Shanshan Chen, Christopher L. Cunningham , John Lach

Extracting Spatio-Temporal information from Inertial Body Sensor Networks

Shanshan Chen, Christopher L. Cunningham, John LachUVA Center for Wireless Health University of VirginiaBSN, 2011

Extracting Spatio-Temporal Information from Inertial Body Sensor Networks for Gait Speed Estimation1

Bradford C. Bennett,1Research Statement2Signal processing challenge to obtain accurate spatial information from inertial BSNsGait speed as an example to extract accurate spatio-temporal informationGait speed is the No. 1 predictor in frailty assessmentrequire high gait speed accuracydesire for continuous, longitudinal gait speed monitoringSo this research is motivated by two things, one is to overcome the challenges in signal processing for getting accurate spatial information, Another is simply because gait speed itself is an important parameter in gait analysis. Esp. now its been identified as no.1 predictor in frailty assessment in geriatrics.Predicting frailty with a couple of tens meters/s.And its an application require high accuracy and high resolutionAnd would benefit a lot from longitudinal continuous monitoring2Prevailing Technology --for Gait Speed EstimationNike+Pedometer, cadence

3Fit-Bit: Accelerometer, cadenceGarmin Forerunner 301Wearable wrist GPS, velocity Stopwatch and Tape

Before we jump into this research, lets look at the current devices on gait speed estimation.Nike+ provides a pedometer solution to assess cadence, and FitBit uses accelerometer for cadence.Both of the two solutions require a predefined calibration to get step length, And the accuracy remains questionable.

Garmin Forerunner provides a GPS solution, and gives an RMS of 0.05m/s in velocity assessment,But its limited to outdoor use only.

And dont forget, in clinic, we can always use stopwatch and tape if you dont have other fancy devices,Needless to say itll be limited to clinic use3Inertial BSN for Gait Speed Estimation4

TEMPO 3.1 inertial BSN platformdeveloped at the University of Virginia4Contributions5Refined human gait model by leveraging biomechanics knowledgeImprove accuracy without increasing signal processing complexityMounting calibration procedure to correct mounting errorPractical in experimentsImproved gait speed estimation accuracy by combining the two methods5Outline6Current Gait Speed Estimation MethodGait Cycle Extraction and Integration Drift CancelationStride Length Computation by Reference ModelRefined Human Gait ModelMounting CalibrationExperiment & Results6Gait Cycle & Integration Drift Cancelation

7Gyroscope signals on the sagittal planeUse foot on ground to find gait cycle boundariesNumerically easy to pick up local maximumHelpful for canceling integration driftShank angle is near zero and does not contribute to the stride length calculation when foot is on groundAssume linear drift

7Stride Length Computation

8Reference Model S. Miyazaki, Long-Term Unrestrained Measurement of Stride Length and Walking Velocity Utilizing a Piezoelectric Gyroscope

Then we compute stride length by this reference model, Where they assume human gait can be considered as symmetric single pendulum model.For example, when the shank swing backwards to the maximum, We obtain d1rs.Forward maximum, we get d2rs.By summing left shank step length and right shank step length we can get the stride length based on the trignometry provided here.

8Outline9Current Gait Speed Estimation MethodGait Cycle Extraction & Integration Drift CancelationStride Length Computation by Reference ModelRefined Human Gait ModelMounting CalibrationExperiments and Results9Inspection of Gait Phase10

By taking this 8 frames/second picture of many gaits, we found when the shank reaches the maximum, the leg isnt straight.1011

11Refined Compound Model12

Reference ModelBased on this observation, we propose this gait model, using only shank length as the hypotenuse at the backward swing, the forward swing remains the same.12Outline13Current Gait Speed Estimation MethodGait Cycle Extraction and Integration Drift CancelationStride Length Computation by Reference ModelRefined Human Gait ModelMounting CalibrationExperiment & Results13Mounting Calibration14Nodes could be rotated 20~30 from ideal orientationAttenuate the signal of interest on the sensitive axis

Ideal MountingNon-ideal Mounting14Mounting Calibration Methods15Validation of Mounting Calibration Algorithm16Mounting Position Rotated Around Y-axisMeasured by Proposed AlgorithmMeasurement Error of Angle0-0.0720.0721516.2861.2863027.8962.1044543.9541.0466058.0781.9227574.7370.2639090.4610.461Pendulum Model to simulate node rotation on shankRotate around z-axis with controlled degreeDetermine the rotation by Mounting Calibration AlgorithmAchieve an average error of ~1

Outline17Current Gait Speed Estimation MethodGait Cycle Extraction and Integration Drift CancelationStride Length Computation by reference modelRefined Human Gait ModelMounting CalibrationExperiment & ResultsTreadmill Control of SpeedIs gait on treadmill different from on ground?Gyroscope signals collected on treadmill show no significant difference from those collected on ground18

18Experiments on TreadmillTwo subjects, a taller male subject and a shorter female subjectTwo trials were conducted for each subject, one with well-mounted nodes and another with poorly-mounted nodes to validate mounting calibrationSpeeds ranging from 1 to 3 MPH with a 0.2 MPH (0.1m/s) increment for 45 seconds at each speed

19

Subject with poorly mounted Inertial BSN nodes performing mounting calibration on treadmillResultsBefore/After Mounting Calibration21Badly mounted nodes causes underestimation of gait speed attenuation of signal due to bad mountingMounting Calibration has correct the significant estimation error

Before Mounting CalibrationAfter Mounting Calibration21Results of Two Subjects22Significantly reduced RMSE compared to the reference model Overestimate at lower speeds and underestimate at higher speedsOverestimate taller subjects speeds more than the shorter subject

Gait Model at Different SpeedsThe thigh angle can be critical for controlling the step length

23Use thigh nodes to increase accuracy if invasiveness is not a concernHow accurate is accurate enough?Depends on application requirementHigh SpeedElimination of thigh angle results in underestimation of stride length at high speedVice versa at low speed

Results of Two Approaches24Double Pendulum at Initial Swing

Single Pendulum Model at Toe-offBetter than the reference modelStill overestimate the gait speed

Single Pendulum at Toe-OffFuture Work25Need more subjects, more gait types, and more gait speedsFor certain types of pathological gait, include those with shuffling, a wide base, and out-of-plane motionMore refined gait models will be developed based on biomechanical knowledgeEvaluate if a training set of data can be used to calibrate the algorithm for each individual subjectConclusion26Achieving an RMSE of 0.09m/s accuracy with a resolution of 0.1m/sProposed model shows significant improvement in accuracy compared to the reference modelMounting calibration corrected the estimation errorLeveraging biomechanical domain knowledge simplifies signal processing

Thanks! Q&A