CN106599795A - Dynamic low-resolution pedestrian re-identification method based on scale distance gradient function interface learning - Google Patents

Abstract

The invention discloses a dynamic low-resolution pedestrian re-identification method based on scale distance gradient function interface learning, and is to obtain a good re-identification effect when images are low in resolution and change in scale. The method comprises the following steps: to begin with, obtaining an image pair, that is, a positive sample pair, from the same person, and image pairs, that is, negative sample pairs, from different persons; then, introducing distance scale gradient functions, wherein the positive sample pair generates a feasible distance scale gradient function, and the negative sample pair generates an infeasible distance scale gradient function, and the two kinds of distance scale gradient functions form a distance scale gradient function space, a positive parameter vector expressing the feasible distance scale gradient function, and a negative parameter vector expressing the infeasible distance scale gradient function; and finally, carrying out classification through a trained random forest classifier to judge whether the images are from the same object according to the information that whether some distance scale gradient function is in a positive domain or a negative domain in the function space.

Description

A kind of dynamic low resolution pedestrian learnt apart from tapering function separating surface based on yardstick Recognition methodss again

Technical field

The invention belongs to monitor video retrieval technique field, is related to a kind of pedestrian recognition methodss again, more particularly to a kind of base In the dynamic low resolution pedestrian recognition methodss again that yardstick learns apart from tapering function separating surface.

Background technology

In the last few years, monitor network and apply and spread to public safety, mobile detection, passenger flow statisticses etc. more and more widely Field, video monitoring play more and more important effect in terms of fighting crime, safeguarding social safety, and video investigation becomes public The effective means of the machine-operated solving criminal cases of peace.But in actual video investigation, investigator needs taken the photograph according to specified pedestrian's object more Suspected target is quickly investigated, is followed the trail of and lock as the moving frame under head and track, need to expend substantial amounts of human and material resources and Time, efficiency of solving a case not only is affected, and easily misses optimal solving a case opportunity.It is that one kind is directed to specific pedestrian that pedestrian recognizes again Across the camera head monitor video automatically retrieval technology of object, i.e., match same a group traveling together under the non-overlapping multi-cam of irradiation area Object, the technology can aid in video investigation person quickly and accurately to find moving frame and the track of suspected target, to the Ministry of Public Security Door improves case-solving rate, safeguards that life property safety of people is significant.

It is in pedestrian's weight identification process of multi-cam, especially in the extraction and matching process of pedestrian's feature, different Often there is visual angle change, change in size, illumination variation etc. in the pedestrian image under photographic head, so as to cause the drop of weight discrimination It is low.Nowadays the crime means of offender are more and more brilliant, hide oneself by various modes, if they avoid shooting Head, so treats to recognize that the number of times that target occurs under photographic head is less again, undoubtedly can make troubles to identification work again.Pedestrian's weight Identification is to carry out target query using image, therefore to the image that obtains mined information to the full extent, to lifting pedestrian Weight recognition performance is significant.

Recognition methodss can be divided into two classes to existing pedestrian again：The first kind mainly constructs the visual signature of robust, then uses The distance function (such as Euclidean distance etc.) of standard carries out similarity measurement；Equations of The Second Kind is mainly entered by learning a suitable yardstick Capable more accurately distance metric.The image adopted based on the heavy technology of identification of said method is to being same resolution and same chi mostly Degree, but in some real scenes, pedestrian image is not only low resolution, and is different scale.

Chinese patent literature CN104298992A, open (bulletin) day 2015.02.01, discloses a kind of based on data The dimension self-adaption pedestrian recognition methodss again of driving, the invention is by obtaining inquiry pedestrian and pedestrian to be measured using sparse method Between cross-domain support concordance come calculate pedestrian to the distance between, using concordance of the training data under different visual angles come Automatic adjusument yardstick, it is change that the method take into account the yardstick of pedestrian image, but does not account for pedestrian image Resolution is different, therefore the result that the algorithm is obtained is nor optimal.

Chinese patent literature CN103793702A, open (bulletin) day 2014.05.14, discloses a kind of based on collaboration The pedestrian of scale learning recognition methodss again, the invention are studied the pedestrian based on scale learning under semi-supervised framework and recognize skill again Art, carries out scale learning by not marking sample auxiliary mark sample, and the method does not account for the yardstick of pedestrian image pair, divides Resolution change meeting counterweight recognition result produces impact, thus heavy recognition performance of the method under few mark sample need to be carried Rise.

Chinese patent literature CN105005797A, open (bulletin) day 2015.10.28, discloses a kind of based on negative and positive The Tai Ji relative distance measure of Bi-objective sample, the method set up virtual the moon sample for each positive sample (original sample), Learn out distance function transformation matrix using relative distance, the method does not account for the yardstick of sample image, change resolution When what can produce on the metric learning method affect, thus the method also optimizes space.

Chinese patent literature CN105138998A, open (bulletin) day 2015.12.09, discloses a kind of based on visual angle The pedestrian of adaptive space learning algorithm recognition methodss and system again, the method pass through visual angle self adaptation sub-space learning algorithm To transformation matrix, row distance is calculated and pedestrian recognizes again to recycle transformation matrix to enter for acquistion.Although the method belongs to pedestrian and knows again Other field, but with the research of the dynamic low resolution pedestrian that learnt apart from tapering function separating surface based on yardstick again recognition methodss Angle is different.

Chinese patent literature CN105095475A, open (bulletin) day 2015.11.25, discloses a kind of based on two-stage Imperfect attribute labelling pedestrian recognition methodss and the system again of fusion；Chinese patent literature CN104462550A, it is open (public Accuse) day 2015.03.25, disclose a kind of pedestrian's recognition methodss again for merging sorting consistence based on similarity and dissimilarity；This Two methods lift the effect that pedestrian recognizes again by being all ranking results of the fusion under two kinds of different situations.In in addition State Patent document number CN104200206A, open (bulletin) day 2014.12.10, discloses a kind of based on double angle sorting consistences Pedestrian's recognition methodss again, the method obtains initial ranking results based on the method for global characteristics and local feature, then carries out It is further to reset, and then preferably weighed recognition effect；And Chinese patent literature CN103810476A, it is open (public Accuse) day 2014.05.21, pedestrian's recognition methodss again in a kind of video surveillance network based on small group information association are disclosed, should Relatedness of the method according to small group information, the key character that pedestrian's groupuscule body characteristicses are recognized again as pedestrian, mainly Solve the problems, such as in video surveillance network that pedestrian's weight recognition accuracy is low, precision is not high；These methods are also all belonging to pedestrian's weight Identification field, but the dynamic low resolution pedestrian recognition methodss again learnt apart from tapering function separating surface based on yardstick with us Research angle be also different.

Chinese patent literature CN104268583A, open (bulletin) day 2015.01.07, discloses based on color region The pedestrian of feature recognition methodss and system again；Chinese patent literature CN104484324A, open (bulletin) day 2015.04.01, The pedestrian retrieval method of a kind of multi-model and fuzzy color is disclosed, this two patents mainly make full use of the color point of pedestrian Improving the accuracy rate that pedestrian recognizes again, this is with us by the dynamic that learnt apart from tapering function based on yardstick for cloth structural information The angle of low resolution pedestrian recognition methodss again is also different.

The content of the invention

In view of the shortcomings of the prior art, the invention provides a kind of learnt apart from tapering function separating surface based on yardstick Dynamic low resolution pedestrian recognition methodss again, the method is by drawing in storehouse graphical rule and inquiry -- the image distance in storehouse Function, i.e. yardstick will be classified apart from tapering function to the yardstick for producing from different images, use one apart from tapering function Kind new model recognized again carrying out pedestrian, and then lifts the accuracy of multi-cam similarly hereinafter a group traveling together's matching.

The technical solution adopted in the present invention is：

A kind of dynamic low resolution pedestrian recognition methodss again learnt apart from tapering function separating surface based on yardstick, its feature It is to comprise the following steps：

Step 1：Extract picture visual properties, specifically by training data be divided into positive sample to and negative sample pair, wherein Each pair image is from different photographic head, if an inquiry -- in storehouse, image is to for (I_i,I_j), do not considering inferior image resolution In the case of different, visual properties xs of the every image I under identical dimensional m is extracted；

Step 2：The visual properties of the picture that step 1 is extracted carry out expression of the yardstick apart from tapering function, the function Expression is made up of some steps, specifically includes following sub-step：

Step 2.1, obtains a series of visual properties of image；

Step 2.2, constitutes yardstick apart from tapering function space；

Step 2.3, represents distance scale tapering function with parameter vector；

Step 3：The result obtained based on step 2 produces test parameter vector, using linear interpolation method to entering in storehouse image Row is up-sampled, and so as to obtain more observation data, online weight cognitive phase has obtained the observation with training stage equal number Data, by the recurrence for observing data, are inquired about -- the test parameter vector of image pair in storehouse；

Step 4：Random forest grader, is specifically included to classifying come to image：

Step 4.1, trains random forest grader；

Step 4.2, the result obtained based on step 4.1 are classified.

In a kind of above-mentioned dynamic low resolution pedestrian learnt apart from tapering function separating surface based on the yardstick again side of identification Method, the concrete grammar of the step 2.1 include：Define I_iVisual properties beWherein, 1 expression sampled images be resolution and The scale ratio of the resolution of original image；Then step by step to I_jDown-sampling, 0.01 scale-value of down-sampling reduction each time, leads to Cross this mode to obtainA series of visual properties, wherein scale ratio illustrate two figures The height and width of picture；

In a kind of above-mentioned dynamic low resolution pedestrian learnt apart from tapering function separating surface based on the yardstick again side of identification Method, the concrete grammar of the step 2.2 include：Every image is divided into 8 rows 3 are arranged 24 pieces, extract 64 WeiHSVTe per block Levy, so every image is showed by 1536 dimensional features, then calculate the resolution in storehouse image and original image of down-sampling Rate scale ratio k, and I_iAnd I_jBetween Euclidean distanceDraw with k as abscissa, d is vertical seat Target yardstick apart from tapering function, for the ease of observation, by the Euclidean distance between image pair be converted to it is another form of away from FromEach image is to (I_i,I_j) conversion be change with scale ratio k apart from d' Change and change；In off-line training step, positive sample is infeasible to producing to producing feasible distance scale tapering function, negative sample Yardstick apart from tapering function, feasible and infeasible yardstick apart from tapering function constitute yardstick apart from tapering function space；

In a kind of above-mentioned dynamic low resolution pedestrian learnt apart from tapering function separating surface based on the yardstick again side of identification Method, the concrete grammar of the step 2.3 include：Yardstick is represented using parameter vector apart from tapering function, Definition Model is f (), w is the parameter vector of this model, and model use K=96 is to training image to (I_i,I_j) observation data (k, d'_i,j (k)), if f is k to d'_i,jUnified model, then

d'_i,j(k)=f (k, w), k ∈ [0.05,1] (1)

Defined parameters vector w=(w₀,...,w_n,...,w_N-1) be N-dimensional, then image is to (I_i,I_j) parameter vector w_i,j For：

Wherein, model f () is model of fit, based on following multinomial model：

In a kind of above-mentioned dynamic low resolution pedestrian learnt apart from tapering function separating surface based on the yardstick again side of identification Method, the concrete grammar of the step 4.1 include：In feasible and infeasible yardstick between tapering function, parameter vector There may not be enough discernments, Random Forest model trains random forest grader using parameter vector, wherein, offline Training stage yardstick represents model from the positive sample parameter vector positive to generation apart from tapering function, from negative sample to producing what is born Parameter vector.

In a kind of above-mentioned dynamic low resolution pedestrian learnt apart from tapering function separating surface based on the yardstick again side of identification Method, the concrete grammar of the step 4.2 include：

Cognitive phase is weighed apart from the online of tapering function space in yardstick, the random forest that trained is by judging a figure As whether whether the yardstick to producing is classified as from same object in positive domain or negative domain apart from tapering function, concrete point Realize for two sub-steps：

Step 4.21, changes the yardstick in storehouse image and obtains visual properties；

Step 4.22, processes the visual properties and distance of retrieval image.

The present invention has advantages below and beneficial effect：1st, invention introduces yardstick is apart from tapering function, when in storehouse figure The resolution of picture is low and during unstable dimensional variation, can also carry out effectively recognizing again；2nd, invention introduces dynamic low resolution Rate weighs recognition performance to improve, and the optimization in image resolution ratio and nahoscale-level causes the expansion of method and the suitability very By force.

Description of the drawings

Fig. 1 is the inventive method flow chart.

Specific embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with the accompanying drawings and embodiment is to this It is bright to be described in further detail, it will be appreciated that enforcement example described herein is merely to illustrate and explains the present invention, not For limiting the present invention.

The present invention is a kind of dynamic low resolution pedestrian learnt apart from tapering function separating surface based on yardstick side of identification again Method.The image from same person is obtained first to i.e. positive sample pair, and from different people image to being negative sample pair；So Afterwards by positive sample to producing feasible yardstick apart from tapering function, by negative sample to producing infeasible yardstick apart from gradual change letter Number, both yardsticks constitute yardstick apart from tapering function space apart from tapering function, and represent feasible by positive parameter vector Apart from tapering function, negative parameter vector represents infeasible yardstick apart from tapering function to yardstick；What is finally trained is random gloomy Woods grader in the function space according to certain yardstick apart from tapering function whether in positive domain or negative domain, be classified as whether coming From same object so that, obtain more when storehouse image is low resolution and dimensional variation when query image is high-resolution Good heavy recognition effect.

Fig. 1 is asked for an interview, the present embodiment adopts MATLAB7 as Simulation Experimental Platform, in the simulation based on VIPeR data sets Pedestrian data set SALR-VIPeR, the pedestrian's data set SALR-PRID simulated based on PRID450S data sets and data set Tested on CAVIAR.VIPeR data sets have 632 pedestrian images pair under two photographic head, deposit between two photographic head In differences such as obvious visual angle, illumination；PRID450S data sets have 450 lists under the disjoint photographic head of two spaces Scape image pair, and the data set has lasting light change and gorgeous degree change；CAVIAR data sets are widely used in pedestrian to be known again Not, the data set has the photo of 72 people under the photographic head of certain shopping mall two.

The present invention flow process be：

Step 1：By training data be divided into positive sample to and negative sample pair, wherein each pair image from different photographic head, If an inquiry -- in storehouse, image is to for (I_i,I_j), in the case where image resolution ratio difference is not considered, extract every image I and exist Visual properties x under identical dimensional m；

Step 2：Expression of the yardstick apart from tapering function, which is implemented including following sub-step：

Step 2.1：In order to generate yardstick apart from tapering function, if I_iVisual signature be(1 represents that sampled images are differentiated The scale ratio of rate and original image resolution), then step by step to I_jDown-sampling, 0.01 yardstick of down-sampling reduction each time Value, obtainsA series of visual signature, wherein scale ratio illustrate the height of two images Degree and width；

Step 2.2：Every image is divided into 8 rows 3 are arranged 24 pieces, extract the 64 dimension HSV features per block, so per a figure As being showed by 1536 dimensional features；Then the resolution-scale in storehouse image and original image of down-sampling is calculated than k, and I_iAnd I_jBetween Euclidean distanceDraw with k as abscissa, d is the yardstick of vertical coordinate apart from gradually Euclidean distance between image pair, for the ease of observation, is converted to another form of distance by varying functionEach image is to (I_i,I_j) conversion be change with scale ratio k apart from d' And change；In off-line training step, to producing feasible yardstick apart from tapering function, negative sample is infeasible to producing for positive sample Yardstick apart from tapering function, feasible and infeasible yardstick apart from tapering function constitute yardstick apart from tapering function space；

Step 2.3：In order to separate above two function, we represent yardstick apart from tapering function using parameter vector, We set model as f (), w is the parameter vector of this model, and model use K=96 is to training image to (I_i,I_j) see Survey data (k, d'_i,j(k)), if f is k to d'_i,jUnified model, then

d'_i,j(k)=f (k, w), k ∈ [0.05,1] (1)

Assume parameter vector w=(w₀,...,w_n,...,w_N-1) be N-dimensional, then image is to (I_i,I_j) parameter vector w_i,j For：

We using model of fit be following multinomial model：

Step 3：Online heavy cognitive phase is arrived, low resolution is appointed in storehouse image and so retain low resolution, tests in storehouse figure Scale ratio k of picture equally carries out down-sampling step by step from 1 to 0.05, and yardstick represents model in this way apart from tapering function A series of observation data are obtained, for regression process, if the resolution in storehouse image is very low, observation data are limited , thus we using linear interpolation method to up-sampling in storehouse image, so as to obtain more observation data, know again online The other stage we obtained and training stage equal number observation data, by observe data recurrence, one inquiry -- The test parameter vector of storehouse image pair is just generated；

Step 4：In feasible and infeasible yardstick between tapering function, parameter vector may not have enough Discernment, using parameter vector, (off-line training step yardstick represents model by positive sample apart from tapering function to Random Forest model To producing positive parameter vector, negative sample is to producing negative parameter vector) training random forest grader；

Step 5：Cognitive phase is weighed apart from the online of tapering function space in yardstick, the random forest that trained is by judging Whether one yardstick is classified as, whether from same object, being divided into two sub-steps in positive domain or negative domain apart from tapering function Suddenly realizing：

Step 5.1：Change the yardstick in storehouse image and obtain visual properties；

Step 5.2：Process the visual properties and distance of retrieval image.

The result tested on data set from said process is it is found that using yardstick yardstick apart from tapering function (SDF) Than using good many of the result of other two methods, such that it is able to illustrate based on yardstick apart from moving that tapering function separating surface learn The retrieval performance of the pedestrian of state low resolution recognition methodss again is significantly improved.

The application is as shown in the table with the contrast of above-mentioned 10 publications：

It should be appreciated that the part that this specification is not elaborated belongs to prior art.

It should be appreciated that the above-mentioned description for preferred embodiment is more detailed, therefore can not be considered to this The restriction of invention patent protection scope, one of ordinary skill in the art are being weighed without departing from the present invention under the enlightenment of the present invention Under the protected ambit of profit requirement, replacement can also be made or deformed, be each fallen within protection scope of the present invention, this It is bright scope is claimed to be defined by claims.

Claims (6)

1. a kind of dynamic low resolution pedestrian recognition methodss again learnt apart from tapering function separating surface based on yardstick, its feature are existed In comprising the following steps：

Step 1：Extract the visual properties of picture, specifically by training data be divided into positive sample to and negative sample pair, wherein each pair Image is from different photographic head, if an inquiry -- in storehouse, image is to for (I_i,I_j), do not considering image resolution ratio difference In the case of, extract visual properties xs of the every image I under identical dimensional m；

Step 2：The visual properties of the picture that step 1 is extracted carry out expression of the yardstick apart from tapering function, the expression of the function It is made up of some steps, specifically includes following sub-step：

Step 2.1, obtains a series of visual properties of image；

Step 2.2, constitutes yardstick apart from tapering function space；

Step 2.3, represents distance scale tapering function with parameter vector；

Step 3：The result obtained based on step 2 produces test parameter vector, using linear interpolation method to carrying out in storehouse image Sample, so as to obtain more observation data, online weight cognitive phase has obtained the observation data with training stage equal number, By the recurrence for observing data, inquired about -- the test parameter vector of image pair in storehouse；

Step 4：Random forest grader, is specifically included to classifying come to image：

Step 4.1, trains random forest grader；

Step 4.2, the result obtained based on step 4.1 are classified.

2. a kind of dynamic low resolution pedestrian learnt apart from tapering function separating surface based on yardstick according to claim 1 Recognition methodss again, it is characterised in that the concrete grammar of the step 2.1 includes：Define I_iVisual properties beWherein, 1 table Show that sampled images are the scale ratios of the resolution of resolution and original image；Then step by step to I_jDown-sampling, is adopted down each time Sample reduces by 0.01 scale-value, obtains in this wayA series of visual properties, Wherein scale ratio illustrates the height and width of two images.

3. a kind of dynamic low resolution pedestrian learnt apart from tapering function separating surface based on yardstick according to claim 3 Recognition methodss again, it is characterised in that the concrete grammar of the step 2.2 includes：Every image is divided into 8 rows 3 are arranged 24 pieces, carry Take out the 64 dimension HSV features per block, so every image shows by 1536 dimensional features, then calculate down-sampling in storehouse figure The resolution-scale of picture and original image is than k, and I_iAnd I_jBetween Euclidean distanceDraw With k as abscissa, d for vertical coordinate yardstick apart from tapering function, for the ease of observation, the Euclidean distance between image pair is turned It is changed to another form of distanceEach image is to (I_i,I_j) conversion apart from d' It is to change with the change of scale ratio k；In off-line training step, the positive sample distance scale tapering function feasible to generation, Negative sample to producing infeasible yardstick apart from tapering function, feasible and infeasible yardstick apart from tapering function composition yardstick away from From tapering function space.

4. a kind of dynamic low resolution pedestrian learnt apart from tapering function separating surface based on yardstick according to claim 1 Recognition methodss again, it is characterised in that the concrete grammar of the step 2.3 includes：Yardstick distance is represented using parameter vector gradually Varying function, Definition Model are f (), w is the parameter vector of this model, and model use K=96 is to training image pair (I_i,I_j) observation data (k, d'_i,j(k)), if f is k to d'_i,jUnified model, then

d'_i,j(k)=f (k, w), k ∈ [0.05,1] (1)

Defined parameters vector w=(w₀,...,w_n,...,w_N-1) be N-dimensional, then image is to (I_i,I_j) parameter vector w_i,jFor：

$w_{i,j}=\underset{w}{\arg \min }\frac{1}{K}\underset{k\∈\[0.05,1\]}{\Σ}|{d^{\′}}_{i,j}\left(k\right)-f(k,w){|}^2+\mathrm{\λ}\underset{n=0}{\overset{N-1}{\Σ}}|w_n|---\left(2\right)$

Wherein, model f () is model of fit, based on following multinomial model：

$f(k,w)=\underset{n=0}{\overset{N-1}{\Σ}}{w}_n{k}^n---\left(3\right).$

5. a kind of dynamic low resolution pedestrian learnt apart from tapering function separating surface based on yardstick according to claim 1 Recognition methodss again, it is characterised in that the concrete grammar of the step 4.1 includes：In feasible and infeasible yardstick distance gradually Between varying function, parameter vector may not have enough discernments, and Random Forest model is trained using parameter vector at random Forest classified device, wherein, off-line training step yardstick apart from tapering function represent model from positive sample to produce positive parameter to Amount, from negative sample to producing negative parameter vector.

6. a kind of dynamic low resolution pedestrian learnt apart from tapering function separating surface based on yardstick according to claim 1 Recognition methodss again, it is characterised in that the concrete grammar of the step 4.2 includes：

Cognitive phase is weighed apart from the online of tapering function space in yardstick, the random forest that trained is by judging an image pair Whether the yardstick of generation is classified as, whether from same object, being specifically divided into two in positive domain or negative domain apart from tapering function Sub-steps are realizing：

Step 4.21, changes the yardstick in storehouse image and obtains visual properties；

Step 4.22, processes the visual properties and distance of retrieval image.