Biometric gait identification based on a multilayer perceptron

Biometric gait identification based on a multilayer perceptron

*Vijay Bhaskar Semwal, Manish Raj and G.C.Nandi Department of Robotics & Artificial Intelligence

Indian Institute of Information Technology Allahabad, Allahabad *[email protected]

Please cite this article as: : V.B. Semwal, M. Raj, G.C. Nandi, Biometric gait identification based on a multilayer perceptron , Robotics and Autonomous Systems (2014),http://dx.doi.org/10.1016/j.robot.2014.11.01

Online Link : http://www.sciencedirect.com/science/article/pii/S0921889014002632

Abstract: This research presents a novel approach for biometric gait identification. A multilayered back propagation algorithm based artificial neural network (ANN) has been designed for gait pattern classification. The results have also been compared with the k- mean and k-nearest neighbor (KNN) algorithms. Novelty in feature extraction lies in the kernel based principle component analysis as the captured real time data have significant nonlinearity. The gait data have been classified into four classes: Normal, crouch-2, crouch-3 and crouch-4. The proposed method promises the gait identification with very good activity recognition accuracy (ARA). The experimental results show that the proposed methodology can recognize accurately different activities in both outdoor and indoor environment while maintaining a high ARA. Identification of disorder and abnormality in gait pattern is the fundamental concern of this research. We have presented an early detection methodology of abnormal gait which can work as a warning of any potential disease related to human walking. Further, this gait based biometric identification research can be utilized for detection of gender, age, race and authentication. Keywords: Machine Learning, Gait Pattern, Activity Recognition Accuracy, Biometric, Artificial Neural network, Authentication.

INTRODUCTION

The Human gait is considered to be a unique biometric identification tool [1] like a finger print [2]. It can be used to identify persons for various security reasons and to detect different abnormalities much earlier before it can usher permanent damage. The data uses for pattern classification and analysis for different walking [3] is worth to study for prediction of numerous upcoming diseases by looking the abnormality in GAIT pattern. Also Gait is a signature of human walking and can be used for her/his identification [4] purposes. However, human gait synthesis is a complex phenomenon, since it is the manifestation of synchronized motion of both upper and lower body parts. This synchronization is made through several months of learning (for a normal kid up to one year is needed for stable walking) and in that learning process different parts of brain gets involved in establishing co-ordination with nerves and muscle i.e. motor system and sensory organ. The mechanism of learning is almost impossible to study with the existing knowledge about human brain functioning. So we collect data and build biped robots in the digital space to run those robots with the data. Such study can also help us in getting greater insight to humanoid push recovery capability [5] and to capture the biometric identity like gender, age and race based on human locomotion [6] [7].The different previous research work used many machine learning techniques to capture such complex nonlinear patterns [8- 9]. Also sometimes identification of abrupt changes in gait patterns may provide timely alarm for inviting enhanced security [10-13] (e.g. in a video surveillance environment) and safety (e.g. in a life-critical environment such as a patient monitoring system). In this paper, we tried to correlate the Gait with human activity recognition. Due to its inherent uniqueness, recently the GAIT research community [14-17] is actively involved in developing GAIT pattern recognition technique using various Machine Learning tools [18] which can also help in recognizing persons from distance [19]. CCTV camera as surveillance tool, sometimes fails to detect potential threat due to time gap between identification and recognition which can be improved by using distance based identification technique. Similar techniques can be useful at various places such as parking lot, crowded market place, pedestrian crossing, bank, etc. The whole classifications aim is to discriminate among individuals based on their locomotion characteristics [20]. Unlike other rather conventional biometric identification technique such as finger print, iris, face reorganization [21], the GAIT is unobtrusive [22] due to its inherent nonlinearity and complexity. Some useful comparison of existing approaches of GAIT recognition has been provided in Table-1. As humanoid are suppose have complex social environment so it is required to develop the humanoid gait for complex terrain [42]. Toward development of such complex robot Chris et al. developed the complex gait design for ice skating humanoid [43]. The chris et al. contribution is revolutionary step for development of sophisticates robot for uneven terrain. Table 1 shows the all the existing approach of activity reorganization using locomotion.

Table.1. Comparison of existing approaches.

Ref. Method Advantage Disadvantage Uses (Bobick et al. 2001) [14]

Activity recognition using temporal templates

Real time application More than one classifier reduces the accuracy

Indoor/Outdoor both

(Vega et al., 2003) [15]

Spectral analysis of human motion

Higher-order Spectral Periodic detection Differentiate between people and vehicular objects

(Vega et al., 2003) [15]

View based motion analysis

Object models are not required

Need to reduce the distribution combinatory

Outdoor scenario

(Wang et al.,2004) [16]

Gait recognition Locomotion human model

Insensitive to noise Indoor scenario

(Noorit et al. 2010) [17]

Model-based Action Recognition

Inclusion of motion texture

Poor performance in walking case

Indoor environment

(Chris et al., 2014) [43]

Relies entirely on motion in the frontal plane to propel the robot

Gait design for an ice skating humanoid robot

classical inverted pendulum-based Walking gait when using the same skates unstable

both ice skates and inline skates

The gait data is classified using different machine learning technique ANN [23], KNN [24] and K-Mean

[25] algorithm and analyzed classified data. Any study related to locomotion will help to understand the problem of disable person and will help to develop the more advance sophisticate humanoid. There are of different types of gait patter jumping, running, dancing, walking, pushing and sitting but all these type of gait are not cyclic. Out of these jumping we considered as coordinated and cyclic but it do not consider as locomotion. The high dimensional feature spaces lead for high computation. To reduce the computational cost we need to reduce the dimension. The reduced dimensionality also improved the classification accuracy. The one which required less training is the best machine learning algorithm. So to reduce computation cost feature reduction is becoming important. Let assume that the data have feature X(i), i=1,2,….,l which make up the l- dimension feature vector 𝑋 = [𝑋(1), 𝑋(2), … . , 𝑋(𝑙)] ∈ 𝑅 . The two major objectives of dimensionality reduction is reduce computational cost and generalized accuracy cost. The benefits of reduced dimensionality will require less time to training and classify it and it minimizes the risk of over fitting. Less dimensionality leads for less training cost and therefore it improve the high generalization capability of classification Algorithms, for reduced dimensionality feature vector. Feature selection is the process of selecting important feature still retains sufficient information for classification. There are of two class of feature selection the filter method and wrapper method. The classification based on distance measure to expressed the how good the classes are separates from each other is filter based classification. Instead of selecting proper subset the wrapper methods are classifier dependent. This method directly calculates the value based on feature. Wrapper method measures the correctness of the algorithm which implies the good accuracy. Due to high computational cost and wrapper method is not used even thought the performance is good.

Fig.1: Flow Diagram of Gait Classification using different machine learning techniques

Human gait is considered as the locomotion with the support of very body parts following stages: lift one leg with the support lifted leg till it will be not in front of your bodybody forward with lifted leg contact groundsuch as a person is walking, running, jumping, jogging etc are important activities in video Surveillance. We contribute the use of ANN for activity recognition [23] with the help of movement of legs only. Experimental results suggest that our method is able to recognize the human activities with a good accuracy rate. coordinated using regular periodic motion of upper or lower body extremity;locomotion of individual. It is considered very difficult to disguiidentification system. In this paper we first selected the principle feature using classified gait data into five different gait data named normal and four types of crouchmachine learning technique [24] and finally point out the performance follow: The first section highlights on theabout the technique of data smoothing, data correction and analysis over various machine learning techniques. third section defines about the experimental result and different pattern found over the different subject. The section presents the performance using different machine learning other performance parameter for this analysis. algorithm for classification of gait.

Flow Diagram of Gait Classification using different machine learning techniques

Human gait is considered as the locomotion with the support of very body parts [19]. The entire gait consist of lift one leg with the support of another leg on ground and move your body forward and swing the

lifted leg till it will be not in front of your body [20]. The cycle of locomotion finally culminates by body forward with lifted leg contact ground. There are different type of activity associate with humanoid locomotion such as a person is walking, running, jumping, jogging etc are important activities in video Surveillance. We contribute the use of ANN for activity recognition [23] with the help of movement of legs only. Experimental results

le to recognize the human activities with a good accuracy rate. coordinated using regular periodic motion of upper or lower body extremity; this is responsible for

It is considered very difficult to disguise and conceal the Gait based biometric first selected the principle feature using KPCA then based on these feature we

classified gait data into five different gait data named normal and four types of crouch and finally point out the performance comparison. The whole paper is

on the important of this study and their benefits. The Second section presents nique of data smoothing, data correction and analysis over various machine learning techniques.

third section defines about the experimental result and different pattern found over the different subject. The ng different machine learning techniques; we designed the confusion matrix and

other performance parameter for this analysis. We concluded with observation that ANN is the best performing

Flow Diagram of Gait Classification using different machine learning techniques

The entire gait consist of of another leg on ground and move your body forward and swing the

finally culminates by moving whole activity associate with humanoid locomotion

such as a person is walking, running, jumping, jogging etc are important activities in video Surveillance. We contribute the use of ANN for activity recognition [23] with the help of movement of legs only. Experimental results

le to recognize the human activities with a good accuracy rate. The human walk is this is responsible for unique

Gait based biometric PCA then based on these feature we

and applied the ANN whole paper is organized as

important of this study and their benefits. The Second section presents nique of data smoothing, data correction and analysis over various machine learning techniques. The

third section defines about the experimental result and different pattern found over the different subject. The Fourth we designed the confusion matrix and

We concluded with observation that ANN is the best performing

Fig.2:

A. Background

The main objective of gait study to understand the problem of disable person, elder people and used to developed prosthesis l[26]. The classification can use for development of biological inspired bipedal machine which can work similar to human environment [27]. In clutter environment push is very common experience phenomena, so push recovery humanoid, to mimic such capability we need intelligence.intelligent agent for machine [28- 29]. Fig.2. shows the earlier implementation of planned activity before processing later Fig.3flow diagram of gait classification using different machine learning techniques

Broadly we used the three memories to allow learning at different phases [part one is perceptual used to perceive the information and store this information of external event, Second one is working arefer as short term and third one is long term memorycycle the information store into long term memory as previous

Fig.2: Human GAIT Sagittal plane view [6]

The main objective of gait study to understand the problem of disable person, elder people and used to developed prosthesis l

The classification can use for development of biological inspired bipedal machine which can work similar to human In clutter environment push is very common experience phenomena, so push recovery ic such capability we need intelligence. To behave as human lot of research is taking place to

. Fig.2. shows the earlier implementation of artificial intelligence Fig.3 describes the coordination of brain and body action together

flow diagram of gait classification using different machine learning techniques

Fig.3: Human GAIT different Phase

Fig.4: Sense- Appraise- Act [30]

to allow learning at different phases [30]. The learning process consist of three memeory part one is perceptual used to perceive the information and store this information of external event, Second one is working arefer as short term and third one is long term memory.. The working memory store the temporarily result. Aftercycle the information store into long term memory as previous experience, this is an entire learning cycle of e

The main objective of gait study to understand the problem of disable person, elder people and used to developed prosthesis leg The classification can use for development of biological inspired bipedal machine which can work similar to human

In clutter environment push is very common experience phenomena, so push recovery [5] is very essential for o behave as human lot of research is taking place to develop the

artificial intelligence into machine which was a the coordination of brain and body action together.. Fig.1 describes the

learning process consist of three memeory part one is perceptual used to perceive the information and store this information of external event, Second one is working also

temporarily result. After process of entire is an entire learning cycle of experience [31].

Fig.5:

1. Methodology Human Tracking and Activity Recognition

To recognize the human activity, we establish the features of each act

Walking feature: It is a common experience phenomena of walking activity

speed in same direction. So, the walking could be

predefined threshold for walking [32-33]. Note that the significant difference between running and walking strides

one of the feet will be in contact with the principal axis (ground) at any given time as shown in Figure

case of jumping activity, every part of human moves only vertically and in the same direction either up or down. Therefore,

jumping action can be identified by the velocities of all the three components to be near or equal to zero

but greater than zero in vertical direction as shown in Figure

running activities were that travelling speed of running is greater than jogging and other difference is

the leg components to the axis of ground as shown in Figure

travelling is greater than jogging and the other difference is of distance ratio between leg compo

shown in Figure 6 (d).Fig.6 is different types of gait for different pattern walking, (b) jumping, (c) jogging and (d) running. It is a

stable gait, which human being perfectly can mimic.

Fig.5: Multi Store Human Memory [31]

I. METHODOLOGY

Activity Recognition

, we establish the features of each activity from the parameters of human model as follows:

t is a common experience phenomena of walking activity, each human body part move generally

could be identified by the velocities of all components superior to zero but lesser than a

. Note that the significant difference between running and walking strides

one of the feet will be in contact with the principal axis (ground) at any given time as shown in Figure

case of jumping activity, every part of human moves only vertically and in the same direction either up or down. Therefore,

jumping action can be identified by the velocities of all the three components to be near or equal to zero

but greater than zero in vertical direction as shown in Figure 6(b). Jogging feature: The only differences between jogging and

running activities were that travelling speed of running is greater than jogging and other difference is

the leg components to the axis of ground as shown in Figure 6(c). Running feature: Similarly in case of running activity, speed of

travelling is greater than jogging and the other difference is of distance ratio between leg components to the axis of ground as

is different types of gait for different pattern walking, (b) jumping, (c) jogging and (d) running. It is a

stable gait, which human being perfectly can mimic.

(a)

(b)

(c)

from the parameters of human model as follows:

each human body part move generally with same

identified by the velocities of all components superior to zero but lesser than a

. Note that the significant difference between running and walking strides is that at least

one of the feet will be in contact with the principal axis (ground) at any given time as shown in Figure 5 (a). Jumping feature: In

case of jumping activity, every part of human moves only vertically and in the same direction either up or down. Therefore,

jumping action can be identified by the velocities of all the three components to be near or equal to zero in horizontal direction

(b). Jogging feature: The only differences between jogging and

running activities were that travelling speed of running is greater than jogging and other difference is of distance ratio between

(c). Running feature: Similarly in case of running activity, speed of

nents to the axis of ground as

is different types of gait for different pattern walking, (b) jumping, (c) jogging and (d) running. It is a

(d)

Fig.6: Silhouette pattern for (a) Walking, (b) Jumping, (c) Jogging and (d) Running [32].

Multi-layer neural networks [34] consist of an input layer, a hidden layer, and an output layer. Multi-layer neural networks can

have several output units. The units of the hidden layer unction as input units to the next layer. However, multiple layers of linear

units still produce only linear functions [35]. The step function in perceptions is another choice, but it is not differentiable, and

therefore not suitable for gradient descent search. The solution is the sigmoid function, a non-linear, differentiable threshold

function [36] [37]. Fig.6 is the model of our multi layer propagation ANN.

Fig.7: Whole process Data fusion(feature level) and classification process

Fig.7 is the process of data fusion at feature level then dimensionality reduction using PCA and its classification using different

machine leaning technique. PCA used for selection of major feature.The units of the hidden layer function as input units to the

next layer. However, multiple layers of linear units still produce only linear functions. The step function in perceptions is another

choice, but it is not differentiable, and therefore not suitable for gradient descent search [38]. The Solution: the sigmoid function,

a non-linear, differentiable threshold function [39]. Fig.7 is the overall model of back propagation based neural network [40].

Vision based GAIT data collection

Sensor based GAIT data collection

Data fusion (Feature level)

Principle Feature Extraction

Physical parameters

Machine Learning

Different GAIT classification

Fig.8: Multi-Layer Back Propagation ANN

Algorithm for Human Activity Recognition 1) Input is fed to the system as a feature of different gait. 2) Select the principal component, which are used for further processing. 3) Reduce the dimensionality. 4) Apply various technique for classification ANN,KNN,K-mean 5) KNN and K-mean used for classification; based on the similarities in the features. 6) The features depend on the following criteria: Walking, Jumping, Jogging and Running. 7) Compare the all type of gait 8) Artificial Neural network (ANN) outperforms the K-nearest neighbor (KNN), K-mean and other existing methods for

classification.

(a) (b)

(c) (d)

Fig.9: Templates of (a) jogging, (b) running, (c) walking, and (d) jumping for human activities [32].

Fig.8 is template of different gait style jogging, running, walking. Algorithm for ANN for GAIT classification [41]: The

algorithm for started with initializing weight of all node, then select the data point and calculate the output for each point,

computer the error and propagate it back. Finally update the weight and keep this loop run till error should be not be below

threshold.

Back propagation Algorithm Neural Networks

Step 1 propagates the input forward through the network

Step 2 propagate the errors backward through the network is similar to the delta rule in gradient descent.

Step 3 sums over the errors of all output units Influence by a given hidden unit (this is because the training data only provides

direct feedback for the output units). A. Classification Methods The main contribution this paper is to classify the gait data into different category based on training of gait data with already known categories and fit the new gait sample, normally we called prediction of new sample category. Principal component analysis (PCA) is used to reduce data dimensionality. PCA converts the high dimensional data into a lower dimension space [44]. PCA ignores the small variability and captures big (principal) variability in the data using higher eigen value.

1) k-Nearest Neighbors The application of k-nearest neighbors (K-NN) reported as GAIT categorization [45]. The idea is to determine the gait

category of given sample gait data based on nearest to it in the gait space and also depend on the categories of k sample gait that are nearest to it. The k-nearest classify the unknown sample on the vote of k nearest data point rather than single neighbor. Each time new sample point joints the cluster the centroids are recomputed of cluster. There are two types of learning method one is stochastic probability based e.g. Bayesian classifier and another one is eager learning the example are neural network and decision tree. The other categories are the K-nearest neighbor and case based reasoning.

Euclidean Distance: it is used for real value attributes and instance x

Distance between two instances xi and xj

𝑑

The KNN algorithm used to classifies class based on the similarityThe K-Means algorithm is a method to cluster objects based on their attributes into k partitions. It assumes that the k clusters exhibit GausIt assumes that the object attributes form a vector space. The objective it tries to achieve is to minimize total intra

Algorithm Fitness Function The K-Means algorithm used to minimize square error of all data The equation of square error (E) for all data point in data set is given by equation 2. Where p is a given data the sum of the square error for all elements in the data set; p is a given element; and m

Algorithm:

1) Select the value of K € Prime && ODD;

Fig.10: Block diagram of Gait Biometric based identification system

Fig.10 is the block diagram of entire process of Gait biometric identification system. It

dimensionality reduction using PCA then created feature vector. Further using different machine learning technique we classifgait data into different category. Which is important component to identify person?

Supervised

it is used for real value attributes and instance x often referred as feature vector. 𝑎 (𝑥), 𝑎 (𝑥), … … . . 𝑎 (𝑥)

𝑥 , 𝑥 = (𝑎 (𝑥 ) − 𝑎 𝑥 ) − (1)

class based on the similarity measure. The K-mean algorithm method to cluster objects based on their attributes into k partitions.

clusters exhibit Gaussian distributions. It assumes that the object attributes form a vector space. The objective it tries to achieve is to minimize total intra-cluster variance.

used to minimize square error of all data point in all cluster. The equation of square error (E) for all data point in data set is given by equation 2. Where p is a given data the sum of the square error for all elements in the data set; p is a given element; and mi is the mean of cluster C

𝐸 = |𝑝 − 𝑚 |

∈

– (2)

€ Prime && ODD;

Block diagram of Gait Biometric based identification system

is the block diagram of entire process of Gait biometric identification system. It starts from feature selection to dimensionality reduction using PCA then created feature vector. Further using different machine learning technique we classif

Which is important component to identify person?

Machine Learning

Supervised Reinforcement Unsupervised

The equation of square error (E) for all data point in data set is given by equation 2. Where p is a given data point: Where E is f cluster Ci

starts from feature selection to dimensionality reduction using PCA then created feature vector. Further using different machine learning technique we classified

Fig.11: Differen

Unsupervised Learning K-mean algorithm: It is used for exact clustering. The Knumerical data. Code Step:

Input: Features F= {𝑓 , 𝑓 , 𝑓 , … . . , 𝑓 } where C =number of cluster head T=Threshold for convergence

BEGIN Assume vector mean [𝑚 , 𝑚 , … … . , 𝑚

REPEAT /*until convergence criteria not Recomputed the mean vector for t=t+1 and classify accordingly Update t=t+1; Calculate the Euclidean distance || 𝑚(𝑡

Now m (t) as solution Else Go to repeat END LOOP UNTILL gait data is not classified/*categoEND

Different machine Learning Technique Classification.

II. IMPLEMENTATION

clustering. The K-mean is very sensitive for noise and outlier

} where 𝐹 ∈ 𝑅

] at t=0, c is no of cluster head. /*until convergence criteria not satisfied threshold do not reached*/

Recomputed the mean vector for t=t+1 and classify accordingly

(𝑡) − 𝑚(𝑡 − 1) || < 𝑇

category of gait classification = {normal, croach1, croach2, croach3,

and outlier. We used k-mean for

croach2, croach3, croach4}*/

Fig.12:

Algorithm Gradient-descent (I, Iteration Count) Initialize all weights

For i=1:1: Iterations Count do select a data point 𝐷 = [

set learning rate α € [0,1] calculate outputs o (p ) for each calculate Error δij using backpropagation update all weights (in parallel) wij wij - α * Θij * x end for update all threshold Θij (in parallel) return weights w end

EXPERIMENT

The ideal locomotion have a exactly sinusoidal curve i.e. oscillatory motion at hip and two sharp humps in knee and two sharphump like mountain at ankle joint during normal walk without any external ankle.

Fig.14: Ideal curve during normal walk

0 20 40 60 80 100 120 140 160 180 200-10

0

10

20

30

40

Gait Cycle

H I

P (

deg

ree

)

0

-60

-40

-20

0

Kn

ee (

Deg

ree

)

Fig.12: Working of the GAIT Classification

Fig.13: Gait Classification System

[𝑖, 𝑗] from matrix D € [0,1]

) for each data point backpropagation

all weights (in parallel) * xj(k-1)

XPERIMENT RESULT AND PERFORMANCE

The ideal locomotion have a exactly sinusoidal curve i.e. oscillatory motion at hip and two sharp humps in knee and two sharphump like mountain at ankle joint during normal walk without any external perturbation. Fig. 14 is ideal curves for hip, knee and

Ideal curve during normal walk. (a) hip, (b) knee and (c) ankle respectively

20 40 60 80 100 120 140 160 180 200 220Gait Cycle

0 20 40 60 80 100-15

-10

-5

0

5

10

15

Gait Cycle

An

kle

(de

gre

e)

The ideal locomotion have a exactly sinusoidal curve i.e. oscillatory motion at hip and two sharp humps in knee and two sharp ideal curves for hip, knee and

. (a) hip, (b) knee and (c) ankle respectively

100 120 140 160 180 200 220Gait Cycle

Human Gait dataset:

Dataset : BUAA_IRIP and OpenSim. D = {Normal, Crouch2, Crouch3, Crouch4}, D€ Rd

This research work reviews the all main approach of gait classification. Here we performed the classification using the different machine learning technique i.e. KNN and K-mean for GAIT data [46]. The performance table shows that the ANN is better performing technique. [47]. Table 6 is accuracy table. The PCA used for dimensionality reduction i.e. it is used to convert the higher dimension data into lower dimension space. The PCA which select the only prominent feature vector for K-Mean and KNN [48].

Table.2. Gait Data for subject 1 normal walk

Right Hip Left Knee Left Ankle Left hip Left Knee Left Ankle 37.5 -3.97 -2.14 -4.5 -13.86 10

37.2 -7 -3.7 -4.2 -16.97 8.5

36.9 -10.52 -4.8 -2.6 -20.96 6

36.2 -14.12 -4.8 0.1 -26 0

35.7 -17.38 -3.7 2.1 -32.03 -5

34.8 -19.84 -2.14 4.1 -38.74 -9

33.5 -21.27 -1 6.8 -45.6 -12.7

30.7 -21.67 0.6 10.2 -52.05 -13.5

27.4 -21.22 1.8 14 -57.54 -12.7

25.3 -20.2 2.8 18.5 -61.66 -10

23 -18.86 3.9 22 -64.12 -7.1

21 -17.35 4.5 25 -64.86 -4.1

18 -15.73 5.2 27.8 -63.95 -2.1

16 -14.08 5.7 30.22 -61.59 0

13.8 -12.5 5.9 33 -57.97 1.3

12.4 -11.09 6.2 35 -53.27 2.2

10.8 -9.91 6.8 36.4 -47.58 2.8

8.8 -8.97 7.5 37.8 -40.94 3.1

6.9 -8.28 8.2 38.3 -33.46 2.8

5 -7.86 8.9 38.2 -25.38 2.3

3 -7.72 9.8 37.8 -17.27 1.5

1.2 -7.94 10.5 37.5 -9.94 0.5

0 -8.6 11.2 37.2 -4.31 0.2

-2.4 -9.76 11.5 36.8 -1.12 0

-4.2 -11.5 10.9 37.2 -0.54 0

-4.5 -13.86 10 37.5 -3.97 -2.14

-4.2 -16.97 8.5 37.2 -7 -3.7

-2.6 -20.96 6 36.9 -10.52 -4.8

0.1 -26 0 36.2 -14.12 -4.8

2.1 -32.03 -5 35.7 -17.38 -3.7

4.1 -38.74 -9 34.8 -19.84 -2.14

6.8 -45.6 -12.7 33.5 -21.27 -1

10.2 -52.05 -13.5 30.7 -21.67 0.6

14 -57.54 -12.7 27.4 -21.22 1.8

18.5 -61.66 -10 25.3 -20.2 2.8

22 -64.12 -7.1 23 -18.86 3.9

25 -64.86 -4.1 21 -17.35 4.5

27.8 -63.95 -2.1 18 -15.73 5.2

30.22 -61.59 0 16 -14.08 5.7

33 -57.97 1.3 13.8 -12.5 5.9

35 -53.27 2.2 12.4 -11.09 6.2

36.4 -47.58 2.8 10.8 -9.91 6.8

37.8 -40.94 3.1 8.8 -8.97 7.5

38.3 -33.46 2.8 6.9 -8.28 8.2

38.2 -25.38 2.3 5 -7.86 8.9

37.8 -17.27 1.5 3 -7.72 9.8

37.5 -9.94 0.5 1.2 -7.94 10.5

37.2 -4.31 0.2 0 -8.6 11.2

36.8 -1.12 0 -2.4 -9.76 11.5

37.2 -0.54 0 -4.2 -11.5 10.9

37.5 -2.21 0 -4.5 -13.86 10

We used OpenSim data set for Gait of four categories Normal, Crouch2, Crouch3 and Crouch4. We used 30 samples of each

category of data for training and 20 data set points for testing. The ANN is applied for classification and the result achieved is

given in table2. D = {Normal, Crouch2, Crouch3, Crouch4}, D€ Rd , D dimension data set of each class and training data 30 data

point and testing data point 20.

Fig.15: Multi-Layer Back Propagation ANN

Fig 9 (a) is the performance curve which is related to training of model, when the mean square error will reach up to threshold

in which no of epoch. As we set the threshold .001. We achieved the target in 11 epochs.

(a)

(b)

Fig. 9 (b) is training state of our model and Fig.9 (c) is the regression model of data fitting.

0 1 2 3 4 5 6 7 8 9 1 0 1 1

1 0-1 0

1 0-8

1 0-6

1 0-4

1 0-2

1 00

1 02

B e s t V a lid a tio n P e r fo rm a n c e is 2 9 8 .5 8 1 4 a t e p o c h 5

Me

an

Sq

ua

red

Err

or

(m

se)

1 1 E p o c h s

Tra in

V a l id a t io n

Te s tB e s t

G o a l

1 01

1 02

1 03

grad

ient

G rad ie n t = 4 4 .9 8 2 4 , a t e p o c h 6

1 0-3

1 0-2

mu

M u = 0 . 0 0 1 , a t e p o c h 6

0 1 2 3 4 5 60

5

1 0

val f

ail

6 E p o c h s

V a lid a t io n C h e c k s = 6 , a t e p o c h 6

(c)

Fig.16: A)- performance curve mean square error B)- Training State c)- Regression

Performance measure: Fig.17 is the accuracy rate (percentage) of gait classifications using K-mean where k=1. We calculated error i.e. total misclassification rate using formula (3).

Error (Total Misclassi ication rate) = total misclassi ied/ total sample test (3)

Fig.17: The Accuracy classification rate of Different GAIT pattern using K-mean

Fig.18 shows the accuracy bar chart of gait classification using KNN for different majority value of K. Fig. 19 is the classification of normal gait using KNN with different majority vote.

-4 0 -2 0 0 2 0

-5 0

-4 0

-3 0

-2 0

-1 0

0

1 0

2 0

T a rg e tO

utp

ut

~=

0.8

6*T

arg

et

+ -0

.91

T ra in in g : R = 0 .7 8 9 9 7

D a ta

F i tY = T

-4 0 -2 0 0 2 0

-5 0

-4 0

-3 0

-2 0

-1 0

0

1 0

2 0

T a rg e t

Ou

tpu

t ~

= 0

.22

*Ta

rge

t +

-0.2

6

V a lid a tio n : R = 0 .1 7 8 7 8

D a t a

F itY = T

-4 0 -2 0 0 2 0

-5 0

-4 0

-3 0

-2 0

-1 0

0

1 0

2 0

T a rg e t

Ou

tpu

t ~

= 0

.93

*Tar

ge

t +

-0.

02

T e s t: R = 0 .9 6 8 6 5

D a t a

F itY = T

-4 0 -2 0 0 2 0

-5 0

-4 0

-3 0

-2 0

-1 0

0

1 0

2 0

T a rg e tO

utp

ut

~=

0.8

7*T

arg

et

+ -

0.7

1

A ll: R = 0 .8 3 9 1 8

D a t a

F itY = T

C R O U C H_2 C R O U C H_3 C R O U C H_4 N O R M A L0

20

40

60

80

100

A C

C U

R A

C Y

K=1

Fig.18: Accuracy curve of Gait classification using KNN

Fig.19: Normal classification of humanoid gait using for different majority count.

The graph represent the normal data is better classification rate. The gradient descent based ANN used for classification. The

above table shows the ANN based classification technique outperform all the previously existed classification technique. The

ANN based model propagates the input forward through the network and propagate the errors backward through the network is

similar to the delta rule in gradient descent. Finally the sums over the errors of all output units Influence by a given hidden unit

(this is because the training data only provides direct feedback for the output which information can used in surveillance and

further used to identify upcoming disease. The whole research work is the classification of GAIT into following four categories:

Normal, crouch2, crouch3, crouch4 using ANN. The result shown in figure describes that ANN is better performer other

machine learning technique. In the flow diagram it is shown that for a user based query to recognize and identify a human gait. It

can be easily done by indexing the gait data in the database. In next step, the matching between the feature extracted by the

system and that available in the database is taken into account for human gait classification.

Performance Matrix

The main performance indicator for any classification or biometric identification system is (receiver operating characteristic)

ROC. It is basically, the curve of true acceptance rate (TAR) against false acceptance rate (FAR), which is the measure of no of

false instance classified as positive among all intruder and imposter cases.

FRR (False Rejection Rate) - The probability of the legitimate claim when biometric system will fail to identify. It is a statistic

biometric performance during verification task.

TRR (True Reject Rate) - Biometric performance in verification task. The percentage of times a system correctly rejects a false

claim of identity.

TAR (True Acceptance Rate) – It is count of true claim of identity when a system correctly verifies.

FAR (False Acceptance Rate) – The false acceptance percentage of system..

In the biometric literature, FAR is sometimes defined such that the "impostor" makes zero effort to obtain a match.

TAR=1-FRR

C R O U C H_2 C R O U C H_3 C R O U C H_4 N O R M A L0

20

40

60

80

100

A C

C U

R A

C Y

k=1k=2k=3k=4k=5

K=1 K=2 K=3 K=4 K=5 K=6 K=7 K=8 K=90

10

20

30

40

50

60

70

A C

C U

R A

C Y

N O O F N e i g h b o r ( m a j o r i t y v o t e)

Where TAR-true acceptance rate and FRR-false rejection rate

To verify the result of biometric system we have four matrix terms True Acceptance Rate (FAR), True Rejection Rate (FAR),

False Acceptance Rate (FAR), False Rejection Rate (FAR).

Verification results are reported in terms of the True Acceptance Rate (TAR), False Accept Rate (FAR), and ROC. The TAR is

measured as the number of occurrences when genuine biometric identity is matched correctly, whereas, FAR is the measurement

of the number of occurrences when imposter or intruder identity is matched falsely. EER is the point where FAR and FRR are

equal, where FRR is the False Reject Rate and measured on the basis of number of false rejections of genuine matches and also

given as,

FRR = 1 – TAR

ROC = TAR vs FAR

Table.3. Confusion matrix for K-Mean{ Data Size 30 training and 20 testing }

Normal Crouch2 Crouch3 Crouch4 TAR

Normal 17 0 2 1 17/20=0.85

Crouch2 0 17 3 0 17/20=0.85

Crouch3 1 2 15 2 15/20=0.75

Crouch4 0 1 0 19 19/20=0.95

FAR 1/18=.055 3/20=0.15 5/20=0.25 3/22=0.13

𝑇𝐴𝑅 =

1720

+1720

+1520

+1920

4× 100 = 85%

FRR=1-TAR

𝐹𝐴𝑅 =

118

+3

20+

520

+3

224

× 100 = 14.79%

Table.4. Confusion matrix for KNN (K=1) {Data Size 30 training and 20 testing }

Normal Crouch2 Crouch3 Crouch4 TAR

Normal 19 1 0 0 19/20=0.95

Crouch2 0 17 3 0 17/20=0.85

Crouch3 1 2 17 0 17/20=0.85

Crouch4 0 1 0 19 19/20=0.95

FAR 1/20=0.05 4/21=0.19 3/20=0.15 0

𝑇𝐴𝑅 =

1920

+1720

+1720

+1920

4× 100 = 90%

FRR=1-TAR

𝐹𝐴𝑅 =

120

+4

21+

320

4× 100 = 39.04%

Table.5. Confusion matrix for ANN {Data Size 30 training and 20 testing}

Normal Crouch2 Crouch3 Crouch4 TAR

Normal 20 0 0 0 20/20=1

Crouch2 1 19 0 0 19/20=0.95

Crouch3 1 1 18 0 18/20=0.90

Crouch4 2 0 1 17 17/20=0.85

FAR 4/24=0.166 1/20=0.05 1/19=0.52 0

𝑇𝐴𝑅 =

2020

+1920

+1820

+1720

4× 100 = 92.5%

FRR=1-TAR

𝐹𝐴𝑅 =

424

+1

20+

119

4× 100 = 6.73%

Fig.20: ROC curve

Table 6. is the accuracy table which showing the ANN is better except the case of crouch 4 gait as it is more complex one gait so exact feature selection and Classification in K-mean is better.

Table.6. The accuracy percentage of Different GAIT data classiifcation.

Method K-Mean KNN(k=1,2,3,4,5) ANN

Normal Gait 74.68 100,100,100,100,100 100 Crouch 2 Gait 51.43 50,50,50,55,55 60 Crouch 3 Gait 33.19 50,50,50,50,50 50 Crouch 4 Gait 81.80 50,40,35,25,15 55

III. CONCLUSION

In this research a new gait identification method has been proposed. Using proposed methodology the practicability of the gait identification can be improved by more accurate spatio-temporal modeling. Extensive simulation we have shown that with more robust feature extraction techniques. The classification rate and ARA (activity reorganization activity) improved substantially. Identification of abnormality, disorder and upcoming disease at the primary stage has been the main concern at this research. The experimental results demonstrated that the proposed method could accurately recognize different activities in both indoor and outdoor scenarios while maintaining a high recognition accuracy rate. The novel classification could classify GAITs effectively into four different classes Normal, crouch-2, crouch-3 and crouch-4. The result has been compared with KNN and K-mean algorithm and it has been shown that our classified outperform the existing classifiers.

0.75 0.8 0.85 0.9 0.95 10

0.2

0.4

0.6

TAR

FA

R

ROC c u r v e

K-mean

KNN

ANN

ACKNOWLEDGMENTS

The authors would like to acknowledge the support of the faculty, students and Research Development team at Robita & AI Lab IIIT Allahabad. The authors would also like to thank master student for contribution in to biometric Gait data collection for this research work.

REFERENCES

[1] Kale, A., Sundaresan, A., Rajagopalan, A. N., Cuntoor, N. P., Roy-Chowdhury, A. K., Kruger, V., & Chellappa, R. (2004). Identification of humans using gait IEEE Transactions on Image Processing, 13(9), 1163-1173.

[2] Daniel Zapico Palacio, Victor Hugo Medina García, Oscar Sanjuán Martínez, Rubén González Crespo: “Generating Digital Fingerprints using Artificial Neural Networks”. IC-AI 2008: 907-912.

[3] Sajid Iqbal, Xizhe Zang, Yanhe Zhu, Jie Zhao, Bifurcations and chaos in passive dynamic walking: A review, Robotics and Autonomous Systems, Volume 62, Issue 6, June 2014, Pages 889-909.

[4] Wang, L., Tan, T., Ning, H., & Hu, W. (2003). Silhouette analysis-based gait recognition for human identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(12), 1505-1518.

[5] Semwal, V.B.; Bhushan, A.; Nandi, G.C., "Study of humanoid Push recovery based on experiments," Control, Automation, Robotics and Embedded Systems (CARE), 2013 International Conference on , vol., no., pp.1,6, 16- 18 Dec. 2013.

[6] Hu, M., Wang, Y., Zhang, Z., & Zhang, D. (2011). Gait-based gender classification using mixed conditional random field. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, 41(5), 1429-1439.

[7] Huang, G., & Wang, Y. (2007). Gender classification based on fusion of multi-view gait sequences. In Springer Computer Vision–ACCV 2007 (pp. 462-471).

[8] Wang, L., Tan, T., Hu, W., & Ning, H. (2003). Automatic gait recognition based on statistical shape analysis. IEEE Transactions on Image Processing, 12(9), 1120-1131.

[9] Wang, C., Zhang, J., Wang, L., Pu, J., & Yuan, X. (2012). Human identification using temporal information preserving gait template. IEEE Transactions on, Pattern Analysis and Machine Intelligence, 4(11), 2164-2176.

[10] Joachim Steigenberger, Carsten Behn, Gait generation considering dynamics for artificial segmented worms, Robotics and Autonomous Systems, Volume 59, Issues 7–8, July–August 2011, Pages 555-562.

[11] Semwal, V.B.; Katiyar, Shiv A.; Chakraborty, P.; Nandi, G.C., "Biped model based on human Gait pattern parameters for sagittal plane movement," Control, Automation, Robotics and Embedded Systems (CARE), 2013 International Conference on , vol., no., pp.1,5, 16-18 Dec. 2013.

[12] Wang, Liang, Huazhong Ning, Tieniu Tan, and Weiming Hu. "Fusion of static and dynamic body biometrics for gait recognition." Circuits and Systems for Video Technology, IEEE Transactions on 14, no. 2 (2004): 149-158.

[13] Argall, Brenna D., et al. "A survey of robot learning from demonstration."Robotics and autonomous systems 57.5 (2009): 469-483. [14] Bobick, A. F., & Davis, J. W. (2001). The recognition of human movement using temporal templates. Pattern Analysis and Machine

Intelligence, IEEE Transactions on, 23(3), 257-267. [15] Vega, I. R., & Sarkar, S. (2003). Statistical motion model based on the change of feature relationships: human gait-based

recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(10), 1323-1328. [16] Wang, L., Ning, H., Tan, T., & Hu, W. (2004). Fusion of static and dynamic body biometrics for gait recognition. IEEE Transactions

on Circuits and Systems for Video Technology, 14(2), 149-158. [17] Noorit, N., Suvonvorn, N., & Karnchanadecha, M. (2010). Model-based human action recognition. In Second International

Conference on Digital Image Processing (pp. 75460P-75460P). International Society for Optics and Photonics. [18] Ekinci, M. (2006). Human identification using gait. Turk J Elec Engin, 14(2), 267-291. [19] M. Hobon, N. Lakbakbi Elyaaqoubi, G. Abba, Quasi optimal sagittal gait of a biped robot with a new structure of knee joint,

Robotics and Autonomous Systems, Volume 62, Issue 4, April 2014, Pages 436-445. [20] Chen, L., Wang, Y., Wang, Y., & Zhang, D. (2009). Gender recognition from gait using radon transform and relevant component

analysis. In Springer Emerging Intelligent Computing Technology and Applications (pp. 92-101). [21] Bartlett, M. S., Movellan, J. R., & Sejnowski, T. J. (2002). Face recognition by independent component analysis. , IEEE Transactions

on Neural Networks13(6), 1450-1464. [22] Semwal, Vijay Bhaskar, Neha Gaud, and G. C. Nandi. "Human Gait State Prediction Using Cellular Automata and Classification

Using ELM." MISP-2017. [23] Sepulveda, F., Wells, D. M., & Vaughan, C. L. (1993). A neural network representation of electromyography and joint dynamics in

human gait Journal of biomechanics, 26(2), 101-109. [24] Semwal, Vijay Bhaskar, et al. "An optimized feature selection technique based on incremental feature analysis for bio-metric gait

data classification." Multimedia tools and applications76.22 (2017): 24457-24475. [25] Selim, S. Z., & Ismail, M. A. (1984). K-means-type algorithms: a generalized convergence theorem and characterization of local

optimality. , IEEE Transactions on Pattern Analysis and Machine Intelligence, (1), 81-87. [26] Fong, Terrence, Illah Nourbakhsh, and Kerstin Dautenhahn. "A survey of socially interactive robots." Robotics and autonomous

systems 42.3 (2003): 143-166. [27] F.; Yamakita, M.; Kamamichi, N.; Zhi-Wei Luo, "A novel gait generation for biped walking robots based on mechanical energy

constraint," IEEE Transactions on Robotics and Automation, , vol.20, no.3, pp.565,573, June 2004. [28] Lawrence, S., Giles, C. L., Tsoi, A. C., & Back, A. D. (1997). Face recognition: A convolutional neural-network approach. IEEE

Transactions on Neural Networks, 8(1), 98-113. [29] M. Hoffmann, K. Štěpánová, M. Reinstein, The effect of motor action and different sensory modalities on terrain classification in a

quadruped robot running with multiple gaits, Robotics and Autonomous Systems, Available online 24 July 2014. [30] Xavier Perrin, Ricardo Chavarriaga, Francis Colas, Roland Siegwart, José del R. Millán, Brain-coupled interaction for semi-

autonomous navigation of an assistive robot, Robotics and Autonomous Systems, Volume 58, Issue 12, 31 December 2010, Pages 1246-1255.

[31] Cristina P. Santos, Vítor Matos, Gait transition and modulation in a quadruped robot: A brainstem-like modulation approach, Robotics and Autonomous Systems, Volume 59, Issue 9, September 2011, Pages 620-634.

[32] Gupta, Jay Prakash, et al. "Human Activity Recognition using Gait Pattern."International Journal of Computer Vision and Image Processing (IJCVIP) 3.3 (2013): 31-53.

[33] Shahid Hussain, Sheng Q. Xie, Prashant K. Jamwal, Control of a robotic orthosis for gait rehabilitation, Robotics and Autonomous Systems, Volume 61, Issue 9, September 2013, Pages 911-919.

[34] Su-Bin Joo, Seung Eel Oh, Taeyong Sim, Hyunggun Kim, Chang Hyun Choi, Hyeran Koo, Joung Hwan Mun, “Prediction of gait speed from plantar pressure using artificial neural networks”, Expert Systems with Applications, Volume 41, Issue 16, 15 November 2014, Pages 7398-7405.

[35] Lee, Heesung, Heejin Lee, and Euntai Kim. "A new gait recognition system based on hierarchical fair competition-based parallel genetic algorithm and selective neural network ensemble." International Journal of Control, Automation and Systems 12.1 (2014): 202-207.

[36] Semwal, Vijay Bhaskar. "Data Driven Computational Model for Bipedal Walking and Push Recovery." arXiv preprint arXiv:1710.06548 (2017).

[37] Sabourin, Christophe, Weiwei Yu, and Kurosh Madani. "Gait Pattern Based on CMAC Neural Network for Robotic Applications." Neural processing letters 38.2 (2013): 261-279.

[38] Semwal, Vijay Bhaskar, and Gora Chand Nandi. "Toward developing a computational model for bipedal push recovery–a brief." Sensors Journal, IEEE 15.4 (2015): 2021-2022.

[39] Shibendu Shekhar Roy, Dilip Kumar Pratihar, Effects of turning gait parameters on energy consumption and stability of a six-legged walking robot, Robotics and Autonomous Systems, Volume 60, Issue 1, January 2012, Pages 72-82.

[40] Semwal, Vijay Bhaskar, et al. "Design of Vector Field for Different Subphases of Gait and Regeneration of Gait Pattern." IEEE Transactions on Automation Science and Engineering (2016).

[41] Pedro Silva, Cristina P. Santos, Vítor Matos, Lino Costa, “Automatic generation of biped locomotion controllers using genetic programming”, Robotics and Autonomous Systems, Available online 5 June 2014.

[42] Brian R. Duffy, “Anthropomorphism and the social robot”, Robotics and Autonomous Systems, Volume 42, Issues 3–4, 31 March 2003, Pages 177-190.

[43] Chris Iverach-Brereton, Jacky Baltes, John Anderson, Andrew Winton, Diana Carrier, Gait design for an ice skating humanoid robot, Robotics and Autonomous Systems, Volume 62, Issue 3, March 2014, Pages 306-318.

[44] Semwal, Vijay Bhaskar, and Gora Chand Nandi. "Generation of joint trajectories using hybrid automate-based model: a rocking block-based approach." IEEE Sensors Journal 16.14 (2016): 5805-5816.

[45] Vijay Bhaskar Semwal, Pavan Chakraborty, G.C. Nandi, Less computationally intensive fuzzy logic (type-1)-based controller for humanoid push recovery, Robotics and Autonomous Systems, Available online 16 September 2014.

[46] Huang, P. S., C. J. Harris, and Mark S. Nixon. "Human gait recognition in canonical space using temporal templates." IEE Proceedings-Vision, Image and Signal Processing 146.2 (1999): 93-100.

[47] Cunado, David, Mark S. Nixon, and John N. Carter. "Using gait as a biometric, via phase-weighted magnitude spectra." Audio-and Video-based Biometric Person Authentication. Springer Berlin Heidelberg, 1997.

[48] Nixon, Mark S., Tieniu Tan, and Rama Chellappa. Human identification based on gait. Vol. 4. Springer, 2010.

Authors Biographics Vijay Bhaskar Semwal obtained his B.Tech. from the College of Engineering Roorkee, Roorkee, in 2008. He received his M.Tech. from IIIT Allahabad in 2010. Currently, he is pursuing a Ph.D. from IIIT Allahabad. Before becoming a research scholar at IIIT Allahabad, he worked as a Senior System Engineer (R&D) with Siemens Gurgaon and Bangalore. He has worked for various major organizations, such as Siemens AG and Newgen. He qualified in the prestigious GATE exam, a record eight times in computer science and achieved 145 AIR in 2008. He is also a Sun-certified java programmer. Currently he is serving as vice chair for IEEE student branch of IIIT-Allahabad. For 2013-14 he served as publicity committee mentor for IEEE student branch IIIT Allahabad. His research interests are machine learning, evolutionary algorithms, analysis of biped locomotion and humanoid push recovery, artificial intelligence, design & analysis of algorithms, biometric identification, Brain wave based authentication.

Manish Raj graduated from Ideal Institute of technology Ghaziabad with Electronics and Communication Engineering. He obtained his M.Tech Degree from Indian Institute of Information Technology, Allahabad with second rank in Robotics &AI department. Currently he is perusing Ph.D. from IIIT Allahabad. His research interests are Biped Locomotion, Control System, Non linear dynamics, Humanoid Robotics, chaos and fractal analysis, Artificial Intelligence, Soft Computing, Artificial Life Simulation, Mathematical Foundation of Robotics, and Hybrid System.

Prof.G.C.Nandi graduated from Indian Institute of Engineering, Science & Technology (Formerly Bengal Engineering College, Sibpore, Calcutta University), in 1984 and post graduated from Jadavpur University, Calcutta in 1986. He obtained his PhD degree from Russian Academy of Sciences, Moscow in 1992. He was awarded National Scholarship by Ministry of Human Resource Development (MHRD), Govt of India in 1977 and Doctoral Fellowship by External Scholarship Division, MHRD, Govt. of India in 1988. During 1997 he was visiting research scientist at the Chinese University at Hong Kong and he was also visiting Faculty with Institute for Software Research, School of Computer Sciences, Carnegie Mellon University, USA, ( 2010- 2011). Currently, he is serving as the senior most Professor and Dean of Academic Affairs of Indian Institute of Information Technology, Allahabad. From January to July 2014, he served as the Director-in-Charge of Indian Institute of Information Technology, Allahabad. Professor Nandi is the Senior Member of ACM, Senior member of IEEE, Chairman, ACM-IIIT-Allahabad

Professional Chapter, (2009-2010), Chartered Member of Institute of Engineers (India),Member of DST (Department of Science and Technology, Govt. of India) Program Advisory Committee member of Robotics, Mechanical and Manufacturing Engineering. He has published more than 100 papers in the various refereed journals and international conferences. His research interest includes robotics specially biped locomotion control & humanoid push recovery, artificial intelligence, soft computing and computer controlled systems.