circular direction

11 11 Types of Acceleration Gravity Pulls bodies towards one another Amplitude depends on the masses of the bodies Earth's gravitational pull is 9.81m/s 2 Forward Thinking: How do we Differentiate between these types of accelerations?

12 12 These Accelerations as Experienced by the Human Body Linear Gravity, standing to walking Angular Bending over to pick up a pencil Spinning like a top Dancing

13 13 But, What Else Does the Sensor Experience? Offset Mechanical Changes Drift in Circuit Performance Noise EMI – AWGN Narrowband (60 Hz) and broadband (RF radiation) Muscle Spasms – AWGN bandpass noise pulses Voices – broadband bandpass Cross-Axis Contamination - nonlinear (strong sensor characterization needed) Circuit Noise – AWGN broadband - well modeled and understood

14 14 Frames of Reference Global Frame of Reference Gravity always points in -Z direction The sensor is fixed with respect to the Earth Ex: Needle of a compass Physiological Frame of Reference The sensor is always aligned with the Patient Gravity can point anywhere

15 15 Current Published Research Two Groups (1) Heavy Emphasis on Biologics, Light Emphasis on DSP Lots of light post processing: low pass filtering with lots of tweaking to obtain data per a particular sample set Lots of Sensors: Magnetometers, gyroscopes, accelerometers, well powered externally Large majority of the papers found lie in this category

16 16 Current Published Applications Gesture Movement Detection – Wii Athletic Optimizations Adaptive Noise Canceling of ECG Signals Human Movement Knee Unlock – Falling Monitoring Heart Movement – HF Rate Response

17 17 Current Published Research (2) Heavy Emphasis on DSP, Light Emphasis on Biologics Intense complex processing No direct application Ex: sensor fusion techniques not applicable to the field Current Methods Simple Processing Simple filtering Thresholding Neural Networks Adaptive Filtering Kalman Filtering

18 18 Published Example of Kalman Filtering of 3-Axis Accelerometers P. Veltink et-al were processing a 3-Axis Accelerometers data stream using Kalman filtering to establish an inclination measurement Inclination is the difference between the global frame of reference and sensor (or patient) frame of reference ARMA Acceleration Modeling, Kalman Filtering of Estimation Errors, Autocalibration of Offset Error Estimation Their application was an external application, however, it had potential to work in an implantable mode H. J. Luinge and P. H. Veltink, \Inclination measurement of human movement using a 3-D accelerometer with autocalibration," IEEE Transactions on Neural Systems and Rehabilitation Engineering, pp. 112{121, 2004.

19 19 Overview of Kalman Filtering: Predict The optimal solution is when state space equations are linear and noise and modeling errors are Gaussian Prediction: Predicted Estimate Covariance:

20 20 Residual (or Innovation): Innovation Covariance: Optimal Kalman Gain Overview of Kalman Filtering: Update Updated State Estimate Updated State Covariance

21 21 The Gravity Acceleration Model Observation a k is the linear and angular accelerations experienced due to physiological movement g k is gravity b k is the offset (b k = b k-1 – ) ( is a constant) is the noise with potentially time varying covariance, A z k is a 3x1 vector of Cartesian coordinates The unknown states are a k, g k, and b k Its very complicated because all three are unknown

22 22 Modeling Options Case 1: Simplified Model Just Gravity with a simplified prediction model x(k) = x(k-1) Case 2: Linear Extrapolation Model Just Gravity linearly extrapolated from past two estimates Slope between x(k-1) and x(k-2) is equal to slope between x(k) and x(k-1)

23 23 Case 1 Model

24 24 Case 2 Model Acceleration, offset and noise were combined for this model

25 25 Simulation Results Gravity Generated test data from polar coordinates Converted test data to Cartesian coordinates Modeling Errors Added AWGN with SNR ranging from 0 – 60 dB A small constant offset was added as well Accelerations were added by varying theta and phi Q = Modeling error constant Varied modeling error to investigate the modeling error effects

26 26 X Component - Model 1 0dB15dB 30dB 45dB60dB

27 27 Gravity X Component - Model 2 0dB 15dB 30dB 45dB 60dB

28 28 Impact of Offset vs. Modeling Error When the SNR is high, the offset becomes the dominating error = 1/(1,000)

29 29 MSE Plots Comparing Models

30 30 Modeling Error vs. MSE – Case 1

31 31 Work Conclusion An accelerometer can feasibly be used to estimate physiological human motion For complex estimates, a Kalman filter may a feasible method to estimate fine physiological states like slouching A more accurate model may be needed (and is in development) Other sensors like gyroscopes and magnetometers are unnecessary

32 32 Future Work More Accurate Models Use more accurate physics in modeling movement Model Depth – (i.e. FIR Filter) Determine Linearity of Signals and Distribution of Noise If model is nonlinear, a Particle Filter is a viable option Synthesizable RTL Implementation Low Power Architectures for Implantable Systems

33 Thank You Estimation and Classification of Human Movement Using 3 Axis Accelerometers

Download ppt "Estimation and Classification of Human Movement Using 3 Axis Accelerometers Eric Cope Advisors: Dr. Antonia Papandreou-Suppappola Dr. Bahar Jalali-Farahani."

Similar presentations

Ads by Google