You are on page 1of 3

# MARKER LESS VIEW INDEPENDENT GAIT ANALYSIS USING DFT

## Prof L.V. Chaitanya Srinivas

Dept. of Biomedical Engineering, VIT University, Vellore, Tamil Nadu, India. E-mail: Chaitanya.lv@vit.ac.in

Abstract-We present a new method for viewpoint independent markerless gait analysis employing Discrete Fourier Transform (DFT) analysis. The system uses a web camera, works with a wide range of directions of walking. These properties make the proposed method particularly suitable for identification by gait. Tests on video sequences with subjects walking freely along different walking directions have been performed. This method shows that markerless gait analysis does not depends on any of internal or external camera parameters and obtained data that can be used for gait biometrics purposes.

z(n)=ankle1(n)+j*ankle2(n), where n is the sample number. Once the complex data is available we can take the FFT of z (n), Z(k) = n z(n) exp(-j2 kn/N) where N=101 and n=0, 1, .., N-1. The discrete Fourier transform (DFT) is used as the most popular tool to obtain a spectral representation of a given physiological signal. The discrete Fourier transforms pair that can be applied to sampled signals can be expressed as: Z(k) = n z(n)exp( j2 kn / N) --- (1) and Z(n) = (1/N) k Z(k)exp(j2 kn/N) --- (2) For n = 0,1,, N 1; k = 0,1,, N 1. where Z(k) represents the frequency domain function and Z(n) the time domain function. Utilizing these pair of formulas we can progress back and forth between a time representation of our complex joint angle signals (z(n)) and its frequency domain representation (Z(k)), that is, the DFT is invertible. Modification of the frequency spectrum is also possible to change the time representation. In our analysis both Z( k) and z (n) are complex series. In the case of Z(k), each value of Z(k) contains information on the complex modulus, or magnitude, and the phase in its real (R (k)) and imaginary (I(k)) parts. These, the magnitude function (|Z (k)|) and phase ( (k)) can be expressed as: | Z(k)| = (R2(k)+ I2(k)) --- (3)

INTRODUCTION

Gait recognition, which identifies people based on the way they walk or run, has recently gained growing research interest. The proposed method is a 2Dgait biometric approach, in which a single camera (webcam) is used. Camera is kept perpendicular to subject walking direction (sagittal plane). Current gait recognition approaches may be explicitly classified into two main classes: model-based approaches and appearance-based approaches. Model-based approaches generally aim to recover gait features, such as stride dimensions, limb lengths and kinematics of joint angles, by model matching in each frame of a walking sequence. Model-based approaches are insensitive to view-point and occlusions. Appearance-based approaches can be further divided into two classes: state-space methods and spatio-temporal methods. The former recognizes gait motion by considering temporal variations of observations with respect to the static poses of each frame. The latter generally characterizes the spatio-temporal distribution generated by gait motion in its continuum. A gait recognition method based upon spatiotemporal silhouette analysis. In this paper we describe a method based on Discrete Fourier transform (DFT) which allows us to identify the person based on gait. II METHOD

(k) = tan-1(I(k)/R(k)) --- (4) and the square of the magnitude function, the power spectral density (PSD) can be defined for periodic signals as: P(k) = |Z(k)|2 --- (5)

Data analysis: For our data analysis we considered joint angles: knee1-knee2, ankle1-ankle2, middle points and edge points. As we know human gait is cyclic and repeats itself, first we have to extract one full gait cycle for all angle data. Here the data are considered as complex data as for example
___________________________________ 978-1-4244 -8679-3/11/\$26.00 2011 IEEE

11

III SILHOUETTE REPRESENTATION Given a fixed camera, the human silhouette can be extracted by background subtraction and thresholding (Fig 1). And then its outer contour can be easily obtained using a border following algorithm based on connectivity. The contour can be further viewed as being in the complex plane (Fig. 2). The origin of the complex plane is set as the centroid of the contour. Each point on the contour can be represented by a complex number. si = xi + j yi, (i = 0, 1, ,N 1),
Fig 3: skeletal image

where N is the number of contour points. We use the same number of points to express each contour. Edge points (Fig. 4), middle points(Fig. 5) and skeleton image (Fig. 3) are shown below.

## Fig 1: human silhouette

Data acquisition From the contour image of the silhouette, we obtain edge point, middle point data in the form of Cartesian coordinates. Angle data is obtained from skeletal image. All these points lies on and inside the contour. Based on these points we gait analysis is done. From the skeletal image, we acquire angle data like both left and right knee and ankle angles. These angles are considered as complex data. For example Z(t) = L.knee-angle + i* R.knee-angle This data is transformed into frequency domain. Magnitude, phase and power spectral density data are obtained from the equations (3,4 & 5). From Edge point image, Cartesian coordinates of points on the contour are obtained. These points again are considered as complex data. Later, middle points of the contour are also considered as complex data. Information from various view independent video sequences are drawn. All this information is used to train the neural network.
Fig 5: middle points

## Designing Neural Network

12

Fig : Error reduction graph plotted for 12 training vectors and 5 test vectors Fig : NN block diagram

Main aim of using NN in this project is to classify different abnormalities. NN will be fed with different inputs. Finally, it has to classify the abnormalities. Classification will be based on knee angle, hip angle, ankle angle, cadence, stride length, frequency domain parameters, middle points of the contour and points on edge. Performance of a NN depends on number of inputs given. As number of inputs increases, error gradually decreases .

designing more sophisticated classifiers, algorithm robustness analysis deserve more attention in future work. These graphs demonstrate reduction in error when NN trained with different inputs of healthy condition and Parkinson's disease. REFERENCES
[1]. [2]. [3]. [4]. [5]. www.am.sanken.osaka.u.ac.jp/GaitDB/index.html www.clinicalgaitanalysis.com/data/index.html.su www.physionet.org/physiobank/database/#gait www.cc.gatech.edu/cpl/projects/hid. Gait analysis for human identification in frequency domain Shiqi Yu, Liang Wang, Weiming Hu, and Tieniu Tan, National Laboratory of Pattern recognition(NLPR), Institute of Automation,2007. Abnormal gait detection using discrete fourier transform Ahmed Mostayed1, Mohammad Mynuddin Gani Mazumder2, Sikyung Kim3 and Se Jin Park4 Korean Research Institute of Standards and Science (KRISS) ,2010. Markerless View Independent Gait Analysis with Self-camera Calibration Michela goffredo, Richard D. Seely, John N. Carter and Mark S. Nixon ECS, University of Southampton (UK), 2007 S. Corazza, L. Mndermann, A. M. Chaudhari, T. Demattio, C. Cobelli, and T. P. Andriacchi. A markerless motion capture system to study musculoskeletal biomechanics: Visual hull and simulated annealing pproach. ABME, 6:1019 1029, 2006. W. T. Dempster and G. R. L. Gaughran. Properties of body segments based on size and weight. Am J Anat, 120:3354, 1965. [8] D. A. Forsyth, O. Arikan, L. Ikemoto, J. O Brien, and D. Ramanan. Computational studies of human motion: part 1, tracking and motion synthesis. J FTCGV, 1:77254, 2005. M. Goffredo, N. Spencer, D. Pearce, J. N. Carter, and M. S. Nixon. Human Perambulation as a Self Calibrating Biometric., volume 4778 of LNCS. Springer, 2007. Rose SA, Deluca PA, Davis RB. Kinematic and kinetic evaluation of the ankle after lengthening of the gastrocnemius fascia in children with cerebral palsy,J Pediatr Orthop 1993,13:pp. 72732. Ounpuu S, Bell KJ, Davis RB 3rd, DeLuca PA. An evaluation of the posterior leaf spring orthosis using joint kinematics and kinetics,J Pediatr Orthop, 1996,16:pp. 37884. 13. Deluzio KD, Wyss UP, Zee B, Costigan PA, Sorbie C. Principal component models of knee kinematics and kinetics: normal vs. pathological gait patterns,Hum Movement Science, 1997,16: pp. 2

[6].

[7].

[8].

Fig : Error reduction graph plotted for 5 training vectors and 5 test vector

[9].

IV CONCLUSION AND FUTURE WORK In this paper an approach for gait detection employing DFT analysis has been presented. In other words we performed gait analysis in Fourier domain. The results indicate that the joint angle combination in the frequency domain is more efficient than other methods. Besides, edge points and middle points are also used. This makes the proposed method more accurate. Our presented method is view dependent. Analysis of recognition performance with view angle changes would be done. Also, seeking better similarity measures,
[10].

[11].

[12].

[13].

13