Human Recognition Based on Kinematics and Kinetics of Gait

Journal of Medical and Biological Engineering, 31(4): 55-63 55 Human ecognition Based on Kinematics and Kinetics of Gait u-chih Lin 1,* Bing-Shiang ang u-tzu Lin 3 i-ting ang 1 Department of Biomedical Engineering, uanpei University, HsinChu 3, Taiwan, OC Department of Mechanical Engineering, National Chiao Tung University, HsinChu 3, Taiwan, OC 3 Graduate Institute of Information and Computer Education, National Taiwan Normal University, Taipei 16, Taiwan, OC eceived 1 Jun 1; Accepted Nov 1; doi: 1.545/jmbe.86 Abstract Human recognition based on gait is a new biometric method that considers both spatial and temporal features which tracks the gait of a human being from a distance. The most widely used techniques for gait recognition at present can be divided into two categories: the model-free method and the model-based method. Both these methods process images of a moving body. Although some effort has been devoted to gait recognition, a high identification rate and low computational cost still need to be achieved. In this study, we investigated gait recognition in terms of biomechanics, which is often overlooked in the fields of computer science and electronics. We present a method for people recognition by using both the three-dimensional parameters for lower limb joints, i.e., kinematic and kinetic parameters. Dynamic data are acquired by tracking marker positions using a motion capture system with force plates. Although a marker-free and cable-free biometric system is more applicable for security purposes, acquiring data by marker tracking is one of the most precise methods currently available. The results obtained from marker tracking can clarify the relationship between the recognition rate and dynamic data of normal walking. In this study, 35 trials were conducted for 1 subjects. For each subject, trials were conducted for training and the remaining 15 were for testing. A self-organizing map (SOM) neural network was employed for data classification. The importance of a specific variable in each dimension for each joint is discussed in detail to investigate the major factors that cause the differences in human gait according to a biomechanical approach. Experimental results showed that gait is a reliable feature for individual identification since a high recognition rate can be achieved by choosing appropriate joints or dynamic parameters in some dimensions of a gait. Kinematic variables in the frontal and transverse planes of the knee or hip joints are recommended due to their higher recognition rates. These findings can be applied to both biomedicine and computer science. Keywords: Biometric, Neural network, Individual identification, Gait 1. Introduction With recent advances in technology and changes in community structure, individual recognition or identification is gaining increased importance. For example, security is needed to access computer systems, ATMs, credit cards, medical records management, buildings, etc. [1,]. Many studies have shown the feasibility and reliability of human recognition on the basis of gait because of the repeatability and uniqueness of human gait [3,4]. In biomedicine, studies have tended to focus on the repeatability and similarity of normal gait to distinguish between patients and normal subjects or assess changes in gait pattern after clinical surgery [5-8]. In contrast, in computer science, investigators have attempted to find discrepancies in normal gait for identity recognition. When all gait parameters are obtained, the gait pattern of each person is considered to be * Corresponding author: u-chih Lin Tel: +886-3-61346; Fax: +886-3-61347 E-mail: d87511@ntu.edu.tw sufficiently unique and thus reliable for individual recognition [4,9]. The most commonly used techniques for gait recognition are the model-free and the model-based methods [4,1,11]. Video images of a person walking in the model-free or silhouette-based method are processed using their silhouette. The silhouette-based method [1-14] directly manages image sequences without modeling any parameters [1]. In this method, binary maps (silhouettes) are extracted for applications without color or texture information. For example, the width w[i] of different heights of the silhouette, the top and bottom projections, and angular representation can be used as the features. It can be used directly or after Fourier transformation. This method depends on the accuracy of contour pixels; hence, noisy silhouettes may lead to inaccurate results. In contrast, the model-based method [15-19] involves parameters fed into a model. It is view- and scale-invariant [1]. The body parameters (height, distance between the head and pelvis, estimation of the inclination of the thigh and leg) may be used as features.

56 J. Med. Biol. Eng., Vol. 31 No. 4 11 Both the silhouette-based and model-based methods are marker-free methods [4,1,11]. Although it is more reasonable and convenient to consider the gait features in marker-free gait sequences [15], the information analyzed from video image processing is not sufficient and accurate. The trend is to use a marker-free motion capture method for gait recognition, but before accurate results can be obtained from the marker-free method, appropriate parameters for image processing must be identified. Presently, the analysis of marker positions obtained from a motion capture system is the technique with which most precise kinematic parameters for gait can be obtained. Since the kinetic characteristics are also important for gait dynamics, kinetic variables are also obtained by incorporating force plate measurements. In this study, we used the kinematic and kinetic variables derived from motion capture data as the biometrics instead of information obtained by using the image processing method. We attempted to separate each dynamic variable for different joints in each plane to thoroughly investigate the influence of different variables on the walking style of a person. As a result, we can suggest the variables, joints, and planes that are most viable for use as biometric features in marker-free image processing methods. mass: 71.18 ± 13.1 kg) took part in the experiment. They were all students from National Chiao Tung University in Taiwan. The ten subjects underwent 35 trials; for each subject, trials were conducted for training, and the remaining 15 were for testing. The dynamic parameters were computed by the commercial software SMAT Analyzer provided by BTS Corp. (BTS Spa Corp., Italy). We investigated the importance of the dynamic parameters for each joint in each plane in detail to determine the major factors that cause differences in human gait from a biomechanical point of view. A self-organizing map (SOM) neural network was used as the algorithm for data classification, which is presented in section.3.. Methods.1 Motion analysis A motion capture system (BTS Spa Corp., Italy) and two force plates (BETEC Corp., USA) were used to acquire the three-dimensional kinematic and kinetic variables, respectively, during the gait. The Davis marker set was used in this experiment. Eleven markers were attached to the pelvis and one of the lower limbs of each of the subjects, as indicated in Fig. 1 (the figure shows the markers for two lower limbs, but the markers were only attached to one lower limb in this experiment). Three markers were on the pelvis: Anterior superior iliac spine (ASIS) (right and left) and, on the back, between the posterior superior iliac spine (PSIS). Three markers were on each thigh: great trochanter, femoral condyle, and lateral bar. Three markers were on each shank: head of the fibula, and lateral malleulus, lateral bar (same way for the bar on the thigh). Finally, two markers were on each foot (fifth metatarsal joint and heel) []. We alternated placing the markers on the right and left legs of each subject to decrease tracking time and increase accuracy by decreasing marker numbers. Further simplification of experiments and analysis procedures is necessary for real-world applications. The subjects were asked to walk at a natural speed along a walkway. There were 5 8 strides per trial and 3 strides between gait initiation and contact with the force plate. The exact number of strides depended on the step length of each subject. We did not ask subjects to purposely step on the force plate but ensured contact with it by adjusting the start position. The cadences and step lengths were measured by the system but not included in the recognition computations. A gait cycle was defined as the interval between two heel strikes. Ten male subjects (age: ± (mean ± SD) years; height: 1.7 ±.7 m; Figure 1. Davis marker set positions. Three markers are on the pelvis: ASIS (right and left) and on the back between PSIS. Three markers are on each thigh: great trochanter, femoral condyle, and lateral bar. Three markers on each shank: head of fibula, lateral malleulus, lateral bar (same way for the bar on the thigh). Two markers are on each foot (fifth metatarsal joint and heel) [].. Statistical analysis The variability values were computed by using modified versions of the equations of Kadaba et al. [1]. We used the modified equations since it is more complicated to analyze the similarity or variability of waveforms, and satisfactory results cannot be obtained if only simple statistics are used [1]. The adjusted coefficient of multiple determination a for between subjects was given by a S T ( i 1 j 1 t 1 1 S T ( i 1 j 1 t 1 ijt ijt ) / T( S 1) t ) /( S T 1) where ijt is the t th time point of the j th run for the i th subject, and, S, and T are the total run number, subject number, and time points, respectively. 1 ST and S i 1 j 1 T t 1 ijt (1) ()

Human ecognition by GAIT 57 1 t ijt S i 1 j 1 S The adjusted coefficient of multiple determination a for within a subject and between trials was given by a where 1 S T ( i 1 j 1 t 1 1 S T ( it ijt j 1 and 1 i 1 j 1 t 1 i ijt T j 1 t 1 T ijt ijt ) / ST( 1) it ) / S( T 1) i The positive square root of the adjusted coefficient of multiple determination was used to describe the repeatability in our experiment, and the variability was the one s complement of repeatability. The paired t-test statistical method was used to compare the recognition rates and variations. SPSS v1. (SPSS Inc., Chicago, IL) was used for the statistical computation. (3) (4) (5) (6) features of certain image appearances using equation (7), where W i (t) is the weight vector of the ith unit in the tth iteration, α(t) is the learning rate, and N c is the list of unit indices that make up the neighborhood. wi ( t) ( t)( x wi ( t)), i Nc wi ( t 1), otherwise. After iterations to update neuron values by Kohonen s algorithm [-4], nodes in the -D array organize themselves in response to the input vectors. The resulting map projects the input vectors onto a -D map, preserving the natural topology of the training data, which contain the feature vectors of various types of samples. In the proposed SOM-based classification scheme, a -D SOM with 4 4 neurons is initially constructed. Each node is randomly assigned to feature values for one of the training samples. The learning algorithm performs the weight-update process by using an approximation to the Mexican-hat function [5]. If one node is too coarse to represent a certain category of training data, it is divided into four child nodes. Further dividing should proceed if nodes after the first division are still too rough. The architecture of our SOM network is shown in Fig. 3. Vectors with similar features are trained to gather together in adjacent nodes. (7).3 SOM architecture A SOM neural network was used as the algorithm for data classification in this experiment. As shown in Fig., each feature vector of each subject was extracted to be one point in the feature space. The SOM nodes are initially positioned arbitrarily in the data space. While the input training sample is feeding the SOM map, the node (neuron) nearest to the training node (black circle on the map) is selected and moved towards the current training sample (white circle in the feature space), as are its neighbors on the grid. After many iterations, the grid tends to approximate the data distribution. Figure 3. Architecture of the adopted SOM network. The dotted lines map neurons to feature vectors expressed in images. SOM architecture Subjects Network learning Feature extraction Mapping Feature space Figure. SOM mapping. Each feature vector of each subject is extracted to be one point in the feature space. The SOM neural network can imitate the mechanism of a cerebral cortex. A -D map is used in our architecture to represent the cortex map memorizing the distortion sensitivity for variant image patterns. During the training period of the network, each unit with positive activity within the neighborhood of the winning unit updates its value toward the In the training stage, the training samples of the dynamics are first categorized manually. Each training vector is classified into a category in which all samples have similar features. If the feature vector of the current training sample is the nearest to the center of a certain category, the coordinate of that center is modified and moved toward the training vector. After the iterations of learning, the SOM eventually memorizes the features for every category of the training data. In the recall process, the input vector is fed to the neural network as input. If features of the current input sample are most similar to features of a certain category, the current input sample is classified into that category. In this experiment, the learning rate α(t) in (7) was initially set to.8 and changed as the number of iterations increased (by a factor of.). In addition, the neighbors centered around the current training node were all modified according to the Mexican-hat function. The training stage then stopped after 1 training iterations.

Angle (degree) Angle (degree) Angle (degree) Angle (degree) Angle (degrees) Angle (degree) 58 J. Med. Biol. Eng., Vol. 31 No. 4 11 3. esults The joint angles in the three planes were obtained at the hip and knee, while only the sagittal and transverse planes joint angles were obtained at the ankle joint in this experiment. The average joint angles and standard deviations for training trials on the sagittal planes of each subject are shown in Figs. 4(a)-4(c). The mean values are shown as solid lines, while the standard deviations are shown as broken lines. The mean value of standard deviations of hip flexion/extension for each curve: Subject1:.41E+, Subject: 1.3E+, Subject3: 1.35E+, Subject4: 1.51E+, Subject5: 1.34E+, Subject6: 1.37E+, Subject7: 1.38E+, Subject8: 1.56E+, Subject9: 1.19E+, Subject1: 1.41E+. The mean value of standard deviations of knee flexion/extension for each curve: Subject1:.3E+, Subject: 1.84E+, Subject3:.19E+, Subject4:.17E+, Subject5: 1.8E+, Subject6:.3E+, Subject7: 1.9E+, Subject8:.59E+, Subject9: 1.65E+, Subject1:.36E+. The mean value of standard deviations of ankle flexion/extension for each curve: Subject1: 1.61E+, Subject: 1.9E+, Subject3: 1.46E+, Subject4: 1.57E+, Subject5: 1.1E+, Subject6: 1.65E+, Subject7: 1.8E+, Subject8: 1.69E+, Subject9: 1.3E+, Subject1: 1.6E+. As shown in these figures, different trials for the same subject had high repeatability and small standard deviations. (a) (b) 4 - -4 8 6 4 - (c) 1-1 - -3-4 Hip flexion/extension 4 6 8 1 Knee flexion/extension 4 6 8 1 Ankle flexion/extension 4 6 8 1 subject 1 subject subject 3 subject 4 subject 5 subject 6 subject 7 subject 8 subject 9 subject 1 subject 1 subject subject 3 subject 4 subject 5 subject 6 subject 7 subject 8 subject 9 subject 1 subject 1 subject subject 3 subject 4 subject 5 subject 6 subject 7 subject 8 subject 9 subject 1 Figure 4. (a) Average hip joint angles (flex/extension) and the standard deviation of the training trials of each subject. (b) Average knee joint angles (flex/extension) and the standard deviation of the training trials of each subject. (c) Average ankle joint angles (flexion/extension) and the standard deviation of the training trials of each subject. The results for the frontal and transverse planes showed similar phenomena as those for the sagittal plane. The average angles and standard deviations of the joint angles in the sagittal planes for the 1 subjects in each trial (trials 1 5) are shown in Figs. 5(a)-5(c). The mean value of standard deviations of hip flexion/extension for each curve: Trial1: 4.3E+, Trial: 4.76E+, Trial3: 4.67E+, Trial4: 4.7E+, Trial5: 4.77E+. The mean value of standard deviations of knee flexion/extension for each curve: Trial1: 5.63E+, Trial: 5.91E+, Trial3: 6.E+, Trial4: 6.18E+, Trial5: 6.33E+. The mean value of standard deviations of ankle flexion/extension for each curve: Trial1: 3.5E+, Trial: 3.1E+, Trial3: 3.3E+, Trial4: 3.63E+, Trial5: 3.75E+. Each curve indicates the mean value and standard deviation of the 1 subjects. The standard deviations between subjects (Figs. 5(a)-5(c)) were larger than those within a subject (Figs. 4(a)-4(c)). The variation curves for the mean values of all subjects for each trial are almost overlapping, and the standard deviations are within similar ranges. The results for all three planes showed the same behavior; hence, only the results for the sagittal plane are presented here for conciseness. The variability of joint angles and joint forces within a subject, between trials and between subjects in each plane is shown in Table 1. (a) 4 - -4 Hip Flex-extension 4 6 8 1 Knee Flex-extension 1 9 8 7 6 5 4 3 1-1 - (b) (c) 1-1 - -3-4 4 6 8 1 Ankle Flex-extension Trial1 Trial Trial3 Trial4 Trial5 Trial1 Trial Trial3 Trial4 Trial5 Trial1 Trial Trial3 Trial4 Trial5 4 6 8 1 Figure 5. (a) Average hip joint angles (flex/extension) and the standard deviation of 1 subjects in trials 1 5. (b) Average knee joint angles (flex/extension) and the standard deviation of 1 subjects in trials 1 5. (c) Average ankle joint angles (flex/extension) and the standard deviation of 1 subjects in trials 1 5.

ecognition ates(%) Human ecognition by GAIT 59 Table 1. Variations in joint angles and joint forces between subjects and within a subject in each plane. Joint angle (degrees) Joint Hip Knee Ankle Flx/ext Abd/add Int/ext rot Flx/ext Abd/add Int/ext rot Flx/ext Int/ext rot (1) Within a subject.43.1.549.68.178.53.195.15 () Between subjects.418.135.367.467.496.3531.17.5175 Joint force (Nt) Joint Hip Knee Ankle Medial/ Forward/back Medial/ Forward/back Vertical Medial/ Vertical force lateral force ward force lateral force ward force force lateral force (1) Within a subject () Between subjects.318.33.356.1668.41.4.1838 *.4535.181.166.45.77.178.4474 * Vertical force The table shows that the variability between subjects was considerably larger than that within a subject (p <.5 by the paired t-test), especially in the frontal and transverse planes. The difference between subjects was relatively larger than that within a subject, as shown in Figs. 4 and 5. These results clearly show that we can differentiate between humans on the basis of their gaits. Hence, we used the kinematic data (joint angles) and kinetic data (joint forces) as the features for human recognition in this study. First, we used 16, 5, 36, 49, and 64 neurons to compute the recognition performances for different joint angles. The recognition rates using joint angles in all three dimensions for the hip and knee joints and two dimensions for the ankle joint were measured. The recognition results for each joint using kinematic data with various numbers of neurons are shown in Fig. 6. 1 1 8 8 6 4 16 neurons 16 neurons 5 neurons 36 neurons 49 neurons 64 neurons Ankle Knee Hip All joints Ankle Knee Hip All joints Joints Joints Figure 6. ecognition rates obtained from three-dimensional joint angles of each joint using different numbers (16, 5, 36, 49, and 64) of neurons. The values were computed from the three-dimensional joint angles, i.e., values of sagittal, frontal, and transverse planes were all included in the computation. Experiments using all three joints were also carried out for the sake of comparison. The recognition rates using the joint angles of the hip, knee, ankle, and all three joints were all very high regardless of the number of neurons used. The average recognition rates for all joint cases were best when 5 neurons were used; Hip: 99.33%, Knee: 98.67%, Ankle: 89.33%, All joints: 99.33%, and Average: 96.67%. The mean recognition rates for all neuron cases were best for the knee joint and worst for the ankle (Hip mean: 94.8%, Knee mean: 96.53%, Ankle mean: 9.4%, All joints mean: 99.7%). The results were the best when all three joints were taken into consideration. However, the differences were not always significant, as the results for different joints were compared by the paired t-test, as shown in Table. Table. esults of the paired t-test of recognition rates obtained using different joints. Joint p Ankle/knee.6* Ankle/hip.88 Ankle/all Joints.1* Knee/hip.343 Knee/all Joints.91 Hip/all Joints.13 The experiment conducted on the ankle joint had the worst result because of the fewer dimensions used. Thus, we needed to compute the results of each plane independently in order to thoroughly investigate the correlation between the joint variables and recognition rates. The recognition results for each joint using different kinematic and kinetic variables in each plane are listed in Tables 3 and 4, respectively. Twenty-five neurons were used to recognize different humans based on joint angles and joint forces. Table 3. ecognition rates obtained by using different joint angles in the sagittal, frontal, and transverse planes (5 neurons used). ecognition rate (%) Flx/ext Abd/add Int/ext rot Ankle angle 76. * 74. Knee angle 88. 96. 91.33 Hip angle 8.67 9.67 98.67 Table 4. ecognition rates obtained by using different joint forces in the sagittal, frontal, and transverse planes (5 neurons used). ecognition Medial/lateral Forward/backward rate (%) force force Vertical force Ankle force 7. * * Knee force 94.67 9.67 76.67 Hip force 84.67 87.33 8.67 We employed the SOM algorithm for determining the physiological or behavioral characteristics of a person by observing the relationship of the neuron topology. One of the characteristics of SOM is its meaningful topological structure; i.e., its topology can reflect the relationship of the input data. For example, if the trials of two people are often classified into the same neuron or two nearby neurons, the walking style of the trial subjects must be similar. Discussing the topology of neurons (dynamic data analysis results) and the characteristics of people may provide us with more information about human

6 J. Med. Biol. Eng., Vol. 31 No. 4 11 Figure 7. (a) SOM training results and the neural topology using the hip joint angle during gait (5 neurons used). Different legends indicate the different subjects, and the results are obtained from trials for each subject. (b) SOM training results and the neural topology using knee joint angle during gait (5 neurons used). Different legends indicate the different subjects, and the results are obtained from trials for each subject. (c) SOM training results and the neural topology using ankle joint angle during gait (5 neurons used). Different legends indicate the different subjects, and the results are obtained from trials for each subject. recognition by gait. The SOM training results and neural topologies obtained using data related to the joint angles and joint forces in each joint are shown in Figs. 7(a)-7(c) and 8(a) (c). The training trials of the 1 subjects were used for analysis. The trials of the same subjects classified into the same neuron were not marked specifically; therefore, we only saw one label for many trials on the same neuronal position if the trials belonged to the same subject. As indicated in Figs. 7(a) and 7(b), when the hip joint and knee joint angles were used as the recognition features, the characteristics of each person were more easily separated. As shown in Fig. 7(c), we found that subjects 3 and 6 had neurons that were located close to one another and were sometimes confused. Although the close positions of the neurons for subjects 1 and 5 in both the hip and ankle joints can be seen in Figs. 7(a) and 7(c), the topology relationships between subjects were found to actually be different when different joint angles were used in this experiment. In other words, the characteristics of

Human ecognition by GAIT 61 Figure 8. (a) SOM training results and the neural topology using the hip joint force during gait (5 neurons used). Different legends indicate the different subjects, and the results are obtained from trials for each subject. (b) SOM training results and the neural topology using the knee joint force during gait (5 neurons used). Different legends indicate the different subjects, and the results are obtained from trials for each subject. (c) SOM training results and the neural topology using ankle joint force during gait (5 neurons used). Different legends indicate the different subjects, and the results are obtained from trials for each subject. different joints are different in each person, and similar properties for the hip joint in two persons do not necessarily indicate similar properties for the knee joint. Based on the kinetics results, the distribution of neurons was concluded to be different from that obtained from the kinematics results. Subjects 3 and 5 and subjects 6 and 1 had closely located neurons, respectively, when both the hip joint and knee joint forces were used, as shown in Figs. 8(a) and 8(b). When the hip joint force was used, as indicated in Fig. 8(a), subjects 6 and 7 also had closely located neurons. The computation results using the ankle joint force were not very good; hence, the neural topologies are more complex in Fig. 8(c). The neurons for subjects 6 and 7 were also closely located in the ankle joint, as in the hip joint. This indicates that the kinetic characteristics of subjects 6, 7, and 1 were similar.

6 J. Med. Biol. Eng., Vol. 31 No. 4 11 4. Discussion In general, the recognition rates using kinetic data (joint forces) were lower than those using kinematic data (joint angles) in this experiment, as indicated in Tables 3 and 4. Kinetic data provided better recognition results with the medial/lateral and forward/backward forces than the vertical force. The recognition rate obtained by the knee joint forces was lower for the vertical force because of the relatively small variation between subjects when compared to the other planes in the knee joint. The recognition rates obtained using the kinematic and kinetic variables of the hip and knee joints were greater than those obtained by using the ankle joints for all planes. This experiment shows that the recognition rates derived from the ankle joint are the least desirable. The worse results for the ankle joint may be caused by the complexity of the joint s structure, which may have made the stability less in this experiment. There are nearly 4 joints in the human feet; hence, the motion is quite complex and difficult to describe and measure. The motion axis (dorsi/plantar flexion and internal/external rotation) that we measured by using the skin marker positions in this experiment is actually slightly different from the real one. Tables 1 and 3 also show that, in general, the greater the variations in joint angles between subjects, the higher the recognition rates that can be obtained for the hip and knee joints. The joint angles are more repeatable between subjects in the sagittal plane; hence, the recognition rates were lower than those for other planes in the hip and knee cases. As shown in Tables 1 and 4, higher variations in the joint forces do not exactly correlate with greater recognition rates. However, they still indicate that the joint reaction forces of the hip and knee joints for the medial/lateral and forward/backward forces were less repeatable between subjects and had higher recognition rates. Moreover, the results for the vertical force were the worst because of the relatively lower variability between subjects. This suggests that hip and knee joint angles in the frontal and transverse planes be used as the lower limb features for individual recognition. In order to reduce the computational cost, using only the knee or hip parameters instead of all three joints is recommended. Based on the SOM training results and neural topologies, we speculate that there is a more interesting connection between the neural topology and walking style of a human being than is currently known. A more thorough investigation covering more subjects is required for realizing the relationships between human walking and identity recognition in both biomedicine and computer science. 5. Conclusions This study has proposed a human recognition scheme that uses gait as a feature and an SOM neural network as a classifier. High recognition rates were obtained when three-dimensional kinematic and kinetic variables of the lower limbs were used as features for human recognition. Better recognition rates were obtained when the knee or hip joint angles were used. Using ankle joint parameters led to the worst results. In general, kinematic parameters were better than kinetic parameters for gait recognition. Exercises in which joint angles of the frontal and transverse planes were used exhibited a better outcome than those in which joint angles of the sagittal plane were used. Furthermore, results obtained by using joint forces in the sagittal and frontal planes of the hip and knee joints were better than those from the transverse plane. In conclusion, gait is a reliable feature for human recognition, and the proposed methodology is suitable for application to research in the fields of biomedicine and computer science. A better marker-free method is still required for practical applications; however, before precise results can be obtained, dynamic data gained by tracking marker positions can provide us with a considerable amount of information on gait recognition and thus help us in developing a more efficient recognition method. Acknowledgements The authors gratefully acknowledge financial support provided by the National Science Council, epublic of China, under Grant NSC 97-1-E-64-1. eferences [1] A. K. Jain, A. oss and S. Prabhakar, An introduction to biometric recognition, IEEE Trans. Circuits Syst. Video Technol., 14: 4-, 4. [] S. Prabhakar, S. Pankanti and A. K. Jain, Biometric recognition: security and privacy concerns, IEEE Secur. Priv., 1: 33-4, 3. [3] S. V. Stevenage, M. S. Nixon and K. Vince, Visual analysis of gait as a cue to identity, Appl. Cogn. Psychol., 13: 513-56, 1999. [4] M. S. Nixon and J. N. Carter, Automatic recognition by gait, Proc. IEEE, 94: 13-4, 6. [5] F. C. Su and W. L. Wu, Design and testing of a genetic algorithm neural network in the assessment of gait patterns, Med. Eng. Phys., : 67-74,. [6] W. L. Wu and F. C. Su, Potential of the back propagation neutral network in the assessment of gait patterns in ankle arthrodesis, Clin. Biomech., 15:143-145,. [7] F. C. Su, W. L. Wu,. M. Cheng and. L. Chou, Fuzzy clustering of gait patterns of patients after ankle arthrodesis based on kinematic parameters, Med. Eng. Phys., 3: 83-9, 1. [8] M. Kohle, D. Merkl and J. Kastner, Clinical gait analysis by neutral networks: issues and experiences, IEEE Symp. on Computer-based Medical Systems, 138-143, 1997. [9] D. K. Wagg and M. S. Nixon, On automated model-based extraction and analysis of gait, Proc. IEEE Int. Conf. Face and Gesture ecognition, 11-16, 4. [1] N. V. Boulgouris, D. Hatzinakos and K. N. Plataniotis, Gait recognition: a challenging signal processing technology for biometric identification, IEEE Signal Process. Mag., 78-9, 5. [11]. Urtasun and P. Fua, 3D tracking for gait characterization and recognition, Proc. IEEE Int. Conf. Face and Gesture ecognition, 138-143, 4. [1]. T. Collons,. Gross and J. Shi, Silhouette-based human identification from body shape and gait, Proc. IEEE Int. Conf. Face and Gesture ecognition, 366-371,. [13] L. Lee and W. E. L. Grimson, Gait analysis for recognition and classification, Proc. IEEE Int. Conf. Face and Gesture ecognition, 148-155,. [14] P. J. Phillips, S. Sarkar, I. obledo, P. Grother and K. Bowyer, Baseline results for the challenge problem of human ID using gait analysis, Proc. IEEE Int. Conf. Face and Gesture ecognition, 137-14,.

Human ecognition by GAIT 63 [15]. Zhang, C. Vogler and D. Metaxas, Human gait recognition, Conf. Comput. Vis. Pattern ecognit. Workshops, : 34-349, 4. [16] A. F. Bobick and A.. Johnson, Gait recognition using static, activity-specific parameters, Conf. Comput. Vis. Pattern ecognit. Workshops, 1: 43-43, 1. [17] C. BenAbdelkader,. Cutler and L. Davis, Stride and cadence as a biometric in automatic person identification and verification, Proc. IEEE Int. Conf. Face and Gesture ecognition, 37-377,. [18] C.. am, M. S. Nixon and J. N. Carter, Automated person recognition by walking and running via model-based approaches, Pattern ecognit., 37: 157-17, 4. [19] L. Wang, H. Ning, T. Tan and W. Hu, Fusion of static and dynamic body biometric for gait recognition, IEEE Trans. Circuits Syst. Video Technol., 14: 149-158, 4. []. B. Davis, S. Ounpuu, D. Tyburski and J.. Gage, A gait analysis data collection and reduction technique, Hum. Mov. Sci., 1: 575-587, 1991. [1] M. P. Kadaba, H. K. amakrishnan, M. E. Wootten, J. Gainey, G. Gorton and G. V. B. Cochran, epeatability of kinematic, kinetic, and electromygraphic data in normal adult gait, J. Orthop. es., 7: 849-86, 1989. [] T. Kohonen, The self-organizing map, Proc. IEEE, 78: 1464-148, 199. [3]. ojas (Ed.), Neural Networks: a Systematic Introduction, New ork: Springer-Verlag, 1996. [4] T. Kohonen, Self-organized formation of topologically correct feature maps, Biol. Cybern., 43: 59-69, 198. [5] J. A. Freeman and D. M. Skapura (Eds.), Neural Networks: Algorithms, Applications, and Programming Techniques, New ork: Addison-Wesley, 199.

64 J. Med. Biol. Eng., Vol. 31 No. 4 11