CN107122795A - A kind of pedestrian integrated based on coring feature and stochastic subspace discrimination method again - Google Patents

Abstract

The present invention relates to a pedestrian re-identification method based on kernelization features and random subspace integration, comprising the following steps: S1, obtaining a training sample set and a test sample set of pedestrian images, and determining a kernelization function between the two samples; S2, Transform the original features of the two sample sets into kernelized features; S3, in the kernelized feature space of the training sample set, randomly select multiple different subspaces, and calculate the kernelized feature differences between different and the same pedestrian image pairs Value covariance matrix and its inverse matrix, to obtain the distribution function of the kernelized feature difference of the image pair; S4, in each subspace, calculate the probability that the sample pair is the same pedestrian and the probability of being a different pedestrian, and the two The ratio of the probabilities is used as the distance between the samples; S5, the distance is integrated to obtain the final distance between each sample pair. Compared with the prior art, the present invention has good pedestrian re-identification ability, is applicable to various features, and has strong robustness.

Description

A Pedestrian Re-ID Method Based on Kernelized Feature and Random Subspace Integration

technical field

The invention relates to a method for feature extraction and distance measurement learning in intelligent analysis of monitoring video, in particular to a pedestrian re-identification method based on kernelized features and random subspace integration.

Background technique

Pedestrian re-identification refers to the problem of matching pedestrian images from different camera perspectives in a multi-camera system. It provides critical assistance in the analysis of different aspects such as pedestrian identity and behavior, and has developed into a key component in the field of intelligent video surveillance.

The main methods in the field of person re-ID can be divided into the following two categories: 1) person re-ID methods based on feature representation; 2) person re-ID methods based on metric learning.

In the method of person re-identification based on feature representation, low-level visual features are the most commonly used features. Commonly used low-level visual features include color histogram, texture, etc. The specific description is as follows: the color histogram describes the color distribution characteristics of the entire image or a small area of it by counting the color distribution on the image. It is robust to viewing angle changes, but is susceptible to brightness changes such as lighting, so it is usually extracted on a specific color space. The structure information of the whole image or a small area described by the texture feature is a good supplement to the color information described by the color histogram feature. Most of the pedestrian re-identification algorithms are implemented based on the underlying visual features. However, when humans perform pedestrian re-identification tasks, they often judge whether two images are not completely based on the underlying visual features, but more on semantic attribute features. Belong to the same pedestrian. For example: hair style, type of shirt, type of coat, shoes and other information. Compared with the underlying visual features, the method based on semantic attribute features has natural advantages: semantic attributes are more robust to the differences in pedestrian appearance characteristics under different surveillance videos, and the description of the semantic attributes of the same pedestrian in different surveillance videos is usually is unchanged; semantic attributes are closer to human understanding, therefore, the results obtained by feature methods based on semantic attributes are more in line with human needs; methods based on semantic attributes are more convenient for human interaction.

After the feature representation method, how to measure the distance between different pedestrian images is also one of the key issues in the field of pedestrian re-identification. Feature-based methods usually use classic distance functions such as Euclidean distance, cosine distance, and geodesic distance when calculating the similarity of feature vectors. These classical distance functions do not take into account the characteristics of samples and thus tend to perform poorly. In recent years, a large number of literatures have adopted the method of distance measurement, and through the training of labeled samples, a distance function that is more in line with the characteristics of the sample is obtained, thereby improving performance. These methods do this by learning a distance function of the Markanobis form. Among them, the pedestrian re-identification method based on the simple and direct strategy (KISS) is at the forefront in terms of performance. However, this method is based on the theoretical assumption that the sample distribution obeys the Gaussian distribution, but in reality, the samples not only do not obey the Gaussian distribution perfectly, but may even deviate seriously, resulting in performance degradation. In addition, in actual situations, the sample size is often much smaller than the feature dimension, which makes the calculation of the Mahalanobis distance in metric learning difficult or even unsolvable.

Contents of the invention

The purpose of the present invention is to provide a pedestrian re-identification method based on the integration of kernelized features and random subspaces in order to overcome the above-mentioned defects in the prior art, so that the extracted features more closely obey the Gaussian distribution and can reconcile the sample size and feature dimension. The contradiction of the number avoids the SSS (small sample size) problem, thereby improving performance.

The purpose of the present invention can be achieved through the following technical solutions:

A pedestrian re-identification method based on kernelized features and random subspace integration, comprising the following steps:

S1, obtain the training sample set and test sample set of pedestrian images, determine the kernelization function between the two samples, the output value of the kernelization function is a one-dimensional real number, in each sample set, the same pedestrian has multiple images; in In the training sample set, the correspondence between each pedestrian and its multiple images is known, but in the test sample set, it is unknown;

S2, respectively convert the original features of the two sample sets into kernelized features, and the dimensions of the kernelized features are the number of samples in the training sample set;

S3. In the kernelized feature space of the training sample set, a plurality of different subspaces are randomly selected, and in each subspace, the covariance matrix and its inverse matrix of the kernelized feature difference of different pedestrian image pairs are calculated, and the calculation The covariance matrix and its inverse matrix of the kernelized feature difference of the same pedestrian image pair are used to obtain the distribution function of the kernelized feature difference of the image pair; the purpose of using the training sample is to learn a suitable distribution function through the training sample;

S4, respectively under each subspace (the subspace here is the subspace selected in step S3), calculate the difference of the kernelization feature of the sample pair in the test sample set, according to the difference covariance matrix and its inverse matrix and distribution function, Calculate the probability of the sample pair being the same pedestrian and the probability of being a different pedestrian, and use the ratio of the two probabilities as the distance between the two samples;

S5. Integrate the distances calculated in different subspaces to obtain the final distance between each sample pair in the test sample set, which is used for pedestrian identification. The smaller the final distance, the higher the possibility that the sample pair is the same pedestrian. In this way, the images belonging to the same pedestrian in the test sample set can be identified.

In the step S1, the kernelization function is a Gaussian kernel function, and the distribution function obtained in the step S3 is a Gaussian distribution function.

In the step S1, the kernelization function is k( _xi , x _j ), Wherein, σ=1, x _i , x _j represent the i-th and j-th training samples respectively.

In the step S2, the specific process of converting the original feature into a kernelized feature includes:

Convert the training set X to kernelized features

Among them, m is the number of samples in the training set, X∈R ^d×m , d is the dimension of sample features,

Convert the test set Z to kernelized features

Among them, n is the number of samples in the test set,

In the described step S3, the covariance Σ ₀ calculation formula of the kernelized feature difference of different pedestrian image pairs is:

Among them, y _ij = 0 represents all sample pairs that do not belong to the same pedestrian, and N ₀ is the total number of samples that meet this condition;

The covariance matrix Σ ₁ calculation formula of the kernelized feature difference of the same pedestrian image pair is:

Wherein, y _ij =1 represents all sample pairs that do not belong to the same pedestrian, and N ₁ is the total number of samples meeting this condition.

In the described step S4, the distance calculation formula between two samples is:

Compared with the prior art, the present invention has the following advantages:

(1) Transform the original features into kernelized features, and the difference distribution between feature pairs will more closely obey the Gaussian distribution, which is the basic theoretical assumption of distance metric learning based on simple and direct strategies;

(2) In the nonlinear space, the kernelized features tend to have stronger discriminative ability;

(3) Project the complex eigenvectors into multiple subspaces with smaller dimensions to calculate the distance respectively, and replace the traditional distance calculation method with the randomly selected subspace projection method, which can significantly improve the performance of the algorithm and optimize The process of calculating the sample distance is simplified, and the overhead of matrix operations is saved;

(4) The number of randomly selected subspace dimensions is much smaller than the sample size, reconciling the contradiction that the sample size is much smaller than the number of feature dimensions in practical applications, and thus makes the distance calculation more accurate and efficient.

Description of drawings

Fig. 1 is the flowchart of the inventive method;

Fig. 2 (a)-2 (d) is the probability distribution figure of the characteristic difference between the sample pair before and after using the present embodiment method, wherein: 2 (a) is the difference distribution of the same pedestrian sample pair under the original feature; 2 ( b) is the difference distribution of different pedestrian sample pairs under the original feature; 2(c) is the difference distribution of the same pedestrian sample pair under the kernelization feature; 2(d) is the difference distribution of different pedestrian sample pairs under the kernelization feature.

Fig. 3 (a), 3 (b) are the CMC curves on VIPeR (P=316) pedestrian re-identification public data set under the situation of using different parameters and different characteristics of the method of this embodiment, wherein: 3 (a) is The CMC curves using different parameters under the LOMO feature; 3(b) is the CMC curve using different parameters under the kCCA feature.

detailed description

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is carried out on the premise of the technical solution of the present invention, and detailed implementation and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

Example

A pedestrian re-identification method based on kernelized features and random subspace integration, comprising the following steps:

Step 1: Transform the original feature into a kernelized feature representation, as described below:

Obtain the training set X∈R ^d×m and the test set Z∈R ^d×n of pedestrian images, where the sample feature dimension is d, the number of samples in the training set is m, and the number of samples in the test set is _n . There are i training samples, and z _i represents the i-th test sample. use the kernel function Convert the training set X to kernelized features Convert the test set Z to kernelized features where σ=1. The specific conversion process can be expressed as

Step 2: In the kernelized feature space, randomly select L different subspaces. The specific description is as follows: After step 1 is completed, the dimension of the kernelized feature space is the same as the number of samples in the training set, that is, m dimension. Randomly select L different subspaces D _k (k=1,2,...,L) in this m dimension, and the dimension of the subspace is

Step 3: In different subspaces, calculate the covariance matrix Σ ₀ of the feature difference between different pedestrian image pairs, and find the inverse matrix Calculate the covariance matrix Σ ₁ of the feature difference between the same pedestrian image pair, and find the inverse matrix The specific description is as follows: in different subspaces, calculate the covariance matrix Σ ₀ of the feature difference between different pedestrian image pairs, and the specific calculation method is

Where y _ij =0 represents all sample pairs that do not belong to the same pedestrian, and N ₀ is the total number of samples meeting this condition. Calculate the covariance matrix covariance Σ ₁ of the feature difference between the same pedestrian image pair, the specific calculation method of Σ ₁ is

Where y _ij =1 represents all sample pairs belonging to the same pedestrian, and N ₁ is the total number of samples meeting this condition.

Step 4: In different subspaces, based on the Gaussian distribution function and two inverse matrices Calculate the probability that the difference between two features belongs to the same pedestrian and the probability that the difference between two features belongs to different pedestrians, and regard the ratio of the two probabilities as the distance between samples. The specific description is as follows: Denotes the test samples in the subspace with The difference of , use H ₀ to represent the hypothesis that the i-th test sample and the j-th test sample belong to different pedestrians, and use H ₁ to represent the hypothesis that the i-th test sample and the j-th test sample belong to the same pedestrian. When the difference between the sample pairs obeys the Gaussian distribution, the difference under the two assumptions can be respectively obtained as The probability

use express The logarithm of the ratio of the probabilities subject to the two hypotheses gives

According to Bayesian formula, formula (5) can be transformed into

which is

By removing the constant, the distance between the i-th test sample and the j-th test sample in the k-th subspace can be obtained, expressed as

Step 5: Integrate the distances calculated in L different subspaces to obtain the final distance. The specific description is as follows: For the L distances obtained in step 4, the weighted average method is used to integrate the distances, and the final distance formula can be Expressed as:

As shown in Figure 1, it is a flow chart of the present embodiment, and the specific implementation is as follows:

1) Determine the kernel function;

2) Convert the original features of the training samples into kernelized features;

3) Convert the original features of the test samples into kernelized features;

4) In order to judge whether z _i and z _j belong to the same pedestrian, use H ₀ to indicate that they are not similar, that is, they do not belong to the same pedestrian, and use H ₁ to indicate that they are similar, that is, they belong to the same pedestrian;

5) Randomly select L subspaces D _k (k=1,2,...,L) in the kernelized feature space;

6) In different subspaces, calculate the covariance matrix Σ ₀ of the feature difference between different pedestrian image pairs, and find the inverse matrix

7) In different subspaces, calculate the covariance matrix Σ ₁ of the feature difference between the same pair of pedestrian images, and find the inverse matrix

8) Calculate the feature difference in different subspaces A probability distribution function that obeys two assumptions, and the logarithmic function of the probability ratio as the distance between samples;

9) In different subspaces, transform the probability ratio distance into Mahalanobis distance d _k (z _i , z _j );

10) Integrate the calculated distances in the L subspaces to obtain the final sample distance.

Fig. 2 (a)-2 (d) is the probability distribution figure of the characteristic difference between sample pairs before and after using the method of this embodiment, the histogram is the actual probability distribution, and the linear graph is the Gaussian distribution curve drawn according to the variance of the data, The four original features compared are LOMO, kCCA, SCNCD, and ELF18, which are widely used in pedestrian re-identification methods, where: 2(a) is the difference distribution of the same pedestrian sample pair under the original features; 2( b) is the difference distribution of different pedestrian sample pairs under the original feature; 2(c) is the difference distribution of the same pedestrian sample pair under the kernelization feature; 2(d) is the difference distribution of different pedestrian sample pairs under the kernelization feature.

Figure 3 (a), 3 (b) is the rank-1 matching rate on VIPeR (P=316) pedestrian re-identification public data set under the situation of using different parameters and different features of the method of this embodiment, and it is compared with the traditional Regularization methods were compared. Among them: 3(a) is the rank-1 matching rate using different parameters under the LOMO feature; 3(b) is the rank-1 matching rate using different parameters under the kCCA feature.

Table 1

Table 1 shows the performance comparison between the method of this embodiment and other metric-based learning algorithms on the public dataset of VIPeR (P=316) pedestrian re-identification.

Table 2

Table 2 shows the performance comparison between the method of this embodiment and other metric-based learning algorithms on the public dataset of PRID 450S (P=225) pedestrian re-identification.

table 3

KRKISS NFST MLAPG wxya MFA wxya KISS LFDA time 5.04 2.48 40.9 3.86 2.58 2.74 7.41 229.3

Table 3 shows the comparison of training time overhead between the method of this embodiment and other metric-based learning algorithms.

Claims (6)

1. a kind of pedestrian integrated based on coring feature and stochastic subspace discrimination method again, it is characterised in that including following step Suddenly：

S1, obtains the training sample set and test sample collection of pedestrian image, the coring function between two samples is determined, in each sample This concentration, same pedestrian has multiple images；

S2, is converted into coring feature, the dimension of coring feature is training sample set by the primitive character of two sample sets respectively In number of samples；

S3, in the coring feature space of training sample set, randomly selects multiple different subspaces, respectively in each subspace Under, the covariance matrix and its inverse matrix of the coring feature difference of different pedestrian images pair are calculated, identical pedestrian image pair is calculated Coring feature difference covariance matrix and its inverse matrix, obtain the distribution function of the coring feature difference of image pair；

S4, respectively under each subspace, calculates the difference that test sample concentrates the coring feature of sample pair, according to difference covariance Matrix and its inverse matrix and distribution function, calculate sample to the probability for identical pedestrian and be different pedestrians probability, and by two The ratio of individual probability is used as the distance between two samples；

S5, carries out integrated to the distance that is calculated in variant subspace, obtains test sample and concentrates final between this pair of various kinds Distance, for pedestrian's identification, finally apart from smaller, sample is higher to the possibility for identical pedestrian.

2. a kind of pedestrian integrated based on coring feature and stochastic subspace according to claim 1 discrimination method again, its It is characterised by, in described step S1, coring function is gaussian kernel function, and the distribution function that step S3 is obtained is Gaussian Profile letter Number.

3. a kind of pedestrian integrated based on coring feature and stochastic subspace according to claim 1 or 2 discrimination method again, Characterized in that, in described step S1, coring function is k (x_i,x_j),Wherein, σ= 1, x_i、x_jI-th, j training sample is represented respectively.

4. a kind of pedestrian integrated based on coring feature and stochastic subspace according to claim 3 discrimination method again, its It is characterised by, in described step S2, the detailed process that primitive character is converted into coring feature includes：

Training set X is converted into coring feature

<mrow> <mover> <mi>X</mi> <mo>~</mo> </mover> <mo>=</mo> <mo>&amp;lsqb;</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mn>1</mn> </msub> <mo>,</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mn>2</mn> </msub> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>m</mi> </msub> <mo>&amp;rsqb;</mo> <mo>=</mo> <mo>&amp;lsqb;</mo> <mi>K</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mn>1</mn> </msub> <mo>,</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>,</mo> <mi>K</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mn>2</mn> </msub> <mo>,</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>K</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>m</mi> </msub> <mo>,</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> <mo>&amp;Element;</mo> <msup> <mi>R</mi> <mrow> <mi>m</mi> <mo>&amp;times;</mo> <mi>m</mi> </mrow> </msup> <mo>,</mo> </mrow>

Wherein, m is training set number of samples, X ∈ R^d×m, d is sample characteristics dimension,

<mrow> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>i</mi> </msub> <mo>=</mo> <mi>K</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>=</mo> <msup> <mrow> <mo>&amp;lsqb;</mo> <mi>k</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mn>1</mn> </msub> <mo>)</mo> </mrow> <mo>,</mo> <mi>k</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mn>2</mn> </msub> <mo>)</mo> </mrow> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>k</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> </mrow> <mi>T</mi> </msup> <mo>&amp;Element;</mo> <msup> <mi>R</mi> <mrow> <mi>m</mi> <mo>&amp;times;</mo> <mn>1</mn> </mrow> </msup> <mo>,</mo> </mrow>

Test set Z is converted into coring feature

<mrow> <mover> <mi>Z</mi> <mo>~</mo> </mover> <mo>=</mo> <mo>&amp;lsqb;</mo> <msub> <mover> <mi>z</mi> <mo>~</mo> </mover> <mn>1</mn> </msub> <mo>,</mo> <msub> <mover> <mi>z</mi> <mo>~</mo> </mover> <mn>2</mn> </msub> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msub> <mover> <mi>z</mi> <mo>~</mo> </mover> <mi>n</mi> </msub> <mo>&amp;rsqb;</mo> <mo>=</mo> <mo>&amp;lsqb;</mo> <mi>K</mi> <mrow> <mo>(</mo> <msub> <mi>z</mi> <mn>1</mn> </msub> <mo>,</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>,</mo> <mi>K</mi> <mrow> <mo>(</mo> <msub> <mi>z</mi> <mn>2</mn> </msub> <mo>,</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>K</mi> <mrow> <mo>(</mo> <msub> <mi>z</mi> <mi>n</mi> </msub> <mo>,</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> <mo>&amp;Element;</mo> <msup> <mi>R</mi> <mrow> <mi>m</mi> <mo>&amp;times;</mo> <mi>n</mi> </mrow> </msup> <mo>,</mo> </mrow>

Wherein, n is test set number of samples,

5. a kind of pedestrian integrated based on coring feature and stochastic subspace according to claim 4 discrimination method again, its It is characterised by, in described step S3, the covariance matrix Σ of the coring feature difference of different pedestrian images pair₀Calculating formula is：

<mrow> <msub> <mi>&amp;Sigma;</mi> <mn>0</mn> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <msub> <mi>N</mi> <mn>0</mn> </msub> </mfrac> <munder> <mo>&amp;Sigma;</mo> <mrow> <msub> <mi>y</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mn>0</mn> </mrow> </munder> <mrow> <mo>(</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>i</mi> </msub> <mo>-</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>j</mi> </msub> <mo>)</mo> </mrow> <msup> <mrow> <mo>(</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>i</mi> </msub> <mo>-</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> </mrow>

Wherein, y_ij=0 represents all samples pair for being not belonging to same a group traveling together, N₀To meet the total sample number of the condition；

The covariance matrix Σ of the coring feature difference of identical pedestrian image pair₁Calculating formula is：

<mrow> <msub> <mi>&amp;Sigma;</mi> <mn>1</mn> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <msub> <mi>N</mi> <mn>1</mn> </msub> </mfrac> <munder> <mo>&amp;Sigma;</mo> <mrow> <msub> <mi>y</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mn>1</mn> </mrow> </munder> <mrow> <mo>(</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>i</mi> </msub> <mo>-</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>j</mi> </msub> <mo>)</mo> </mrow> <msup> <mrow> <mo>(</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>i</mi> </msub> <mo>-</mo> <msub> <mover> <mi>x</mi> <mo>~</mo> </mover> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> </mrow>

Wherein, y_ij=1 represents all samples pair for being not belonging to same a group traveling together, N₁To meet the total sample number of the condition.

6. a kind of pedestrian integrated based on coring feature and stochastic subspace according to claim 5 discrimination method again, its It is characterised by, in described step S4, between two samples is apart from calculating formula：

<mrow> <mi>d</mi> <mrow> <mo>(</mo> <msub> <mi>z</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>z</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <msup> <mrow> <mo>(</mo> <msub> <mover> <mi>z</mi> <mo>~</mo> </mover> <mi>i</mi> </msub> <mo>-</mo> <msub> <mover> <mi>z</mi> <mo>~</mo> </mover> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> <mrow> <mo>(</mo> <msubsup> <mi>&amp;Sigma;</mi> <mn>1</mn> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msubsup> <mo>-</mo> <msubsup> <mi>&amp;Sigma;</mi> <mn>0</mn> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msubsup> <mo>)</mo> </mrow> <mrow> <mo>(</mo> <msub> <mover> <mi>z</mi> <mo>~</mo> </mover> <mi>i</mi> </msub> <mo>-</mo> <msub> <mover> <mi>z</mi> <mo>~</mo> </mover> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mo>.</mo> </mrow> 2