CN109801326A - It is a kind of for obtaining the image measuring method of human somatotype data - Google Patents

Abstract

The present invention relates to the data processing methods of a kind of design, in particular to a kind of for obtaining the image measuring method of human somatotype data.The present invention is to record the coordinate information of each characteristic point by choosing the inflection point on the positive side of human body as characteristic point；Dimension-reduction treatment is carried out to the data in characteristic point sample set again, obtains the average template of body shape；It is labeled in the neckline for the clothes worn for user, cuff, bust, waistline, hip circumference and ankles bit；The distance between two cuffs in S3, double-legged ankle, the position of both legs and the position of waistline are obtained according to mark Face Detection, human body front shape is obtained after being computed；According to ankle position in the S3 detected, similar Block- matching is carried out, obtains human body side view.The present invention can be feasible, effective, and can get the three-dimensional figure data of human body well, is widely applied for fields such as dress designing, virtual fittings.

Description

It is a kind of for obtaining the image measuring method of human somatotype data

Technical field

The present invention relates to the data processing methods of a kind of design, in particular to a kind of for obtaining the figure of human somatotype data As measurement method.

Background technique

The three-dimensional data acquisition of human body key position either still all has in research field in actual life important Research significance, if can have a kind of conveniently mode that can more rapidly and accurately get human body key position automatically Three-dimensional data, such as bust, waist and hip circumference, then this method is in e-commerce, the neck such as the fitting of 3D human body and dress designing Domain will generate very big impetus.In addition, fast and easily human body three-dimensional DATA REASONING is for the big of human body type's research field The acquisition statistics of sample data can also generate great impetus.

In the world without 2 identical human bodies, carrying out anthropological measuring research earliest is that Quelet was measured in 1870 Each position average-size of male.Thereafter lot of domestic and international experts and scholars much probe into, in order to find a kind of side Just the quick method that can accurately obtain human body type's critical data again simultaneously, the main three classes of these methods, contact measurement method, Contactless measurement and spatial digitizer scanning.Contact type measurement method is divided into Martin's method (Manin), sliding gauge method (sliding gauge), replica (replica)；Non-contact type measuring method is divided into photographic process, Mohr diagram method (Moire Topography), spatial digitizer scanning is the development based on various kinds of sensors in recent years and the new measurement side of one kind for rising Method.

Contact measurement method uses conventional tool tape measure, measures to characteristics of human body position.It is developed so far, compares into Ripe measuring tool is Martin measure instrument, the body surface by measuring human body punctuate and then human body sets datum mark and baseline It is long, projector distance, the critical quantities such as angle.But since human somatotype is different, this method is complex, measurement process it is cumbersome because And it is only limitted to clothing industry.Meanwhile either Martin's method, or sliding gauge method, replica are in current data quick obtaining Epoch, most situations do not adapt to.Therefore sight has just been turned to non-connect by the scholar of foreign countries in last century the eighties Touch measurement.

What non-contact measurement was carried out at first is camera method, is broadly divided into two dimension camera shooting and three-dimensional camera shooting.Two dimension camera shooting Method requirement survey crew estimates the figure data of the measured according to the measured in picture relative to surrounding enviroment.It is this Whether experience and object of reference setting of the method based on survey crew are reasonable, thus are extremely difficult to preferable precision effect, can only A kind of measurement means as auxiliary.And three-dimensional camera shooting method is to be taken the photograph based on the three-dimensional presentation principle in computer vision using CCD Camera obtains a series of two dimensional images of a 3 D human body, it is indicated in space coordinates, finally proposes completely to retouch State the indicatrix of human somatotype feature.This method requires CCD camera constantly to clap from different perspectives measured personnel It takes the photograph, complex steps and since human dressing is affected to measurement result, the final result precision is not high.

English physicist Rayleigh combination moire pattern has invented a kind of measuring technique, referred to as Mohr diagram method.He according to The striped form of object evaluates and tests the spacing uniformity between each strain line of grating scale.In the method, two blocks of gratings are usually utilized The overlapping of two pictures of (referred to as grating is paid) or grating generates Moire fringe, to obtain various measured information.Its essence is Light passes through the diffraction in grating time and the result of interference.Environment is often sufficiently complex in real life is extremely difficult in laboratory Requirement, thus this method is difficult to popularize.

At this stage, occur all kinds of with the promotion of the invention of various kinds of sensors and precision in large enterprise and laboratory Human body three-dimensional data acquisition instrument, such as 3D CaMega DCS series Whole Body (half body) scanning system, Visbody tri- Tie up body-scanner etc..Measured human body is placed in multiple scanners under, and the sensor on each scanner passes through rangefinder The spatial position of human body a part is obtained, then summarizes the spatial model for sketching the contours of human somatotype, this method precision is high, and speed is fast The figure data of people can really be obtained.But because instrument is mounted with a large amount of sensor and probe, and need with into The host that row calculates is connected, thus cost is huge, can not use in business and dress designing field on a large scale.Therefore we are uncommon A kind of human body measurement method simple and fast in the case where guaranteeing precision is found in prestige.

In recent years, Cootes et al. proposes a kind of active shape model (Active Shape Models towards image Method, abbreviation ASM) method, it has obtained answering extensively in fields such as digital picture understanding, computer vision and Medical Image Processings With.Active shape model (ASM) is a kind of Object shape description technology, can be used for solving the shape system of the target homing in image Model is counted, model can be established for any object shape, there is preferable robustness, especially extracted in features of human face images On, the coordinate position of multiple characteristic points can be accurately obtained on face according to the sample set that a large amount of training obtain. This method achieves successful application on the fields such as recognition of face and signature analysis.Thereafter Cootes et al. is in order to calculate ASM The positioning of characteristic point is more clear in method, proposes active appearance models (Active Appearance Models, abbreviation AAM) Method, core concept are to find the relationship between model parameter and shapes textures variation, in the search phase, root in the training stage Model is constantly adjusted according to the relationship until it converges on target shape.But this method for the texture of target shape and illumination very Environment is extremely complex when shooting for sensitivity, especially human body, and the texture of human dressing is also different, carries out human body using AAM It is accurate unlike the effect of ASM method when feature extraction and measurement.

In consideration of it, we have proposed a kind of human body image figure data measuring method based on active shape model, and with Correspondence devise the measurement clothes of a set of characteristics of human body position coloring.By training, standardized human body is being calculated just in PCA dimensionality reduction Side average template will put on the people of measurement clothes further according to mahalanobis distance and color-match in the template of acquisition and photo to be measured Body is matched.Human body three-dimensional coordinate points have been obtained, the curve obtained after these point fittings approximate can have been obtained crucial as human body The figure curve at position.In addition, can be greatly lowered when measured personnel wear measurement and take according to the mark point on measurement clothes Due to external environment bring measurement error.This method compares existing image measuring method, improves measurement accuracy；With three-dimensional The methods of scanner scanning is compared, and is measured more convenient.

Summary of the invention

The present invention aiming at the shortcomings in the prior art, proposes that one kind is feasible, effective, and can get the three of human body well The method for tieing up figure data.

The application be based on active shape model (ASM) human body image measure, ASM algorithm realization be broadly divided into training and Search for two main parts.Wherein trained part mainly passes through a large amount of artificial punctuate, obtains the positive side of N group human body Sample information, each group information contained the coordinate information of positive 27 characteristic points of 65 characteristic points and side of human body.And It is handled by PCA, obtains the average template of human body.

The present invention is achieved by following technical proposals:

It is a kind of for obtaining the image measuring method of human somatotype data, characterized by the following steps:

S1: the inflection point on the positive side of human body is chosen as characteristic point, wherein 65 characteristic points are chosen in front, in side 27 characteristic points are chosen, the coordinate information of each characteristic point is recorded；Because human body front is complex, 65 features are chosen altogether Point, and side is relatively simple, chooses 27 characteristic points.Specific reconnaissance such as Fig. 1 and Fig. 2；

S2: by carrying out dimension-reduction treatment to the data in characteristic point sample set obtained in step S1, body shape is obtained Average template；

S3: referring to the position of characteristic point in S1, the neckline for the fitted garment worn for user, cuff, bust, waistline, Hip circumference and ankles bit are labeled；

S4: the position of the distance between two cuffs in S3, double-legged ankle and waistline, warp are obtained according to mark Face Detection Human body front shape is obtained after calculating；

S5: according to ankle position in the S3 detected, similar Block- matching is carried out, obtains human body side view.

Preferably, feature selected in the step S1 of the above-mentioned image measuring method for obtaining human somatotype data Point, in the outside boundary line of human body contour outline；The data record of each characteristic point are as follows:

X_i=(x_i0, y_i0, x_i1, y_i1..., x_ik, y_ik..., x_i(n-1), y_i(n-1))^T, wherein (x_ik, y_ik) represent kth in figure A coordinate, human body front sample n_Just=65, side n_Side=27, i.e. sample set X=[X of the acquisition containing N group training sample₁, X₂... X_N], N=92=n_Just+n_Side。

As more preferably selecting, registration process is carried out to data obtained in S2；I.e. by first sample X₁=(x₁₀, y₁₀, x₁₁, y₁₁..., x_1k, y_1k... x_1(n-1), y_1(n-1))^TAs master sample, to each sample X thereafter_i=(x_i0, y_i0, x_i1, y_i1..., x_ik, y_ik..., x_i(n-1), y_i(n-1))^TCarry out scaling and position translation；Since human body is deposited in every photo Antipode in position and size, if the human body feature point marked is directly carried out statistical modeling.It would become hard to obtain correct Rule.Therefore, it before being modeled, needs to carry out registration process to obtained point, i.e., by first sample

X₁=(x₁₀, y₁₀, x₁₁, y₁₁..., x_1k, y_1k..., x_1(n-1), y_1(n-1))^TIt is every to thereafter as master sample A sample X_i=(x_i0, y_i0, x_i1, y_i1..., x_ik, y_ik..., x_i(n-1), y_i(n-1))^TCarry out scaling and position translation.

Detailed process is as follows:

S21: the translational movement of each sample is calculated:

The center of gravity of first sample is calculated first are as follows:

The then translational movement of i-th of sample are as follows:

S22: the scaling of each sample is calculated:

Horizontal and vertical average conduct scaling S is taken, is calculated as follows:

Wherein: Max (X_i(x), Min (X_i(x), Max (Y_i(x), Min (Y_i(x) it indicates in i-th of sample on the direction x most It is worth greatly, minimum value, the maximum value and minimum value on the direction y；According to T_iAnd S_iCalculate i-th of sample:

S23: being iterated, and determines final sample set, after carrying out primary above-mentioned transformation, has obtained a new sample Each sample is aligned by this collection, this sample set according to first sample on the basis of original sample set.At this point, I Carry out same operation step again, in order to so that in the case where guaranteeing sample rigid body shape invariance so that sample it Between difference it is sufficiently small, so as to obtain accurately and effectively body shape rule:

Same operation step is carried out again by step S21, S23, finally obtains sample set:

S24: dimension-reduction treatment therein is that sample obtained in step S23 is handled as follows:

S241: the average shape of sample set is acquired:

S242: the deviation of each sample and average shape in sample set is sought:

S243: constituting the covariance matrix of sample set, enables:

D=(dx₀|dx₁|...|dx_n-1)

The then covariance matrix of sample set:

S244: the characteristic value and feature vector of covariance matrix are calculated, and characteristic value is sorted

∑p_k=λ_kp_k,

Wherein, λ_kIt is the big characteristic value of K of covariance matrix；

Principal Component Analysis is a kind of technology of simplified data set, it is a linear transformation, and in characteristic processing, mode is known Not, the application in the fields such as data analysis, this method can by one group of index with correlativity, be reassembled into one group it is mutual Unrelated index replaces original index, so functionally, principal component analysis is a kind of dimension reduction method mathematically.Root According to S241 and S242, we know there is sample containing N in the sample set finally obtained, and each human body front sample contains 2 × 65 Dimensional feature, human body side sample contain 2 × 27 dimensional features, and so big data volume certainly will will cause ASM search phase calculation amount Excessive, in order to improve efficiency of algorithm, we are carrying out PCA dimension-reduction treatment to obtained new sample set,

S245: choosing corresponding feature vector, t characteristic value before choosing first in k characteristic value, so that:

There is a feature vector corresponding the characteristic value of each acquirement, is formed new matrix P

P=(p₀|p₁|...|p_t-1)

Then any one body shape is represented by

Wherein b is the vector of t × 1, and different b, which brings above formula into, can be obtained different body shapes.

Preferably, for the feature of the position of both legs in the above-mentioned image measuring method for obtaining human somatotype data Point matching with the following method: first find both feet ankle position (x_left, y_left) and (x_right, y_right), then according to double Labeling position (the x of foot ankle and waistline_crotch, y_crotch) calculate two legs the position θ that strides, then it is right after Pan and Zoom The characteristic point of two legs is subject to lower operation again:

x_new=(x_i-x_crotch)cosθ-(y_i-y_crotch)sinθ+x_crotch

y_new=(x_i-x_crotch)sinθ+(y_i-y_crotch)cosθ+y_crotch。

Preferably, increasing by one between the step S4 and S5 of the above-mentioned image measuring method for obtaining human somatotype data A step S42 establishes characteristic point local gray level model, and calculates the gray scale and the gray level model of this feature point in training of each point Distance；

The local gray level model is realized by following step:

S421: sample each method of characteristic point to two sides respectively take k point, calculate the gray scale of the point, then share 2k+1 A gray value, including characteristic point itself；Then the gray scale layer of j-th of characteristic point of i-th of sample can be denoted as:

g_ij=(g_ij1, g_ij2..., g_ij(2k+1))^T；

Wherein, i=1,2 ... N, for human body front j=0,1,2 ..., 64, human body side j=0,1,2 ..., 26；

S422: to g_ijDifference is asked to obtain:

dg_ij=(g_ij2-g_ij1, g_ij4-g_ij3..., g_ij(2k+1)-g_ij(2k))^T；

After standardization:

S423: the mean value of j-th of characteristic point gray scale layer is calculatedWith variance cov_j:

As more preferably selecting, the ash of above-mentioned each point Spend is to calculate to determine by following manner at a distance from the gray level model of this feature point in training:

The smallest point of each point mahalanobis distance in normal direction is chosen, as the new position of this feature point, after all the points update, Attitude parameter b is updated, new model is generated, operating process as above is repeated, is less than until with the matched error rate of human body in figure 5%.

In order to do not done at one punctuate work picture on mark human body feature point, according to opencv carry classifier, It can accurately detect the position of human eye at this stage.Therefore, we first shine into row ASM algorithm operating to human body front. The position that people's eyes are first found out according to the classifier that opencv is carried, by our people in the average template that the training stage obtains Eye makes the difference to obtain translation distance l with it, then in standardized human body front is shone, is opened according to the available people's both hands of Face Detection Armband length, the mould using the armband lenth ratio in arm length degree and template as pantograph ratio s, then after being translated through translation scaling Plate is

WhereinFor the average template obtained in training process；It, will be in obtained average template and figure by above step The human body side of non-punctuate has carried out thick matching.It needs to be adjusted details at this time.

Preferably, similar Block- matching is in the step S5 of the above-mentioned image measuring method for obtaining human somatotype data The people foot range determined by during the positive template matching of human body is first passed through, the point within the scope of foot is subjected to statistical Analysis, obtains the provincial characteristics within the scope of this, including gray scale maximum value G_max, minimum value G_min, average value G_avg, tri- color institute accounting of RGB Example R_r, R_g, R_b, formed vector:

R=(G_max, G_min, G_avg, R_r, R_g, R_b)

It is standardized:

R=(G_max/ 255, G_min/ 255, G_avg/ 255, R_r, R_g, R_b)；

The nose position feature of side, foot position feature are determined again, using the ratio in its length and template as contracting Ratio is put, then the foot of target side dough figurine body and template foot are made the difference as translation distance, then is carried out by scaling is translated with front Same operation obtains side matching result.

In this group of human body photo, people's foot of human body side photo can obtain foot position progress by human body front Similar Block- matching obtains, and after carrying out the matched experimental result of similar block and obtaining side foot position, we can be according to the colour of skin The position where people's side face is detected, at this point, can determine nose position due to the feature of people's side face nose.At this point, people's nose Distance to foot can determine, using in its length and template nose and foot length ratio as pantograph ratio s, then by target side The foot and template foot of dough figurine body make the difference as translation distance l, then can translate scaling with front and carry out same operation.

The utility model has the advantages that the present invention can be feasible, effective, and can get the three-dimensional figure data of human body well, for clothes The fields such as installing meter, virtual fitting are widely applied.

Detailed description of the invention

Fig. 1 human body front punctuate schematic diagram

Fig. 2 human body side punctuate schematic diagram

Fig. 3 not registered human body positive feature point set

Fig. 4 not registered human body lateral feature point set

Human body positive feature point set after Fig. 5 registration

Human body lateral feature point set after Fig. 6 registration

Fig. 7 human body front average template slightly matches

Fig. 8 human body front similar block matching result

Fig. 9 human body side similar block matching result

The result of Figure 10 ASM matching characteristic point

The foot leg of the non-wear measurement clothes of Figure 11 matches figure

The foot leg of Figure 12 wear measurement clothes matches figure

The front matching result of Figure 13 label garment feature point

The side matching result of Figure 14 label garment feature point

The positive side and planar survey of Figure 15 human body take experiment effect

Specific embodiment

Implementation of the invention is illustrated below:

Embodiment 1

It is a kind of for obtaining the image measuring method of human somatotype data, include the following steps:

S1: the inflection point on the positive side of human body is chosen as characteristic point, wherein 65 characteristic points are chosen in front, in side 27 characteristic points are chosen, the coordinate information of each characteristic point is recorded；Because human body front is complex, 65 features are chosen altogether Point, and side is relatively simple, chooses 27 characteristic points.Specific reconnaissance following Fig. 1 and Fig. 2；

In the outside boundary line of human body contour outline；The data record of each characteristic point are as follows:

X_i=(x_i0, y_i0, x_i1, y_i1..., x_ik, y_ik..., x_i(n-1), y_i(n-1))^T, wherein (x_ik, y_ik) represent kth in figure A coordinate, human body front sample n_Just=65, side n_Side=27, i.e. sample set X=[X of the acquisition containing N group training sample₁, X₂... X_N], N=92=n_Just+n_Side。

S2: by carrying out dimension-reduction treatment to the data in characteristic point sample set obtained in step S1, body shape is obtained Average template；

S3: referring to the position of characteristic point in S1, in the neckline for the clothes worn for user, cuff, bust, waistline, hip circumference It is labeled with ankles bit；

S4: the distance between two cuffs in S3, double-legged ankle, the position of both legs and waist are obtained according to mark Face Detection The position enclosed obtains human body front shape after being computed；

S5: according to ankle position in the S3 detected, similar Block- matching is carried out, obtains human body side view.

Embodiment 2

Method similar with embodiment, it is a kind of for obtaining the image measuring method of human somatotype data, including such as

Lower step:

S1: the inflection point on the positive side of human body is chosen as characteristic point, wherein 65 characteristic points are chosen in front, in side 27 characteristic points are chosen, the coordinate information of each characteristic point is recorded；Because human body front is complex, 65 features are chosen altogether Point, and side is relatively simple, chooses 27 characteristic points.Specific reconnaissance following Fig. 1 and Fig. 2；

In the outside boundary line of human body contour outline；The data record of each characteristic point are as follows:

X_i=(x_i0, y_i0, x_i1, y_i2..., x_ik, y_ik..., x_i(n-1), y_i(n-1))^T, wherein (x_ik, y_ik) represent kth in figure A coordinate, human body front sample n_Just=65, side n_Side=27, i.e. sample set X=[X of the acquisition containing N group training sample₁, X₂... X_N], N=92=n_Just+n_Side。

S2: by carrying out dimension-reduction treatment to the data in characteristic point sample set obtained in step S1, body shape is obtained Average template；

Registration process is carried out to data obtained in S2；I.e. by first sample

X₁=(x₁₀, y₁₀, x₁₁, y₁₁..., x_1k, y_1k..., x_1(n-1), y_1(n-1))^TIt is every to thereafter as master sample A sample X_i=(x_i0, y_i0, x_i1, y_i1..., x_ik, y_ik..., x_i(n-1), y_i(n-1))^TCarry out scaling and position translation；By There are the antipodes of position and size for human body in every photo, build if the human body feature point marked is directly carried out statistics Mould.It would become hard to obtain correct rule.Therefore, it before being modeled, needs to carry out registration process to obtained point, i.e., by first A sample

X₁=(x₁₀, y₁₀, x₁₁, y₁₁..., x_1k, y_1k..., x_1(n-1), y_1(n-1))^TIt is every to thereafter as master sample A sample X_i=(x_i0, y_i0, x_i1, y_i1..., x_ik, y_ik..., x_i(n-1), y_i(n-1))^TCarry out scaling and position translation.

Detailed process is as follows:

S21: the translational movement of each sample is calculated:

The center of gravity of first sample is calculated first are as follows:

The then translational movement of i-th of sample are as follows:

S22: the scaling of each sample is calculated:

Horizontal and vertical average conduct scaling S is taken, is calculated as follows:

Wherein: Max (X_i(x), Min (X_i(x), Max (Y_i(x), Min (Y_i(x) it indicates in i-th of sample on the direction x most It is worth greatly, minimum value, the maximum value and minimum value on the direction y；According to T_iAnd S_iCalculate i-th of sample:

S23: being iterated, and determines final sample set, after carrying out primary above-mentioned transformation, has obtained a new sample Each sample is aligned by this collection, this sample set according to first sample on the basis of original sample set.At this point, I Carry out same operation step again, in order to so that in the case where guaranteeing sample rigid body shape invariance so that sample it Between difference it is sufficiently small, so as to obtain accurately and effectively body shape rule, wherein Fig. 3, Fig. 4 be not registered human body A front surface and a side surface feature point set, Fig. 5, Fig. 6 are human body a front surface and a side surface feature point set after registration:

Same operation step is carried out again by step S21, S23, finally obtains sample set:

S24: dimension-reduction treatment therein is that sample obtained in step S23 is handled as follows:

S241: the average shape of sample set is acquired:

S242: the deviation of each sample and average shape in sample set is sought:

S243: constituting the covariance matrix of sample set, enables:

D=(dx₀|dx₁|...|dx_n-1)

The then covariance matrix of sample set:

S244: the characteristic value and feature vector of covariance matrix are calculated, and characteristic value is sorted

∑p_k=λ_kp_k,

Wherein λ_kIt is the big characteristic value of K of covariance matrix；

Principal Component Analysis is a kind of technology of simplified data set, it is a linear transformation, and in characteristic processing, mode is known Not, the application in the fields such as data analysis, this method can by one group of index with correlativity, be reassembled into one group it is mutual Unrelated index replaces original index, so functionally, principal component analysis is a kind of dimension reduction method mathematically.Root According to S241 and S242, we know there is sample containing N in the sample set finally obtained, and each human body front sample contains 2 × 65 Dimensional feature, human body side sample contain 2 × 27 dimensional features, and so big data volume certainly will will cause ASM search phase calculation amount Excessive, in order to improve efficiency of algorithm, we are carrying out PCA dimension-reduction treatment to obtained new sample set,

S245: choosing corresponding feature vector, t characteristic value before choosing first in k characteristic value, so that:

There is a feature vector corresponding the characteristic value of each acquirement, is formed new matrix P

P=(p₀|p₁|...|p_t-1)

Then any one body shape is represented by

Wherein b is the vector of t × 1, and different b, which brings above formula into, can be obtained different body shapes.

S3: referring to the position of characteristic point in S1, in the neckline for the clothes worn for user, cuff, bust, waistline, hip circumference It is labeled with ankles bit；

S4: the distance between two cuffs in S3, double-legged ankle, the position of both legs and waist are obtained according to mark Face Detection The position enclosed, obtains human body front shape after being computed, Fig. 7 is that human body front average template slightly matches；

It is above-mentioned for obtaining the matching of the characteristic point in the image measuring methods of human somatotype data for the position of both legs With the following method: first finding the ankle position (x of both feet_left, y_left) and (x_right, y_rigiht), then according to double-legged ankle and Labeling position (the x of waistline_crotch, y_crotch) calculate two legs the position θ that strides, then after Pan and Zoom, to the spy of two legs Sign point is subject to lower operation again:

x_new=(x_i-x_crotch)cosθ-(y_i-y_crotch)sinθ+x_crotch

y_new=(x_i-x_crotch)sinθ+(y_i-y_crotch)cosθ+y_crotch。

Preferably, increasing by one between the step S4 and S5 of the above-mentioned image measuring method for obtaining human somatotype data A step S42 establishes characteristic point local gray level model, and calculates the gray scale and the gray level model of this feature point in training of each point Distance；

The local gray level model is realized by following step:

S421: sample each method of characteristic point to two sides respectively take k point, calculate the gray scale of the point, then share 2k+1 A gray value, including characteristic point itself；Then the gray scale layer of j-th of characteristic point of i-th of sample can be denoted as:

g_ij=(g_ij1, g_ij2..., g_ij(2k+1))^T；

Wherein, i=1,2 ... N, for human body front j=0,1,2 ..., 64, human body side j=0,1,2 ..., 26；

S422: to g_ijDifference is asked to obtain:

dg_ij=(g_ij2-g_ij1, g_ij4-g_ij3..., g_ij(2k+1)-g_ij(2k))^T；

After standardization:

S423: the mean value of j-th of characteristic point gray scale layer is calculatedWith variance cov_j:

As more preferably selecting, the gray scale of above-mentioned each point with The distance of the gray level model of this feature point is to calculate to determine by following manner in training:

The smallest point of each point mahalanobis distance in normal direction is chosen, as the new position of this feature point, after all the points update, Attitude parameter b is updated, new model is generated, operating process as above is repeated, is less than until with the matched error rate of human body in figure 5%.

In order to do not done at one punctuate work picture on mark human body feature point, according to opencv carry classifier, It can accurately detect the position of human eye at this stage.Therefore, we first shine into row ASM algorithm operating to human body front. The position that people's eyes are first found out according to the classifier that opencv is carried, by our people in the average template that the training stage obtains Eye makes the difference to obtain translation distance l with it, then in standardized human body front is shone, is opened according to the available people's both hands of Face Detection Armband length, the mould using the armband lenth ratio in arm length degree and template as pantograph ratio s, then after being translated through translation scaling Plate is

WhereinFor the average template obtained in training process；It, will be in obtained average template and figure by above step The human body side of non-punctuate has carried out thick matching.It needs to be adjusted details at this time.

S5: according to ankle position in the S3 detected, similar Block- matching is carried out, obtains human body side view.It is above-mentioned to be used for Obtaining similar Block- matching in the step S5 of the image measuring method of human somatotype data is first passed through by the positive template of human body It is with identified people foot range in the process, the point within the scope of foot is for statistical analysis, show that the region within the scope of this is special Sign, including gray scale maximum value G_max, minimum value G_min, average value G_avg, tri- color proportion R of RGB_r, R_g, R_b, formed to Amount:

R=(G_msx, G_min, G_avg, R_r, R_g, R_b)

It is standardized:

R=(G_max/ 255, G_min/ 255, G_avg/ 255, R_r, R_g, R_b)；

Fig. 8, Fig. 9 are human body positive side face similar block matching result；Again by the nose position feature of side, foot position feature It determines, its length is made the difference into work as pantograph ratio, then by the foot of target side dough figurine body and template foot with the ratio in template For translation distance, then by scaling is translated with front carry out same operation, obtain side matching result, Figure 10 be matched through ASM it is special Levy the result of point.

Embodiment 3

Operating process same as Example 2, since background is sometimes sufficiently complex, we are difficult simple dependence colour of skin inspection It surveys to obtain the position where the colour of skin, therefore, we carry out trained template slightly with the human body for not carrying out punctuate on photo When matching, colour of skin matching is first carried out, the yellow position of our cuffs is detected near the colour of skin position obtained on picture.Once two The yellow position of a cuff determines, we can be by its distance d ' as the arm exhibition on our photos to be measured, average mould at this time The scaling multiple of plate is then s=d '/d, and wherein d is the arm length degree on average template.

The stance of people is varied, different people due to stand when both feet stance be it is different, particularly with Angle between both legs is different, this has resulted in the characteristic point of the position of both legs after carrying out translation scaling.It may be due to angle The deviation of degree makes the place that template point location is far in target point, and as shown in figure 11, therefore we measure on clothes according to close-fitting vest Punctuate first find both feet ankle position (x_left, y_left) and (x_right, y_right), then according to double-legged ankle and waistline White positions (x_crotch, y_crotch) the position θ that strides that two legs can be calculated, then after Pan and Zoom, to the characteristic point of two legs It operates again plus as follows

x_new=(x_i-x_crotch)cosθ-(y_i-y_crotch)sinθ+x_crotch

y_new=(x_i-x_crotch)sinθ+(y_i-y_crotch)cosθ+y_crotchFigure 12 is aobvious plus the result after the rotation of two legs Show, both legs have biggish improvement than before.

Similarly, in the human body of side, the feature for being previously noted available foot areas carries out similar Block- matching.This method is certain Can find the position of people's foot in the photo of side, but due to people's foot of a front surface and a side surface be not it is identical, find People's placement of foot still have deviation, we carry out auxiliary search using the ankle that above-mentioned yellow spot navigates at this time, Search range can be further reduced, thus the more accurate position for navigating to people's foot.

In real life, according to measurement method herein come when measuring somatic data, we take in no measurement In the case where can also complete in fact.For example, we can be marked on the clothes of our pure colors with corresponding color, Such as bound on a corresponding position using colored ribbon bandage to make marks.The human body front side of shooting is taken as measurement picture, Characteristic point can be updated on corresponding color position by color-match in the algorithm, as shown in figs. 13 and 14, utilization is this Method, also available preferable measurement result.

Whether the experimental measurements that the application is carried out are accurate, whether can converge to human body by judging characteristic point Contour edge, and guarantee shape composed by last characteristic point can with target body as far as possible mutually recently judgement.It is as follows It, all preferably will be on the key position of Feature Points Matching to human body on positive side when human body dress measurement clothes measure shown in Figure 15. Chest, waist, the position of stern key position characteristic point fitting are also relatively more accurate.Thus method shown in this article can be preferable, more convenient The human somatotype coordinate data got in picture.

Claims (7)

1. a kind of for obtaining the image measuring method of human somatotype data, characterized by the following steps:

S1: the inflection point on the positive side of human body is chosen as characteristic point, wherein choose 65 characteristic points in front, choose in side 27 characteristic points, record the coordinate information of each characteristic point；

S2: by carrying out dimension-reduction treatment to the data in characteristic point sample set obtained in step S1, the flat of body shape is obtained Equal template；

S3: referring to the position of characteristic point in S1, in the neckline for the clothes worn for user, cuff, bust, waistline, hip circumference and foot Ankle position is labeled；

S4: the distance between two cuffs in S3, double-legged ankle, the position of both legs and waistline are obtained according to mark Face Detection Position obtains human body front shape after being computed；

S5: according to ankle position in the S3 detected, similar Block- matching is carried out, obtains human body side view.

2. as described in claim 1 for obtaining the image measuring method of human somatotype data, it is characterised in that: selected in S1 The characteristic point taken, in the outside boundary line of human body contour outline；The data record of each characteristic point are as follows:

X_i=(x_i0, y_i0, x_i1, y_i1..., x_iky_ik... x_i(n-1)y_i(n-1))^TWherein (x_ik, y_ik) represent k-th point of seat in figure Mark, human body front sample n_Just=65, side n_Side=27, i.e. sample set X=[X of the acquisition containing N group training sample₁, X₂, ...X_N], N=92=n_Just+n_Side。

3. as claimed in claim 2 for obtaining the image measuring method of human somatotype data, it is characterised in that: to institute in S2 The data of acquisition carry out registration process；I.e. by first sample X₁=(x₁₀, y₁₀, x₁₁, y₁₁..., x_1k, y_1k..., x_1(n-1), y_1(n-1))^TAs master sample, to each sample X thereafter_i=(x_i0, y_i0, x_i1, y_i1..., x_ik, y_ik..., x_i(n-1) y_i(n-1))^TCarry out scaling and position translation；Detailed process is as follows:

S21: the translational movement of each sample is calculated:

The center of gravity of first sample is calculated first are as follows:

The then translational movement of i-th of sample are as follows:

S22: the scaling of each sample is calculated:

Horizontal and vertical average conduct scaling S is taken, is calculated as follows:

Wherein: Max (X_i(x), Min (X_i(x), Max (Y_i(x), Min (Y_i(x) maximum value in i-th of sample on the direction x is indicated, Minimum value, maximum value and minimum value on the direction y；According to T_iAnd S_iCalculate i-th of sample:

S23: being iterated, and determines final sample set:

Same operation step is carried out again by step S21, S23, finally obtains sample set:

S24: dimension-reduction treatment therein is that sample obtained in step S23 is handled as follows:

S241: the average shape of sample set is acquired:

S242: the deviation of each sample and average shape in sample set is sought:

S243: constituting the covariance matrix of sample set, enables:

D=(dx₀|dx₁|...|dx_n-1)

The then covariance matrix of sample set:

S244: the characteristic value and feature vector of covariance matrix are calculated, and characteristic value is sorted

∑p_k=λ_kp_k,

Wherein λ_kIt is the big characteristic value of K of covariance matrix；

S245: choosing corresponding feature vector, t characteristic value before choosing first in k characteristic value, so that:

There is a feature vector corresponding the characteristic value of each acquirement, is formed new matrix P

P=(p₀|p₁|...|p_t-1)

Then any one body shape is represented by

Wherein b is the vector of t × 1, and different b, which brings above formula into, can be obtained different body shapes.

4. as described in claim 1 for obtaining the image measuring method of human somatotype data, it is characterised in that: for both legs Position characteristic point matching with the following method: first find both feet ankle position (x_left, y_left) and (x_right, y_right), then according to the labeling position (x of double-legged ankle and waistline_crotch, y_crotch) the position θ that strides that calculates two legs, then flat After moving and scaling, lower operation is subject to again to the characteristic point of two legs:

x_new=(x_i-x_crotch)cosθ-(y_i-y_crotch)sinθ+x_crotch

y_new=(x_i-x_crotch)sinθ+(y_i-y_crotch)cosθ+y_crotch。

5. as described in claim 1 for obtaining the image measuring method of human somatotype data, it is characterised in that: in S4 and S5 Between increase a step S42, establish characteristic point local gray level model, and calculate each point gray scale and training in this feature point Gray level model distance；

The local gray level model is realized by following step:

S421: sample each method of characteristic point to two sides respectively take k point, calculate the gray scale of the point, then share 2k+1 it is grey Angle value, including characteristic point itself；Then the gray scale layer of j-th of characteristic point of i-th of sample can be denoted as:

g_ij=(g_ij1, g_ij2..., g_ij(2k+1))^T；

Wherein, i=1,2 ... N, for human body front j=0,1,2 ..., 64, human body side j=0,1,2 ..., 26；

S422: to g_ijDifference is asked to obtain:

dg_ij=(g_ij2-g_ij1, g_ij4-g_ij3..., g_ij(2k+1)-g_ij(2k))^T；

After standardization:

S423: the mean value of j-th of characteristic point gray scale layer is calculatedWith variance cov_j:

6. as claimed in claim 5 for obtaining the image measuring method of human somatotype data, it is characterised in that: each point Gray scale is to calculate to determine by following manner at a distance from the gray level model of this feature point in training:

The smallest point of each point mahalanobis distance in normal direction is chosen, as the new position of this feature point, after all the points update, is updated Attitude parameter b, generates new model, repeats operating process as above, until with the matched error rate of human body in figure less than 5%.

7. as described in claim 1 for obtaining the image measuring method of human somatotype data, it is characterised in that: similar in S5 Block- matching is to first pass through the people foot range determined by during the positive template matching of human body, by the click-through within the scope of foot Row statistical analysis, obtains the provincial characteristics within the scope of this, including gray scale maximum value G_max, minimum value G_min, average value G_avg, RGB tri- Color proportion R_r, R_g, R_b, formed vector:

R=(G_max, G_min, G_avg, R_r, R_g, R_b)

It is standardized:

R=(G_max/ 255, G_min/ 255, G_avg/ 255, R_r, R_g, R_b)；

The nose position feature of side, foot position feature are determined again, using the ratio in its length and template as pantograph ratio, The foot of target side dough figurine body and template foot are made the difference as translation distance again, then by behaviour identical as front translation scaling carry out Make, obtains side matching result.