CN106803063A - A kind of metric learning method that pedestrian recognizes again - Google Patents

Abstract

The invention discloses a kind of metric learning method that pedestrian recognizes again, comprise the following steps：Initially set up the set of pedestrian's target signature under different cameras；Then the metric form of mahalanobis distance is used, and adds constraints and the same clarification of objective metric range under different cameras is constrained to 0, the characteristic distance of different target is constrained to constant μ (μ>0)；Finally, set up loss function and optimize its structure, the optimal metric matrix for meeting constraints is drawn by Projected descent method iteration, complete measurement learning process.The problems such as present invention efficiently solves over-fitting present in existing metric learning method, metric matrix to noise-sensitive, it is adaptable to the application scenario that pedestrian recognizes again under complex scene.

Description

A kind of metric learning method that pedestrian recognizes again

Technical field

The invention belongs to mode identification technology, more particularly, to a kind of metric learning recognized for pedestrian again Method and system.

Background technology

The heavy recognizer of pedestrian is one of key areas of image procossing and pattern identification research, is conceived to without public The identification work of specific objective pedestrian under the camera of the ken.Method more general at present is to find a kind of based on pedestrian's appearance Feature representation, mainly include the information such as color and texture, reuse the phase between a kind of suitable measure calculating target Like degree, and output of sorting.The factor such as being blocked by visual angle, illumination, object under different cameras due to same target is influenceed, its Under different visual angles often there is deviation in the expression of feature, and the feature representation target except choosing more robust, selection is a kind of suitable Measure be also solve matching problem core.For the special scenes of a certain database, by training set sample Practise and produce a more suitable metric space to carry out Similarity Measure, this method is referred to as metric learning.Existing degree In amount learning method, often pursue inter- object distance in the training stage and minimize and between class distance maximization, ignore target signature Part, cause weight be extended of the information such as background, noise in measurement, the result for obtaining often to there is over-fitting Risk, cause test less effective.The present invention proposes a kind of metric learning method based on iso-distance constraint, to a certain degree On solve the problems, such as traditional measure learn over-fitting.

The content of the invention

The present invention proposes a kind of metric learning method that pedestrian recognizes again, it is therefore intended that provide a kind of based on iso-distance constraint Metric learning method, solves the problems, such as the over-fitting that existing metric learning technology is present.

The metric learning method that a kind of pedestrian proposed by the present invention recognizes again, comprises the following steps：

(1) set up goal set (extending to multiple) under two different cameras, respectively constitute set X and set Z, X, Element in Z set is characteristic vector (including color and Texture eigenvalue) of each target in the camera hypograph；

(2) mahalanobis distance measure is used, the distance between X, Z set any two element is calculated：In formula, x_iIt is the target i characteristic vectors of set X, z_jIt is the target j of set Z Characteristic vector, sets up positive sample to set S, and the element in S is sample to (x_i,z_j), x_i、z_jBelong to same under different cameras The characteristic vector of individual pedestrian target；Negative sample is set up simultaneously to set D, and the element in D is sample to (x_i,z_j), x_i、z_jBelong to The feature representation of different pedestrian targets under different cameras；Metric matrix M is initialized as unit matrix；

(3) constrain that to belong to the distance of same target signature in all X, Z set be 0, the characteristic vector of different target away from From being a constant constant μ, span is [2,4], sets up loss function：

In formula, feature of the same pedestrian target under different cameras is to (being defined as positive sample in | S | expression set X, Z It is right) number, | D | represents the number that the feature of same person is not belonging in set X, Z to (be defined as negative sample to)；γ is To the weight for loss function, interval is [0.5,0.7] to positive and negative samples；For F norms square, be regularization , for avoiding algorithm over-fitting；λ spans are [5 × 10^-6, 1 × 10^-4], with the intensity for making adjustments regularization；

(4) Projected descent method, iteration is used to seek the optimal value M of matrix M loss function L (M)^*, make loss function L (M^*) minimum；；In+1 iteration of kth, to kth time iteration optimal value M_kOptimized by gradient direction, obtaineds_kIt is step-size factor, value [0.05,0.5]；

(5) discrimination matrix C_kWhether positive semidefinite, be to make(6) are gone to step, the positive semidefinite being otherwise projected into Most like matrix in space, specific method is that the positive semidefinite square with two Norm minimums of its difference is found in positive semidefinite space Battle arrayInstead of C_k；

(6) willWith the M before iteration_kWeight α and (1- α) are assigned respectively, and new M is formed after linear combination_k+1, as kth The Metzler matrix final result of+1 suboptimization, α spans for (0,1]；

(7) every time after iteration the absolute value of the difference of counting loss function and a preceding loss function ratio, if being less than A certain default threshold epsilon, then judge that iteration terminates, and obtains the optimal result of M, otherwise goes to step (4) and continues to iterate to calculate；It is described ε values are according to computational accuracy and calculate balance determination, can interval [1 × 10^-6, 1 × 10^-4]。

Further, in the step (5), most like matrix method is found as follows：

To C_kSingular value decomposition is carried out, is obtainedI is unit matrix, Λ_kIt is by C_KVery The diagonal matrix of different value composition；

According to C_kObtain most like matrixOrderWherein

According toAnd M_k, obtain M_k+1, calculating formula is：

Obtain most like positive semidefinite matrix；

In formula, α_kBe (0,1] a step-length, represent update weight.

The present invention provides a kind of object matching of metric learning method based on iso-distance constraint to process pedestrian recognizes again in Problem.Wherein, the present invention directly enters row constraint to the positive and negative samples during metric learning the distance between feature.Positive sample Characteristic distance is mapped to the same of object space to being that similarly hereinafter a group traveling together's clarification of objective distance restraint is 0 to different cameras Point.Meanwhile, it is that the characteristic distance under different cameras between different pedestrian targets is constrained to one to characteristic distance by negative sample Constant μ (μ>0), it is therefore an objective to ensure the maximization of minimum between class distance, weight accuracy of identification is improved.Instructed under such constraints Metric matrix is practised, robustness is stronger.

In general, by the above technical scheme that the present invention is contemplated, compared with prior art, mainly possess following Technological merit：

1st, the present invention is different from the constraints of the metric learning of usual use, but proposes more superior iso-distance constraint Metric learning method.Its feature is that the characteristic distance constrained between same pedestrian target is 0, all of different pedestrian targets it Between characteristic distance all be a specific constant μ (μ>0), rather than the distance allowed under usual selected binary constraints between class Difference with inter- object distance is more than a constant, or a method for threshold value is respectively provided with ternary constraints.Do so is managed By according to being, under conditions of it is assumed that total between class distance is fixed, iso-distance constraint can just allow minimum between class distance to maximize, so that Further reduce the possibility divided by mistake, the accuracy that raising is recognized again.

2nd, the algorithm generalization ability is strong, reduces measurement over-fitting and occurs, can be to the strong adaptability of feature, particularly Contain noisy feature for some, its measurement effect is better than other algorithms.

3rd, the algorithm construction simple structure, efficiency is higher under the computational methods of Projected descent method, and normal at each Database (such as VIPer, CUHK03, CUHK01 storehouse) for recognizing again tests effect excellent performance.

Brief description of the drawings

Fig. 1 is general flow chart of the invention；

Fig. 2 is the flow chart solved by iteration optimization；

Fig. 3 is the characteristic vector distribution schematic diagram directly perceived that the present invention obtains positive and negative samples in metric space.

Specific embodiment

In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and It is not used in the restriction present invention.As long as additionally, technical characteristic involved in invention described below each implementation method Not constituting conflict each other can just be mutually combined.

Hereinafter the term that the present invention is used is explained and illustrated first.

Positive semidefinite matrix：The Hermitian matrix M of one n*n is that the condition of positive definite is that and if only if for each non-zero Complex vector z, there is z^*Mz>0, then M is called positive definite matrix, wherein z^*Represent the conjugate transposition of z.Work as z^*Mz>0 reduction is z^*Mz≥0 When, claim M to be positive semidefinite matrix.Because M is Hermitian matrix, it is computed understanding, for arbitrary complex vector z, z^*Mz certainty It is real number, such that it is able to compare size with 0

Loss function：For estimating predicted value f (x) of model and the inconsistent degree of actual value y, it is a non-negative reality Value function, often represents, loss function is smaller, and the robust performance of model is better using L (y, f (x)).

As shown in figure 1, the general flow chart of the metric learning for iso-distance constraint of the present invention, the inventive method specifically includes following Step：

(1) all training samples to database extract feature (including color and texture information), set up set { X；Z； Y }, X={ x₁,x₂.....x_n}∈R^d*nRepresent the set of n sample characteristics under camera A, the wherein feature dimensions of each sample It is d to spend.Z={ z₁,z₂.....z_m}∈R^d*mRepresent the set of m sample characteristics under camera B, the dimension of each sample characteristics Degree remains as d.Y∈R^n*mIt is the matching matrix of set X and Z.Such as y_ij=1 represents x_iAnd z_jIt is same pedestrian target not With the feature representation under camera, otherwise y_ij=-1.Same pedestrian target sample is set up to (positive sample to) set S={ (x_i, z_j)|y_ij=1 }, while setting up set the D={ (x of negative sample pair_i, z_j)|y_ij=-1 }.Our target is that study produces one Metric space based on mahalanobis distance, range formula is as follows：

The wherein constraint of M is a positive semidefinite matrix, is initialized as unit matrix, and we come round and continue study later, obtain One matrix M for being best suitable for characteristic measure.

(2) iso-distance constraint condition is proposed, the characteristic distance between all negative samples pair is set to a constant constant μ, The distance of the feature of positive sample pair is set to 0.

Wherein parameter μ span is [2,4].This shows feature of the same target under different cameras in metric space On be collapsed into a point, without the characteristic distance between same pedestrian target while more than characteristic distance between same target, also need Ensure that characteristic distance of the characteristic distance between all different pedestrian targets in metric space is a constant value μ.

(3) in order to obtain matrix M by sample learning, we set up a loss function：

Wherein | S | and | D | be respectively positive sample pair and negative sample to number, γ is a parameter for belonging to [0,1], is represented The positive and negative influence to the weight for training.Loss function is further optimized, it has been found that inter- object distance it is secondary Convergence rate of the function near origin is very slow, and speed is even below linear function, and by quadratic function to one list of left Position, convergence rate just substantially quickening.The loss function in class is become and is turned to (d+1)²- 1, the final damage after constant term is removed Lose function as follows：

WhereinIt is that, in order to prevent metric matrix over-fitting and increased regularization term, λ spans are [5 × 10^-6, 1 × 10^-4].Finally our optimization problem is changed into following form：

min(L(M)).s.t.M≥0

(4) we select wide variety of Projected descent method to solve this problem, and gradient is calculated first：

Calculate for convenience, can be with equivalence to more succinct matrix form：

Wherein

G₂,G₁It is the element and arrangement diagonal matrix on the diagonal of the row and column of matrix G.In kth time iteration In, we are first by M_kDirectly decline by gradient direction, and be projected into positive semidefinite space, the matrix for obtainingIt is specific public Formula is：

In formula, []⁺It is the computational methods matrix projection to positive semidefinite matrix space P={ M | M >=0 }.Specific practice is, If we are needed non-positive semidefinite matrix C_kTo project to be equivalent in positive semidefinite space and one is found in the P of positive semidefinite matrix space The minimum matrix of the F norm squared poor with it：

Its conventional method is by C_kSingular value decompositionWhereinAnd Λ_kIt is by C_kSingular value is constituted Diagonal matrix, orderWhereinThenIt is that we to be found and C_kMost like half Positive definite matrix.

According to formula

α_kBe (0,1] a step-length, represent update weight, finally i.e. can obtain the measurement square after an iteration process Battle array M_k+1。

So far, we complete all steps of metric learning during an iteration, and repeat the above steps judgement loss Function changes, if being unsatisfactory for iteration termination condition(ε values are 1 × 10^-4,), then repeat step (4) after Continuous iteration.Otherwise, iteration is stopped, the M that current iteration is obtained is exactly our final optimization metric matrix M.

Fig. 3 is distributed schematic diagram directly perceived, spy of the same target under different cameras in figure for the characteristic vector of positive and negative samples The symbol for taking over same shape for use is represented；Wherein Fig. 3 (a) directly extracts the distributed effect schematic diagram of feature, and Fig. 3 (b) is equidistant The feature distribution effect diagram of the metric space produced under constrained procedure；As can be seen that by perform the present invention in etc. The metric learning method of distance restraint, the feature aggregation extent of feature positive sample pair in the metric space of M of test sample is more Intensive, negative sample more disperses to the distance between feature, meets our ideal effect, is conducive to weight accuracy of identification to improve.

The advantage of the algorithm is simple constraints, and all using steady state value, optimization method uses Projected descent method, Convergence rate is very fast.The actual purpose of the way of iso-distance constraint is that of avoiding the generation of algorithm over-fitting so that image is carried on the back The interference informations such as scape, noise are weakened in metric space, and weight accuracy of identification is improved, and is more suitable under the more complex scene of noise Pedestrian recognize again.

As it will be easily appreciated by one skilled in the art that the foregoing is only presently preferred embodiments of the present invention, it is not used to The limitation present invention, all any modification, equivalent and improvement made within the spirit and principles in the present invention etc., all should include Within protection scope of the present invention.

Claims (2)

1. a kind of metric learning method that pedestrian recognizes again, it is characterised in that comprise the following steps：

(1) goal set under two different cameras is set up, the element respectively constituted in set X and set Z, X, Z set is each Characteristic vector of the individual target in different camera hypographs；

(2) mahalanobis distance measure is used, the distance between X, Z set any two element is calculated： In formula, x_iIt is the target i characteristic vectors of set X, z_jIt is the characteristic vector of the target j of set Z；Positive sample is set up to set S, in S Element for sample to (x_i,z_j), x_i、z_jBelong to the characteristic vector of same pedestrian target under different cameras；Set up simultaneously negative Sample is sample to (x to set D, the element in D_i,z_j), x_i、z_jBelong to the mark sheet of different pedestrian targets under different cameras Reach；Metric matrix M is initialized as unit matrix；

(3) sample is 0 to the distance restraint of feature during all S are gathered, and the distance restraint of the characteristic vector in set D is one Constant constant μ, span is [2,4], sets up loss function L (M)：

$L\left(M\right)=\frac{\mathrm{\γ}}{\left|S\right|}\underset{(x_i,z_j)\∈S}{\Σ}{\left(D_M^2\right(x_i,z_j)+1)}^2+\frac{1-\mathrm{\γ}}{\left|D\right|}\underset{(x_i,z_j)\∈D}{\Σ}{\left(D_M^2\right(x_i,z_j)-\mathrm{\μ})}^2+\frac{\mathrm{\λ}}{2}\left|\right|M|{|}_F^2,$

In formula, | S | represents the element number of set S, i.e. to number, | D | represents element number in set D to positive sample, that is, bear sample This to number；γ is positive and negative samples to the weight for loss function, and interval is [0.5,0.7]；It is F norms Square, be regularization term, for avoiding algorithm over-fitting；λ spans are [5 × 10^-6, 1 × 10^-4], with making adjustments canonical The intensity of change；

(4) Projected descent method, iteration is used to seek the optimal value M of matrix M loss function L (M)^*, make loss function L (M^*) It is minimum；

In+1 iteration of kth, to kth time iteration optimal value M_kOptimized by gradient direction, obtain C_k=M_k-s_k▽L(M_k), s_kIt is step-size factor, value [0.05,0.5]；

(5) discrimination matrix C_kWhether positive semidefinite, be to make(6) are gone to step, otherwise by C_kIn projecting to positive semidefinite space With C_kMost like matrix, specific method is found and C in positive semidefinite space_kThe positive semidefinite matrix of two Norm minimums of differenceInstead of C_k；

(6) willWith the M before iteration_kWeight α and (1- α) are assigned respectively, and new M is formed after linear combination_k+1, as kth+1 time The result of the Metzler matrix of optimization, α spans for (0,1]；

(7) every time after iteration the absolute value of the difference of counting loss function and a preceding loss function ratio, if less than a certain Default threshold epsilon, then judge that iteration terminates, and obtains the optimal result of M；(4) are otherwise gone to step to continue to iterate to calculate；The ε takes Value is according to computational accuracy and calculates balance determination, can interval [1 × 10^-6, 1 × 10^-4]。

2. metric learning method according to claim 1, it is characterised in that in the step (5), finds most like square Battle array method is as follows：

To C_kSingular value decomposition is carried out, is obtainedWhereinI is unit matrix, Λ_kIt is by C_kIt is unusual It is worth the diagonal matrix of composition；

According to C_kObtain most like matrix Wherein

According toAnd M_k, obtain M_k+1, calculating formula is：

Obtain most like positive semidefinite matrix；

In formula, α_kValue (0,1] a step-length, representAnd M_kThe weight of difference.