34th Annual International Conference of the IEEE EMBS San Diego, California USA, 28 August - 1 September, 2012 Gait Cycle Spectrogram Analysis using a Torso attached Inertial Sensor Mitchell Yuwono, Steven W. Su, Bruce D. Moulton, and Hung T. Nguyen Abstract-Measurement of gait parameters can provide important information about a person's health and safety. Automatic analysis of gait using kinematic sensors is a newly emerging area of research. We describe a new way to detect walking, and measure gait cadence, by using time-frequency signal processing together with spectrogram analysis of signals from a chest-worn inertial measurement unit (IMU). A pilot study of 11 participants suggests that this method is able to distinguish between walk and non-walk activities with up to 88.70% sensitivity and 97.70% speciicity. Limitations of the method include instability associated with manual ine-tuning of local and global threshold levels. I. INTRODUCTION The ability to measure walking activities can be important for monitoring certain medical conditions and assessing treatment eficacy [1]. For example, in a clinical study involving 488 chronic hert failure patients, Passantino reported strong corelation between the chance of survival and changes in the distance walked [2]. Walking and balance involves complex coordination of the limbs and torso [3]. Gait cadence (step rate) can be determined from small swings in torso angle in the sagittal plane due to the periodic shifting of moment of inertia that occurs on each phase [4]. Previous attempts to measure gait using spectra from torso-attached accelerometers include Brralon's estimation of spectral power magnitude using a Short-time Fourier transform (STFT) and Discrete Wavelet Transform (DWT) [4]. Brralon reported detection sensitivity of 78% and specificity of 68.7%. Another approach using accelerometer and wavelet decomposition was reported by Bidrgaddi to distinguish walking from other high impact activities with 89.14% sensitivity and 89.97% specificity [1]. This research aims to develop more-accurate gait cycle analysis for ambulatory monitoring systems, such as those we re working on at University of Technology Sydney Centre for Health Technologies [5]. This paper describes a new approach for processing and analysing signals from a chest-mounted IMU. Section II gives an overview of the hrdware. Section III describes the method. Section IV presents the data collection. Section V presents results and analysis,and Section VI provides conclusions. II. OVERVEW This pilot study used a Shimmer MEMS kinematic module with a 9DoF daughterboard. The 9DoF board has a Freescale M. Yuwono, S.W. Su, B. Moulton, and H. Nguyen are with the Faculty of Engineering and Information Technology, University of Technology, Sydney, Ultimo, 2007, NSW, Australia. (e-mail: mitchellyuwono@gmail.com). MMA7361 tri-axial accelerometer, a Honeywell HMC5843 magnetometer, and an InvenSense500 gyroscope. Data collection and an Attitude Heading Reference System (AHRS) is run externally in Java 2 Standrd Edition (J2SE) running a custom driver adapted rom original Shimmer driver, with 3-D visualization under jMonkeyEngine 3.0 [6]. The IMU samples at 50 Hz [7]. We do algorithmic prototyping in MATLAB. The device is strapped on the participant's chest in a way that the torso angle can be observed directly from the pitch measurement. During walking, the requency of torso-swing ranges from 0.6 to 2.5 Hz, [4,8],so we sample data for processing at 20Hz. III. METHOD The method includes three processes: torso angle estimation, time-frequency signal processing, and spectrogram image processing. First, the orientation quaternion q of the sensor is estimated using the explicit complementary filter (ECF) [9] applied to measurements of angular velocity o and acceleration a. Pitch information 6 is calculated rom the sensor orientation. The signal is then convolved with a digital band pass ilter (bp) with cutof frequencies of 0.5Hz and 5Hz to yield 6bp- Autocorrelation is used to minimize noisy signals on 6bpf• The Discrete Fourier Transform (DFT) is then be applied to the autocorrelated signal using Bartlett's method to extract the spectrogram S,t). Overlapped spectrograms re averaged. Spectrogram image processing techniques are then used,starting with smoothing using a Gaussian Filter. A Gabor filter and median-C thresholding re then used. This is done by applying a median ilter to the Gabor iltered image and subtracting the result from the normalized image to get the logical mask M2,t). A morphological ilter performing smoothing and erosion is then applied to the logical mask. Gait cycles such as wal/not-walk and cadence are then easily extracted. A block diagram is given in Figure 1. A. Torso Angle Estimation Torso angle is estimated using ECF applied to the information from the gyroscope and accelerometers [9]. Our initial tests suggest that the information provided by these two sensors is suficient to estimate torso angle. The output of the rate gyroscope o and normalized accelerometer reading G can be represented in vector form as in (1) and (2). The symbol A denotes a (1-norm) unit vector. )= [0 )x ) y )zf (1) G = [0 ax Gy G zf (2) 978-1-4577-1787-1/12/$26.00 ©2012 IEEE 6539