ABNORMAL GAIT CLASSIFICATION USING SILHOUETTES S. M. H. Sithi Shameem Fathima and R. S. D. Wahida Banu 2 Syed Ammal Engineering College, Ramanathapuram, Tamilnadu, India 2 Government College of Engineering, Salem, Tamilnadu, India E-Mail: samimnasu@gmail.om ABSTRACT This paper proposes a new methodology to lassify the person with normal walk or abnormal walk for surveillane purposes. Reognizing human walk is emerging as a ritially important biometris, hallenging omputer vision problem. However, the inlusion of abnormal gait dataset with normal gait databases has to be very useful to lassify the normal and abnormal walking style of a person. The silhouettes are trained and tested with K nearest neighbor lassifier. We introdue a more hallenging abnormal walk patterns like Antalgi gait, Charlie haplin gait, steppage gait, sissor gait, irumdution gait, inlusive with normal gait data base. The database onsists of about 5000 frames with 5 different walk styles. Manual seletion of persons with different walking styles resulted in high degree of variability in pose and illumination. The method starts with the extration of human silhouettes from input videos. Initially the ontinuous input videos are onverted into frame-by-frame by means of onversion algorithm. Eah frame onsists of noises and shadows. Then silhouettes are removed from noises and disontinuities to produe an abnormal gait database. From the gait data base, parameters are measured by segmenting into six portions from head to nek, nek to torso, hip to knee of both right and left leg, knee to toe of both legs, height of the blob and width has also taken as features for training. The same features extrated with test data has to be ompared with trained data for lassifiation. The proposed methodology ahieves 77% lassifiation rate for abnormal gait. Keywords: abnormal walking style, abnormal gait lassifiation, KNN lassifier, reognition rate.. INTRODUCTION Due to ever-inreasing rime rate, identifiation using biometris has turn out to be a signifiant field of study. When a person wears helmet, spetales, loves and mask then it is impossible to apture fae, iris, figerprints et., and then identifiation using gait may be a suitable and an effetive tool to identify a person from a long distane. This paper presents a method whih makes a distintion between normal walk and abnormal walk. A person having abnormal gait may be ategorize as suspiious and alarming ations may be taken. There are so many experiments have been done on realisti data and system has been trained mostly for normal walk patterns and partiular onditions like walk with luggage arrying onditions, walk with different dress odes, different speeds, different walking surfaes and illumination hanges with respet to subjets. Experiments have been done on the signifiane about reognition and identifiation about the person. In general gait analysis an be ategorized as linial gait analysis and biometri gait analysis. Clinial gait analysis uses olletion of kinemati data in ontrolled environments using motion analysis systems. Motion analysis provides information regarding gait yle, speed, walking events et. Biometri gait analysis an be used for authentiation purposes beause reliable authorization and authentiation has beome an integral part of every individual s life for a number of routine appliations. Biometri is an automated method of reognizing a person based on a physiologial or behavioral harateristi. Gait is a vision-based human identifiation at a distane and has reently gained wider interest from the omputer vision ommunity due to day by day inreasing rime rate and intimidation. In the present senario, Banks, shopping malls, ATM enters, railway stations, airports, bus stands et, requires protetion from unwanted ativities. Every person posses his own walking style. Gait yle of a person is unique, whih made gait as one of the biometri. It is the only biometri does not require any nearness sensors. It is the only biometri does not give up any ue to suspeted person, like other biometris, whih gives an alert over measuring person. 2. LITERATURE REVIEW In general gait analysis tehniques are lassified into two group, they are three dimension model and two dimension model. In two dimensional model approahes, the gait is again lassified into two types, one is model based approah and the other is model free (holisti) approah. Various algorithms and parameters are applied into these two types of approahes to reognize the human. Prinipal Component Analysis uses Eigen values and Eigen vetors to onstrut the 3D linear model from a set of Fourier oeffiients derived from projeting 2D motion sequenes on to 3D. In the Sixteenth entury, the person an be identified from a long distane by his walking style itself. In the nineteenth entury initially gait and its patterns are analysed only for various medial reasoning and their rehabilitation alone. Gait pattern was used to analyse the mental, physial health of a person. After that gait reognition mainly used for the analysis of health monitoring of athleti and sports men. In reent era gait is a new biometri aimed at reognizing persons by the way they walk. Based on stati visual features, swing distanes and joint angles of human limbs, the system identifies the patient with Parkinson diseases [4]. The onept of pereived exertion was introdued in the late 950s with 376
methods measuring loal fatigue or breathlessness. Pereived exertion is defined by sensations of effort, onstraints, unomforted and fatigue felt by a person when exerising. However, the relevane of measuring pereived exertion in obese patients is still poorly known. So verifiation of exhaustion and exerise safety and rehabilitation programs for obese patients is required. Genu reurvatum affets between 40 and 68% of hemipareti stroke patients [6]. From a biomehanial point of view, it ours during the stane phase. In patients with quadrieps weakness, this phenomenon generates a knee extensor moment whih avoids ollapse during the stane phase. Fibromyalgia (FM) is an impairment disease that involves systemi hroni pain and its pathogenesis and etiology are still not fully understood. Funtionally, FM is a ondition frequently aompanied by diminished physial work apaity and musular fatigue [7]. Indeed, when ompared to a ontrol group, subjets with FM show altered gait parameters, haraterized by redued walking speed, yle frequeny, and stride length whih are also observed in the elderly. Kinemati deviations are observed in swing phase inlude dereased peak hip flexion, dereased peak knee flexion throughout swing. Based on stati visual, early researh was motivated by Johansson s [] and Barlay s [2] psyhologial experiments, where partiipants were able to reognize the type of movement of pedestrians simply from observing the 2D motion pattern generated by light bulbs attahed to several joints over their body. The earliest researh into omputer-vision based gait analysis tehniques was published by Sourabh A et al. [8], whih were based on spatio-temporal analysis and model fitting. Later that year, an Guo et al. [9] published an algorithm based upon a 0 stik model and neural network lassifiation. James J et al. [0] published a gait analysis tehnique based upon the spatial distribution of optial flow and how it varied over time. Hiroshi Murase et al. [4] proposed a tehnique that ompared the Eigen spae trajetories between subjets. This onept was later extended by Huang et al. [3]. Cunado et al. [6] published a model based tehnique that used the Hough transform to a model to the video frames; results were published on a small dataset reorded indoors, whih was to beome the first gait dataset widely used by others. James J et al. [2] published results of their previous algorithm, James J et al. [0] applied to a new dataset reorded outdoors, and this dataset also beame very popular in the researh ommunity. Other signifiant milestones inlude the release of the Gait Challenge dataset and baseline algorithm, provided by south Ampton University (the University of South Ampton's Human ID dataset), Jamie D. Shutler et al [3] publily available datasets and are still used extensively by researhers around the world. Reently CASIA database for gait (hina), OU-ISIR gait data base (Osaka university-japan) are also utilized. This paper inludes abnormal gait database ombined with normal gait database. A. Data aquisition We have onsidered subjets for our experimentation, having ages 8 and above. The purpose and the neessity about the experiments were explained to individual person who have taken into experiments. All are of male andidates. During experiments subjets were wearing different types of lothes like pant, shirt. The andidates wore shoes and slippers during experiments. They were told to walk on the trak shown in Figure-. The Proess Video has been reorded using SON Handy am HDR-C240.It has CMOS sensor with size of /5.8 inh. Resolution of "920 x 080", "440 x 080", frame rate of 60 pixels. The distane between amera and trak was around 25 feet and length of trak was 20 feet. Every person s walk was reorded. Gait yle of different walk pattern also alulated. It is alulated by identifying the duration from toe on to toe on of same leg. Video to frame onversion Three minutes of ontinuous walk onverted around200 frames in JPEG (. jpg) format with frame size of 920x080. The bakground model has subtrated from original frame. The resultant frame onsist only the foreground objet. Then the foreground image is onverted into silhouettes. The silhouettes are onsisting of noises and shadows. Real time hallenges in the real video sequenes, we have faed the diffiulties with shadows, illumination hanges. B. Researh gap In real time emotions of different persons may be different. So the requirement about the real time should adopt the real time onditions like various speed of walk, various dress ode, and different walking onditions. But still there is a gap between how system will ategorize the gait of an identified person as abnormal or dubious person. 3. PROPOSED METHOD A. Bakground subtration a. Aim In order to perform the gait analysis from the real time video frames, the subjet needs to be taking out from the video sequene. Image segmentation is the proess whih is used to separate dynami objets suh as people. One of the most ommon and easiest methods for performing segmentation is to perform image subtration tehnique.the known bakground image is subtrated from the urrent piture frame whih initially ompare the intensities of onneting pixels, then threshold is applied into this. A pixel is onsidered as a part of the foreground when the urrent pixel value differs from its mean value by more than a pre defined threshold value. b. Feature extration The proposed methodology uses height and width; six angles suh as head to nek, nek to torso, hip 3762
to knee of both legs, knee to foot angle of both legs were omputed. W max = max (W, W 2,.W N). The entroid ( ) of the human silhouette is, alulated by using the following equations N i N i N i N j () (2) where (, )represents the average ontour pixel i i position.(, ) represents points on the human blob, N- Total number of points on the ontour. Figure-. Flow hart for normal and abnormal gait lassifiation. ) Height and width alulation Maximum height (H) and maximum width (W) an be measured from silhouettes. When a person moves towards the amera, the height value is inreased. Similarly, when a person is away from the amera, the frame height ontinuously dereases. Variation of height and width an be evaluated by defining the bounding box and the entroid point for eah observed body is alulated. The height and width of box alters in a gait yle. Let H, H 2,.H N be the height of skeleton in a gait yle. Then the maximum height of the person in the entire silhouettes is denoted as H max = max (H, H 2.H N). The variation of the width is important ue for gait analysis, as it ontains strutural and dynamial information about the gait. When the person is in middle stane position, the spae between the two legs is small and hene the width is minimum. The maximum width is attained, when a person walks by swinging his arms. Let W, W 2.W N be the width of skeleton in a gait yle. Then the maximum spaing between two legs for a person is denoted as 2) Angle alulation The angles are alulated by the following proedure. The input silhouettes are applied into Fourier transform. The resultant Fourier oeffiients are given to Radon transform, to produe a rotation invariant image features. Applying Radon transform, it is possible to reate a mapping between ( x, y ) domain to the radon domain of ( r, ). During human walking, a large variation exists at leg portion. Sine the variation is higher, the output of the Radon oeffiients is also higher in its value, and it is very suitable for feature vetors. The radon transform of a skeleton image f x, y denoted as R ( r, ) where r defined by a normal distane from the origin, as a normal angle. Radon transform point R r, f x, y f( x, y) ( r xos ysin ) dxdy (3) r, 0 Where The angles from head to nek, nek to torso, hip to knee of both legs, knee to toe of two legs were alulated as feature vetors for eah person. It has been alulated and stored in database. Nearest neighbor tehnique has been Figure-2. exhaustive angle from head to nek -, nek to torso - 2, hip to knee ( 3, 4), knee to toe ( 5, 6). used to lassify the image belongs to lass ( person with normal walk), lass 2 (person with abnormal walk) or other one (a doubtful person). MATLAB programming 3763
has been used to simulate the proess. The Figure-2. Shows separation of six angles proess has been shown in Figure-2 B. Charlie haplin Figure-3. Bakground. C. Pattern lassifiation One the gait feature has been extrated from the person, it will be projeted into a feature spae and then lassified. A lassifier defines boundaries in a feature spae whih are used to separate different sample lasses from eah other in the data. The simplest type of lassifier is a linear lassifier. This is a straight line whih is defined in the feature spae, points above the line are in one lass, and points below the line are plaed into another lass. This is a simple method and unable to provide the best results due to small non-linear flutuations around the boundary region whih result in poor lassifiation results. An improvement on this method is alled the K-Nearest Neighbour (KNN) lassifiation. D. K-Nearest Neighbour lassifier A feature vetor of unknown lass an be lassified as belonging to a lass by using the k-nearest neighbour rule. A training set of points (i.e. feature vetors projeted in eigenspae) T is used to determine the lassifiation of feature vetor, by the following method:. Calulate the k-nearest points to the unlassified feature vetor in the feature vetor set T. There are a number of distane measures whih an be used to alulate the separation of two points in n- dimensional spae. 2) Determine the lass whih has the most points in the k seleted points, from set T C. Cirumdution Gait D. Sissors Gait E. Steppage Gait 4. RESULTS AND DISCUSSION A. Antalgi Gait 3764
Person Table-. Classifiation rate. Classifiation rate (KNN) Normal gait 90 Abnormal gait 77 5. CONCLUSIONS In this proposed work we introdue the abnormal gait database. The gait pattern was lassified as normal or abnormal using KNN lassifier. The lassifier reognizes walk with 77% lassifiation rate. Lower lassifiation rate was obtained beause of irregularity made by the two legs. REFERENCES [] Mark.S. Nixon, John.N.Carter, Automati reognition by gait, Proeedings of the IEEE, vol. 94, No, November 2006. [2] Barlay C., Cutting J., and Kozlowski, L., Temporal and Spatial Fators in Gait Pereption That Influene Gender Reognition, Pereption and Psyhophysis, vol. 23, no. 2, pp. 45 52, 978. [3] Huang P.S., Harris C.J., and Nixon M.S., Human Gait Reognition in Canonial Spae Using Temporal Templates, IEEE Proeedings on Vision, Image and Signal Proessing, vol. 46, no. 2, pp. 93-00, 999. [4] Cho CW, Chao WH, Lin SH, Chen. A Visionbased System for Gait Reognition in Patients with Parkinson s disease. Expert Syst Appliat 2009; 36: 7033-7039. [9] an Guo, Gang u and Saburo Tsuji. Understanding human motion patterns. In Proeedings of 2th International Conferene on Pattern Reognition, Volume 2, pages 325{329, Otober 994. [0] James J. Little and Jeferey E. Boyd. Desribing motion for reognition. In Proeedings of International Symposium on Computer Vision, pages 235{240, November 995. [] G. Johannson. Visual pereption of biologial motion and a model for its analysis. Pereption and Psyhophysis, 4:20{2, Otober 973. [2] James J. Little and Jeferey E. Boyd. Reognizing people by their gait: The shape of motion. Videre: Journal of Computer Vision Researh, (2), 998. [3] Jamie D. Shutler, Mihael G. Grant, Mark S. Nixon, and John N. Carter. On a large sequene-based human gait database. In Proeedings of Fourth International Conferene on Reent Advanes in Soft Computing, pp. 66-72, 2002. [4] Hiroshi Murase and Rie Sakai. Moving objet reognition in eigen spae representation: gait analysis and lip reading. Pattern Reognition Letters, 7(2): 55-62, February 996. [5] Cunado D. Nixon M.S., and Carter J.N., Automati Extration and Desription of Human Gait Models for reognition purposes, Computer Vision and Image Understanding, Vol. 90, no., pp. -4, 2003. [5] J.-B. Coquart, C. Tourny-Chollet, F. Lemaıˆtre, C. Lemaire, J.-M. Grosbois, M. Garin Relevane of the measure of pereived exertion for the rehabilitation of obese patients 55 (202) 623 640. [6] C. Bleyenheuft,. Bleyenheuft, P. Hanson, T. Deltombe Treatment of genu reurvatum in hemipareti adult patients: A systemati literature review, Annals of Physial and Rehabilitation Mediine 53 (200) 89-99. [7] Suelen M. Góes, Neiva Leite, Riardo M. de Souza, Diogo Homann, Ana C.V. Osieki, Joie M.F. Stefanello, André L.F. Rodaki Gait harateristis of women with fibromyalgia: premature aging pattern 2255-502/ 204 Elsevier Editor ltd. [8] Sourabh A. Niyogi and Edward H. Adelson. Analyzing and reognizing walking figures. In Proeedings of IEEE Computer Soiety Conferene on Computer Vision and Pattern Reognition, pp. 469-474, June 994. 3765