KR20100084066A - Gait recognition using contourlet transform and fuzzy lda - Google Patents

Abstract

PURPOSE: A method for recognizing a gait using coutourlet transform, fuzzy linear discriminate analysis, and fused algorithm are provided to extract an energy image based on a silhouette from a gait image and apply coutourlet and fuzzy-LDA to the image. CONSTITUTION: A silhouette based energy image is extracted from a gait image. The gait image is divided by a directional component by a coutourlet. A feature vector is generated from a sub image divided by a directional component using a fuzzy LDA(Linear Discriminate Analysis). Final gait recognition is performed by a fused method which effectively fuses matching maps generated by each band.

Description

Gait Recognition using Contourlet Transform and Fuzzy LDA}

The present invention proposes a personal identification algorithm using contour and fuzzy-LDA after extracting silhouette-based energy images from gait images. The proposed method first obtains the energy of silhouette image through gait periodic analysis. In the next step, the gait image is divided by the direction components by the contour, and the feature vectors are calculated from the sub-images divided by the direction components using the fuzzy-LDA technique. In the final step, personal identification is performed by using a fusion technique that can effectively fuse the matching degree calculated for each band. The developed algorithm is applied to various gait image data to show its validity.

Recently, researches on authentication systems using biometrics technology have been actively conducted. Existing biometric technologies such as face, fingerprint and iris have already been studied and put into practical use. However, such a conventional biometric method not only requires an artificial environment for recognition, but also has a problem in that an object to be recognized is forced to a specific action for recognition. In the case of face recognition, there is a problem that the recognition rate is greatly lowered for the long distance image. In the case of iris recognition or fingerprint recognition, the subject to be recognized should touch the sensor or the eye close to the sensor. However, in the case of gait recognition, it is possible to recognize the subject only if it is photographed by the camera. In addition, unlike other biometrics, the object to be recognized is a technology that can be recognized without any force for recognition. In addition, the interest in gait recognition has increased significantly in recent years due to the increase in surveillance cameras and the recognition of the remote image.

The proposed gait recognition technique is a feature that can represent the characteristics of each individual. It is a model-based method that analyzes motion information, optical flow of gait images, angle of legs, height and stride length, and uses it for recognition. And gait can be largely classified into a non-model-based method using the entire silhouette image as a pattern.

The present invention proposes a personal identification system using contour and fuzzy-LDA after extracting silhouette-based energy images from gait images. The contourlet applied in the present invention divides an image by frequency band using a multiscale concept, such as a two-dimensional wavelet transform, and then, by band and in the divided region using a directional filter bank. Images are obtained for each directional component. Contourlet represented by such multi-resolution and multi-directional information has an advantage that can effectively express the smooth contour (smooth contour) that is the main feature of the natural image compared to the two-dimensional wavelet. Therefore, in the present invention, after the silhouette image of the gait is obtained, the gait image is divided by the direction components using the above-described contourlet, and then a feature vector is obtained from the sub-images divided by the direction components using the fuzzy-LDA technique. In the final stage, gait recognition is implemented using a fusion technique that can effectively fuse the matching degree calculated for each band.

An object of the present invention is to obtain the energy of the silhouette image through gait periodic analysis and to divide the gait image by the direction component by the contour, and then to the feature vector in the sub-image divided by the direction component using the fuzzy-LDA technique. In addition, the present invention provides a gait recognition method having excellent gait recognition by providing gait recognition performed by a fusion algorithm that can effectively fuse the matching degree calculated for each band.

In the present invention, the contour transform and the fuzzy linear discrimination analysis technique are applied as the main means for implementing the personal identification algorithm having excellent performance by using the gait exhibiting unique characteristics for each individual. Looking at the means for solving the above problems in the present invention, the step of extracting a silhouette-based energy image from the gait image; Dividing the gait image by direction components by a contour; Calculating a feature vector from sub-images divided by direction components using a fuzzy-LDA technique; A final step is recognized by a fusion method that can effectively fuse the matching degree calculated for each band.

According to the present invention, a contour transform that can be decomposed by direction components for gait feature extraction and a fuzzy linear discriminant analysis method applied for feature extraction for each decomposed direction component, and also for each direction component for final recognition of gait There is an effect of providing a recognition method of excellent performance by providing a recognition method performed by a fusion algorithm that can effectively fuse the calculated similarity.

(1) extracting a silhouette-based energy image from the gait image;

(2) dividing the gait image by direction components by a contour;

(3) calculating a feature vector from sub-images divided by direction components using fuzzy-LDA technique;

(4) the step of the final step recognition by the fusion method that can effectively fuse the matching degree calculated for each band;

The present invention provides a gait recognition method using a contour transform and fuzzy linear discriminant analysis technique.

1. Gait Energy Video Extraction

The gait energy image is obtained from the average of the silhouette images during one gait cycle. Even if the silhouette image has the same person and the same environment, the silhouette image has a different size according to the photographed distance. The silhouette image is resized to a constant image size to reduce the change according to the distance. In addition, to obtain a gait energy image, the period of stride is required, and the stride length information is determined by the silhouette width of the lower part of the silhouette, and one period means from the maximum of the stride to the next maximum. do. Based on the information obtained, N _Gait is the total number of frames in the Nth step, and S _N = {S _N (1), ..., S _N (N _Gait )}. The gait energy image GEI is expressed as in Equation 1 below.

(식 1)

(Equation 1)

In addition, when generating gait energy images, the sum of all images is obtained based on the X-axis of the center of gravity of the upper body to generate a feature image that is robust to changes in hands and insteps during the gait cycle. 1 shows a gait energy image obtained by the above method. As shown in FIG. 1, it includes all of the static features related to the body shape of a person and the dynamic features related to physical exercise such as walking habits and speed.

2. On the contour  by Energy  Division

개의 부대역(subband)을 생성하는

-레벨 이진 트리 분해(binary tree decomposition)를 통하여 효과적으로 구현할 수 있다. 여기서, 방향성 필터뱅크는

이며,

개의 실제 쐐기 모양의 주파수 대역이 존재한다. 부대역 0-3은 대부분 수평방향에 대응하는 반면, 부대역 4-7은 대부분 수직방향에 대응한다.The contourlet divides the image by frequency band using a multiscale concept, such as 2D wavelet transform, and then obtains the directional information of the image in the divided region using a directional filter bank. Contourlets represented by such multi-resolution and multi-directional information can effectively express smooth contours, which are the main features of natural images, compared to two-dimensional wavelets. 2 illustrates a directional filter used to construct a contour filter bank. Directional filters have a wedge shaped frequency division as shown in Figure 2.

Creating subbands

It can be implemented effectively through level binary tree decomposition. Here, the directional filter bank

,

There are four actual wedge shaped frequency bands. Subbands 0-3 mostly correspond to the horizontal direction, while subbands 4-7 mostly correspond to the vertical direction.

3. Feature Extraction by Fuzzy Linear Discriminant Analysis

Since the gait energy image is represented by a very high dimensional pattern, it is required to be represented by low dimensional data for feature extraction and classification. LDA is a linear transformation method that maximizes the ratio between a class within a class (Within-Scatter Matrix) and a class between class (Between-Scatter Matrix). Principal Component Analysis is a low-dimensional feature in image space. Based on linear projection into space, it finds the projection direction that maximizes all gait energy images in the entire database. However, since unwanted changes caused by changes in lighting conditions and external environment are also included, Principal Component Analysis is the best method in terms of reconstruction from low-order basis vectors, but LDA for identification and recognition of images with changes in lighting or external environment. The technique shows excellent recognition performance. In the present invention, the contoured transformed gait energy image is reduced to a low dimension by applying a principal component analysis method, and then a feature vector is extracted using fuzzy-LDA.

는 연속적인 행들의 연결로 의 벡터로서 고려된다. 그러면 N개 걸음걸이 영상의 학습 집합은

로 표시된다. 우선, 주성분 분석기법에 의한 축소된 걸음걸이 영상의 특징값을 구하기 위한 공분산 행렬을 정의하면 식 2와 같다.The fuzzy-LDA algorithm applied in the present invention to extract the features of gait energy image (GEI) is reduced the input dimension of the GEI by Principal Component Analysis (PCA), and then the reduced feature vector and fuzzy Calculation using the degree function. Assume that the image is a two-dimensional array (n × n) with respective pixel values. video

Is considered as a vector of concatenations of consecutive rows. Then the training set of the N steps video is

Is displayed. First, a covariance matrix for obtaining feature values of a reduced gait image by the principal component analysis method is defined as in Equation 2.

(식 2)

(Equation 2)

(식 3)

(Equation 3)

라 하자. 영상의 집합 z에 대해서, 그것들의 대응되는 특징벡터

는 다음과 같이 Z를 PCA-변환된 공간으로 투영함으로서 얻어진다.Then, the eigenvalues and eigenvectors of the covariance matrix are calculated. Where r eigenvectors corresponding to r largest eigenvalues

Let's do it. For a set z of images, their corresponding feature vectors

Is obtained by projecting Z into the PCA-converted space as follows.

(식 4)

(Equation 4)

는 i번째

벡터이다.here

Is the i

Vector.

Fuzzy-LDA calculates the average value of each class of reduced feature vector and fuzzy membership using the feature vector and fuzzy membership.

(식 5)

(Equation 5)

는 j번째 레벨된 샘플 집합의 i번째 클래스의 소속도이다. 위의 결과를 이용하여 퍼지 클래스간 분산 행렬

와 퍼지 클래스내 분산 행렬

는 각각 다음과 같이 정의한다.here

Is the membership of the i th class of the j th level sample set. Dispersion matrix between fuzzy classes using the above result

And distribution matrices in fuzzy classes

Are defined as follows.

(식 6)

(Equation 6)

(식 7)

(Equation 7)

는 i번째 클래스

에서의 벡터의 수이고,

는 모든 특징벡터들의 평균값을 나타낸다. 최종적으로 최적의 퍼지 투영행렬

와 퍼지-LDA 방법에 의해 투영된 특징벡터는 각각 다음과 같이 계산된다.Where,

Is the i class

Is the number of vectors in,

Denotes an average value of all feature vectors. Finally the optimal fuzzy projection matrix

The feature vectors projected by and the fuzzy-LDA method are respectively calculated as follows.

(식 8)

(Expression 8)

(식 9)

(Eq. 9)

After the feature of gait image is extracted using fuzzy-LDA for the training image and the verification image, the matching degree is obtained by measuring the Euclidean distance between the training image and the verification image.

4. Gait recognition by fusion algorithm

3 shows a configuration diagram of gait recognition by a fusion algorithm. As shown in FIG. 3, a step of extracting a silhouette-based energy image from a gait image (S1), dividing the gait image by direction components by a contour (S2), and using a fuzzy-LDA technique Comprising the step of calculating the feature vector from the sub-image divided by the direction component (S3), the final step is recognized by the fusion method that can effectively fuse the matching degree calculated for each band. 5 shows the proposed gait recognition algorithm. As shown in FIG. 3, an image decomposition unit for transforming a gait image to a contour is transformed, a feature extraction unit for extracting a feature of each direction component, and a fusion unit for fusion matching of bands.

    As a final fusion method for each band, a weight sum rule was applied. Prior to fusion of the matching degree for each band, the matching degree normalization process is required, and the matching degree is normalized through Z-Score normalization.

    Next, after calculating the matching degree of each normalized class, recognition is performed by selecting the class having the best similarity by using the sum of the matching weights that are appropriately weighted for each band, and the weight sum rule. Decision based on is performed by calculating the band-specific output value calculated by the sum of the similarities of each class for all bands by Eq. 10, and then selecting the class with the best similarity among them by Eq. .

(식 10)

(Eq. 10)

(식 11)

(Eq. 11)

5. Application method and step recognition result

    The present invention implements a method of improving the recognition rate of the gait by simultaneously considering the matching degree by the direction obtained by the contour in addition to the matching degree calculated by the fuzzy-LDA. To this end, it is necessary to determine which subband to use among the various directional components obtained by the contour.

The image used for the performance evaluation of the algorithm is an NLPR DB which is widely used in the world, and the gait image DB is composed of four sets obtained from a total of 20 people. A gait energy image corresponding to one step was acquired from the above images, and a total of 728 images were acquired. Of these, three sets per person were used for learning and one set was used for verification.

으로 총

개의 대역으로 영상을 분해하였다. 각각의 DB에서 대역별로 거의 동일한 패턴의 인식률 변화를 확인할 수 있었으며, 이를 통해 걸음걸이 에너지 영상에서의 컨투어렛을 통한 방향별 부대역 영역에서의 공통적인 특징을 확인할 수 있었다. 본 발명에서는 이러한 각각의 부대역중 인식률이 우수한 2, 6, 7의 부대역 영역을 선택해 융합하여 실험하였다.4 shows the recognition rate through fuzzy-LDA for each decomposition image after decomposing the image into subbands for each direction through contour transform on the DB. The directional filter bank used in the experiment

By gun

The image was decomposed into two bands. In each DB, we could see the change of recognition rate of the same pattern for each band, and through this, we could confirm the common features in the subband region by direction through the contour in the gait energy image. In the present invention, the subband regions of 2, 6, and 7 having excellent recognition rates among these subbands were selected and fused.

Table 1 shows the gait recognition results. As shown in Table 1, the gait image captured at 0 ° showed recognition rate of 89.35% as a result of PCA, 97.14% as a result of fuzzy-LDA, and 9.96% as a result of development method. The gait image taken at 45 ° showed a recognition rate of 82.47% as a result of PCA, 92.98% as a result of fuzzy-LDA, and 94.82% as a result of the development method. According to the present invention, the proposed method is superior to the existing methods. In addition, the recognition rate of the subband region by direction is lower than the image obtained at 0 ° when obtained at 45 ° as a whole, but in terms of the improvement of the recognition rate as a result of fusion results in a higher recognition rate than I could confirm the giving.

{Table 1}

PCA Fuzzy
LDA Contour
Fuzzy-LDA Final recognition rate

89.35% 97.14% LL 97.14% 97.96% 2 49.25% 6 55.20% 7 58.05%

82.47% 92.98% LL 92.98% 94.82% 2 29.45% 6 41.81% 7 42.63%

1 shows a gait energy image.

2 shows a frequency division band of a contour filter bank.

3 is a flowchart illustrating a gait diagnosis process by a fusion algorithm.

4 shows the gait recognition rate of the subband region for each direction obtained by the contour transform.

Claims (1)

(1) extracting a silhouette-based energy image from the gait image; (2) dividing the gait image by direction components by a contour; (3) calculating a feature vector from sub-images divided by direction components using fuzzy-LDA technique; (4) the step of the final step recognition by the fusion method that can effectively fuse the matching degree calculated for each band; Gait Recognition Method using Contourlet Transform and Fuzzy Linear Discrimination

Images