CN106682087A - Method for retrieving vehicles on basis of sparse codes of features of vehicular ornaments - Google Patents

Abstract

The invention discloses a method for retrieving vehicles on the basis of sparse codes of features of vehicular ornaments. The method includes acquiring vehicle image data and detecting vehicle areas in images by the aid of gray co-occurrence matrix features and support vector machines; acquiring vehicular ornament area images in the vehicle images; creating over-complete dictionaries of the vehicular ornament area images; solving sparse vectors of the vehicular ornament area images corresponding to the to-be-inquired vehicles; reconstructing to-be-inquired vehicular ornament area images and computing Euclidean distances from reconstructed images to the to-be-inquired vehicular ornament area images; updating the sparse vectors of the to-be-inquired vehicular ornament area images, sorting elements of the sparse vectors according to the magnitude and setting thresholds to obtain retrieval results of the vehicular ornament area images. Compared with the prior art, the method has the advantages that the vehicles are retrieved on the basis of the features of the vehicular ornaments, accordingly, the retrieval precision and reliability can be greatly improved, and important bases can be provided for effectively solving problems in the aspect of illegal vehicle tracking.

Description

Vehicle retrieval method based on vehicle-mounted decoration feature sparse coding

Technical Field

The invention relates to the field of intelligent traffic research, in particular to research on a vehicle image retrieval method.

Background

In order to solve the urban development problem and realize the sustainable development of the city, the construction of a 'smart city' becomes the irreversible historical trend of the urban development in China; meanwhile, in order to meet the requirements of urban public security prevention and control and urban management, the government and law committee initiates the joint ministry of public security and the industry and the credit to jointly construct a skynet project. The skynet project is to utilize equipment and control software such as image acquisition, transmission, control, display and the like to carry out real-time monitoring and information recording on a fixed area, and provides technical support for extracting evidences of illegal criminal activities of vehicles in cities. However, the main problem in the process of evidence collection is to search for illegal criminal vehicles by adopting a manual searching method in the currently stored mass monitoring data, and the searching process not only consumes a large amount of manpower, material resources and financial resources, but also has great unreliability in manual searching. At present, the automatic search of suspicious vehicle information from video or picture data is mainly based on the automatic identification of the number plate number, the brand and the color of the vehicle inherent to the vehicle. However, in the case of illegal cases involving vehicles in real life, the case-involved vehicles are often fake (fake) branded vehicles, and the investigation according to the vehicle number cannot play any role; meanwhile, retrieving suspicious vehicles according to the brand and color of the vehicle plays little role in reducing workload. Therefore, the search method based on the vehicle intrinsic attribute features does not achieve the expected search effect on fake (fake) brand vehicles. An effective solution to this problem is to search for the suspect vehicle by using the features of the vehicle-mounted ornaments of the suspect vehicle, such as the vehicle ornaments and the annual inspection labels.

Disclosure of Invention

The invention aims to effectively solve the problem of vehicle image retrieval. In order to solve the technical problems, the invention adopts the following technical scheme:

a vehicle retrieval method based on vehicle-mounted decoration feature sparse coding comprises the following steps:

1) acquiring vehicle image data, and detecting a vehicle region in an image by adopting a gray level co-occurrence matrix characteristic and a support vector machine;

2) positioning the vehicle-mounted decoration area image in the vehicle image obtained in the step 1) according to the relative position of the vehicle front windshield and the vehicle;

3) constructing an over-complete dictionary of the vehicle-mounted ornament region image by adopting a K-singular value decomposition method;

4) solving sparse vectors of vehicle-mounted ornament area images corresponding to the vehicle to be inquired by adopting a sparsity self-adaptive matching tracking algorithm;

5) reconstructing a vehicle-mounted ornament region image to be inquired according to the overcomplete dictionary of the vehicle-mounted ornament region image obtained in the step 3) and the sparse vector obtained in the step 4), and calculating the Euclidean distance between the reconstructed image and the vehicle-mounted ornament region image to be inquired;

6) updating the sparse vector of the vehicle-mounted ornament region image to be inquired according to the Euclidean distance obtained in the step 5), sequencing elements of the sparse vector according to the size, setting a threshold value, obtaining a retrieval result of the vehicle-mounted ornament region image, and obtaining the vehicle to be inquired in the image database.

The vehicle image data obtained in the step 1) is obtained by a 300w pixel bayonet camera installed at an urban road intersection in a time period of 8: 00-18: 00 per day; the image data set contains different weather, illumination, vehicle angle and other conditions, each picture contains a complete vehicle image, and the original image size is 2048 × 1536 pixels.

The method for detecting the vehicle region in the image by adopting the gray level co-occurrence matrix characteristics and the support vector machine in the step 1) comprises the following specific steps:

11) selecting a picture only containing a vehicle positive sample and a picture only containing a background negative sample, and training to obtain a detected feature library;

12) carrying out graying processing on the vehicle image, and gridding the image into M multiplied by N sub-images with the size of W multiplied by H; calculating the gray level co-occurrence matrix characteristics of each grid window image, predicting whether the grid image belongs to a vehicle part by using the trained characteristic library in the step (1), and counting the coordinates of the grid centers of all identified vehicle areas to obtain the mean value of the grid center coordinates;

13) calculating Euclidean distance according to the grid center coordinates and the coordinate mean value of each identified vehicle area, and setting a certain threshold value; if the calculated distance is larger than the set average value, the calculated distance is determined as background noise; and obtaining the detected vehicle image according to the mean value of the coordinates.

In the step 2), the vehicle-mounted decoration is positioned in a front windshield area of the vehicle, namely the front windshield area of the vehicle is an interested area; and positioning the vehicle front windshield area according to the relative position of the vehicle front windshield area and the whole vehicle.

The vehicle front windshield area positioning specifically comprises:

21) finally obtaining an image of a roughly positioned vehicle front windshield area according to the relative position relation; and preprocessing the obtained image of the front windshield area of the vehicle: the preprocessing comprises graying, edge detection by using a Canny operator and projection calculation in the transverse and longitudinal directions of the image;

22) finding the left, right, upper and lower boundaries of a front windshield area according to the statistical transverse and longitudinal projection histograms, performing straight line fitting on an edge image of the vehicle front windshield by adopting Hough transformation, calculating the rotation angle of the image, and rotating the original image of the vehicle front windshield area to obtain a horizontal image of the vehicle front windshield area.

The step 3) of constructing the overcomplete dictionary of the vehicle-mounted ornament region image by adopting a K-singular value decomposition method specifically comprises the following steps:

31) initializing a dictionary D by using a training sample set, and solving an optimal coefficient vector of an input sample by using an orthogonal matching pursuit algorithm;

assuming that all sample sets of the dictionary to be trained are represented by Y, solving the dictionary D as an unknown quantity, x represents a sparse vector to be solved, and T₀Representing sparsity constraint parameters, the dictionary learning can be converted into the following formula to solve, namely

32) Updating the dictionary, keeping the other columns of the dictionary D unchanged, and updating the new column D_kNew column d_kBy singular value decomposition, and the coefficients of the new columnAssuming that sparse vector X and dictionary D are fixed, for a certain column D of the dictionary_kCoefficients corresponding to the column, penalty term for the target signal being

Wherein, the matrix E_kIndicates to remove d_kError of all samples after vector; definition of ω_kColumn d for successfully storing dictionary updates_kOf (2), i.e. coefficient vectorsIs not zero, i.e.

Define the matrix omega_k＝N×|ω_kI, orderThe penalty term for the target signal is expressed as

At this time, directly face toPerforming singular value decomposition to obtain:wherein the first column of U isV multiplied by Δ (1,1) is

33) And step 32) is repeated, and iteration is carried out until the maximum iteration times defined by initialization is reached, so that the final dictionary is obtained.

The process of solving the sparse vector of the vehicle-mounted ornament region image to be inquired by adopting the sparsity self-adaptive matching tracking algorithm in the step 4) is specifically as follows:

in the algorithm, L represents the initialization sparsity, r_tRepresenting the signal residual, t the number of iterations, Λ_tSet of indices representing t iterations, α_jColumn j, A, representing dictionary A_tRepresenting a collection Λ of pass indexes_tA set of columns selected from dictionary A, the number of columns being L_t，θ_tIs L_t× 1 column vectors, the input of the algorithm is signal y, the size of step S;

a) initialization signalResidual r₀＝y，Λ₀Is empty set, L is S, t is 1;

b) let h be equal to<r_t-1,A>And calculating the first L maximum value elements in the vector h, and forming a set S by the column sequence numbers of the values in the dictionary A_k；

c) Let C_k＝Λ_t-1∪S_k，A_t＝{α_jJ is the set C_kAll subscript sets in;

d) finding y as A_tθ_tIs least squares solution of, i.e.

e) FromThe L term with the maximum absolute value obtained by the calculation is recorded asCorresponding to A_tL in (1) is listed as A_tLL in corresponding A is listed as Λ_tL；

f) Residual update

g) Such as r_newGo to the next step if 0 r_new||₂≥||r_t-1||₂Updating the step length to L + S and returning to b) to continue iteration;

h) to obtain a final least squares solutionAt Λ_tLWith a non-zero term having a value obtained in the last iterationNamely the obtained sparse vector.

In the step 5), the established overcomplete dictionary and the established sparse vector are divided into k types, namely

A＝[A₁,A₂,…,A_j,…,…,A_n]＝[D₁,D₂,…,D_k]

Reconstruction of image C with segmented elements_k＝D_kx₀。

In the step 6), the reconstructed image C is calculated according to the Euclidean distance calculation formula_kWith the image S to be queried_kDistance between, two M × N-dimensional images C_kAnd S_kIs calculated as follows

Wherein,andrespectively represent images C_kAnd an image S_kThe d-th pixel point of (1).

From the calculated reconstructed image C_kUpdate sparse vectors with Euclidean distance from the image to be queried, i.e.

And sequencing according to the updated sparse vectors, and setting a threshold value to obtain a retrieval result.

Firstly, selecting a picture only containing a vehicle positive sample and a picture only containing a background negative sample to carry out gray level co-occurrence matrix characteristic extraction, training to obtain a detected characteristic library, secondly, quantizing the gray level of the image, wherein the quantized level is divided into 16 levels, and the dimensionality of the gray level co-occurrence matrix is 16 × 16 dimensions because of the relative relationship of pixel pairsThe method comprises the steps of calculating the probability distribution of relative position deviation of a vehicle image, calculating a gray level co-occurrence matrix according to the gray level value, performing gray level processing on the input image, gridding the image into M × N sub-images with the size of W × H, calculating the characteristic of the gray level co-occurrence matrix of each grid window image, predicting whether the grid image belongs to a vehicle part or not by using a trained characteristic library, counting the coordinates of the grid centers of all identified vehicle areas to obtain the mean value of the coordinates of the grid centers, calculating the Euclidean distance according to the coordinates of the grid centers and the mean value of the coordinates of each identified vehicle area, setting a certain threshold, determining the vehicle image as background noise if the calculated distance is larger than the set mean value, and finally obtaining the detected vehicle image according to the mean value of the coordinates.

Compared with the prior art, the technical scheme carries out vehicle retrieval based on the characteristics of the vehicle-mounted ornaments, greatly increases the retrieval precision and reliability, and provides important basis for effectively solving the problem of illegal vehicle tracking.

Drawings

FIG. 1 is a process for vehicle front windshield ROI localization according to the present invention.

Detailed Description

The technical solution is further explained below with reference to specific embodiments:

the method comprises the steps of firstly training a feature library, selecting a picture only containing a vehicle positive sample and a picture only containing a background negative sample to perform gray level co-occurrence matrix feature extraction, training to obtain a detected feature library, quantizing the gray level of an image, wherein the quantized grade is divided into 16 grades, so that the dimension of a gray level co-occurrence matrix is 16 × 16 dimensions, and selectingFirstly, carrying out graying processing on an input image, gridding the image into M × N subgraphs with the size of W × H, then, calculating the gray level co-occurrence matrix characteristic of each grid window image, utilizing the trained characteristic library to predict whether the grid image belongs to a vehicle part, carrying out statistics on the coordinates of the grid centers of all the identified vehicle areas to obtain the mean value of the grid center coordinates, then, carrying out Euclidean distance calculation according to the grid center coordinates and the mean value of each identified vehicle area, setting a certain threshold value, if the calculated distance is greater than the set mean value, determining the grid area as background noise, and finally, obtaining the detected vehicle image according to the mean value of the coordinates.

The second step is that: and positioning the vehicle-mounted decoration according to the relative position of the vehicle front windshield and the whole vehicle. Vehicle ornaments such as vehicle ornaments, labels are mainly located the front windshield region of vehicle, so need carry out the regional location of vehicle front windshield after detecting the vehicle, the location of the region of interest (ROI) in this patent promptly, the vehicle front windshield location need be according to its and the holistic relative position of vehicle and fix a position. For example, as shown in fig. 1(a), for the detected vehicle picture, the start coordinate is located at the lower left corner of the whole picture, the Width of the whole picture is Width, the Height of the whole picture is assumed to be Height, the start position of the front windshield of the vehicle is 0.4 × Height from the bottom end of the picture, the Height of the whole region of interest is 0.35 × Height, and the Width is 0.85 × Width. The roughly positioned front windshield portion finally obtained from the relative positional relationship thereof is shown in fig. 1 (b). After the rough positioning of the front windshield area is obtained, it will be precisely positioned. A series of pre-processing of the image is required before accurate positioning: 1) carrying out gray processing on the front windshield part; 2) carrying out edge detection on the image by applying a Canny operator; 3) the projection in the lateral and longitudinal directions of the image is calculated to find the left and right and upper and lower boundaries of the front windshield portion. Since there is a sudden change in the projection histograms in which there are distinct edges at the left and right and upper and lower boundaries of the front windshield, the left and right and upper and lower boundaries of the front windshield portion can be clearly found from the statistical lateral and longitudinal projection histograms. However, in some vehicle front windshield images, the front windshield portion found by coordinate positioning is rotated by a certain angle, so that a certain correction is required after obtaining a partial image of the front windshield, thereby restoring the front windshield portion to a horizontal position. The method comprises the steps of calculating the inclination angle of the vehicle windshield part when correcting the image, adopting Hough transformation to perform linear fitting on the edge image of the front windshield, calculating the rotation angle of the image, and then rotating the original vehicle front windshield image to obtain a horizontal front windshield image.

The third step: constructing a dictionary of sparse representations is an important prerequisite for sparse coding implementations. This patent adopts K Singular Value Decomposition (KSVD) as the dictionary updating algorithm. The main processes of the KSVD algorithm are as follows:

a) initializing a dictionary D by using a training sample set;

b) solving a sparse coefficient X by using an orthogonal matching pursuit algorithm based on an over-complete dictionary D;

c) define the matrix omega_kDecomposing the penalty term;

d) to pairSingular value decomposition is carried out, and a super-complete dictionary and a sparse coefficient are updated;

e) and repeating the steps b) to d), and iterating to the maximum iteration times defined by initialization to obtain the final dictionary.

The dictionary represented sparsely is obtained by training known samples, the dictionary A is assumed to be formed by preferentially arranging sample images according to rows or columns and is known, in the solving process of sparse vectors, the images need to be expanded into one-dimensional vectors according to the rows or columns, the dimension of expanded image signals is large, the calculation amount in the solving process is large, and therefore a dictionary learning method needs to be selected in the first step of sparse coding. The dictionary learning algorithm selected by the patent is a commonly used K Singular Value Decomposition (KSVD) algorithm.

The main processes of the KSVD algorithm are as follows:

(1) initializing a dictionary D by using a training sample set, and solving an optimal coefficient vector of an input sample by using an orthogonal matching pursuit algorithm;

assuming that all sample sets of the dictionary to be trained are represented by Y, now the solution is performed for dictionary D as an unknown quantity, x represents the sparse vector to be solved, T₀Representing sparsity constraint parameters, the dictionary learning can be converted into the following formula to solve, namely

(2) Updating the dictionary, keeping the other columns of the dictionary D unchanged, and updating the new column D_kNew column d_kBy singular value decomposition, and the coefficients of the new columnAssuming that the sparse vector X and the dictionary D are fixed, the object of interest at this time is a certain column D of the dictionary_kThe coefficient corresponding to the column, the penalty term of the target signal can be expressed as

Wherein, the matrix E_kIndicates to remove d_kError of all samples after the vector. However, the coefficient obtained at this timeIf the dictionary is not sparse, the dictionary is not updated, so that the definition omega is not satisfied_kColumn d for successfully storing dictionary updates_kI.e. coefficient vectorsIs not zero, i.e.

Define the matrix omega_k＝N×|ω_kI, orderThen the formula (2) can be represented as

At this time, directly face toPerforming singular value decomposition to obtain:wherein the first column of U isV multiplied by Δ (1,1) is

(3) And (5) repeating the step (2), and iterating to the maximum iteration times defined by initialization to obtain the final dictionary.

The fourth step: the self-adaptive matching tracking algorithm based on the sparsity does not need the prerequisite of the sparsity, but estimates and updates the sparsity in the algorithm processing process, initializes the sparsity to be L at the beginning of the algorithm, and continuously estimates the sparsity and a source image so as to achieve the purpose of reconstructing the image.

In the algorithm, r_tRepresenting the signal residual, t the number of iterations, Λ_tSet of indices representing t iterations, α_jColumn j, A, representing dictionary A_tRepresenting a collection Λ of pass indexes_tA set of columns selected from dictionary A, the number of columns being L_t，θ_tIs L_t× 1 column vector, the input of the algorithm is signal y, the built dictionary A, the step size S, the process of the sparsity-based adaptive matching pursuit algorithm is as follows:

a) initialization residual r₀＝y，Λ₀Is empty set, L is S, t is 1;

b) let h be equal to<r_t-1,A>And calculating L maximum values in h, and forming a set S by using the column sequence numbers of the values in the dictionary A_k；

c) Let C_k＝Λ_t-1∪S_k，A_t＝{α_j(wherein j is C)_kAll subscript sets in);

d) finding y as A_tθ_tIs least squares solution of, i.e.

e) FromThe L term with the maximum absolute value obtained by the calculation is recorded asCorresponding to A_tL in (1) is listed as A_tLL in corresponding A is listed as Λ_tL；

f) Residual update

g) Such as r_newGo to the next step if 0 r_new||₂≥||r_t-1||₂Updating the step length to L + S and returning to b) to continue iteration;

h) to obtain a final least squares solutionAt Λ_tLWith a non-zero term having a value obtained in the last iterationNamely the obtained sparse vector.

The fifth step: the built overcomplete dictionary and sparse vectors are classified into k classes, i.e.

A＝[A₁,A₂,…,A_j,…,…,A_n]＝[D₁,D₂,…,D_k](6)

Reconstruction of image C with segmented elements_k＝D_kx₀。

And a sixth step: calculating a reconstructed image C according to a Euclidean distance calculation formula_kWith the image S to be queried_kDistance between, two M × N-dimensional images C_kAnd S_kIs calculated as follows

Wherein,andrespectively represent images C_kAnd an image S_kThe d-th pixel point

From the calculated reconstructed image C_kUpdate sparse vectors with Euclidean distance from the image to be queried, i.e.

And sequencing according to the updated sparse vectors, and setting a threshold value to obtain a retrieval result.

The above description is only an embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes performed by the present specification and drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims (9)

1. A vehicle retrieval method based on vehicle-mounted decoration feature sparse coding is characterized by comprising the following steps: the method comprises the following steps:

1) acquiring vehicle image data, and detecting a vehicle region in an image by adopting a gray level co-occurrence matrix characteristic and a support vector machine;

2) positioning the vehicle-mounted decoration area image in the vehicle image obtained in the step 1) according to the relative position of the vehicle front windshield and the vehicle;

3) constructing an over-complete dictionary of the vehicle-mounted ornament region image by adopting a K-singular value decomposition method;

4) solving sparse vectors of vehicle-mounted ornament area images corresponding to the vehicle to be inquired by adopting a sparsity self-adaptive matching tracking algorithm;

5) reconstructing a vehicle-mounted ornament region image to be inquired according to the overcomplete dictionary of the vehicle-mounted ornament region image obtained in the step 3) and the sparse vector obtained in the step 4), and calculating the Euclidean distance between the reconstructed image and the vehicle-mounted ornament region image to be inquired;

6) updating the sparse vector of the vehicle-mounted ornament region image to be inquired according to the Euclidean distance obtained in the step 5), sequencing elements of the sparse vector according to the size, setting a threshold value, obtaining a retrieval result of the vehicle-mounted ornament region image, and obtaining the vehicle to be inquired in the image database.

2. The vehicle retrieval method according to claim 1, characterized in that: the vehicle image data obtained in the step 1) is obtained by a 300w pixel bayonet camera installed at an urban road intersection in a time period of 8: 00-18: 00 per day; the image data set contains different weather, lighting and vehicle angle conditions, and each picture contains a complete vehicle image, with the original image size being 2048 × 1536 pixels.

3. The vehicle retrieval method according to claim 1, characterized in that: the method for detecting the vehicle region in the image by adopting the gray level co-occurrence matrix characteristics and the support vector machine in the step 1) comprises the following specific steps:

11) selecting a picture only containing a vehicle positive sample and a picture only containing a background negative sample, and training to obtain a detected feature library;

12) carrying out graying processing on the vehicle image, and gridding the image into M multiplied by N sub-images with the size of W multiplied by H; calculating the gray level co-occurrence matrix characteristics of each grid window image, predicting whether the grid image belongs to a vehicle part by using the trained characteristic library in the step (1), and counting the coordinates of the grid centers of all identified vehicle areas to obtain the mean value of the grid center coordinates;

13) calculating Euclidean distance according to the grid center coordinates and the coordinate mean value of each identified vehicle area, and setting a certain threshold value; if the calculated distance is larger than the set average value, the calculated distance is determined as background noise; and obtaining the detected vehicle image according to the mean value of the coordinates.

4. The vehicle retrieval method according to claim 1, characterized in that: in the step 2), the vehicle-mounted decoration is positioned in a front windshield area of the vehicle, namely the front windshield area of the vehicle is an interested area; and positioning the vehicle front windshield area according to the relative position of the vehicle front windshield area and the whole vehicle.

5. The vehicle retrieval method according to claim 4, characterized in that: the vehicle front windshield area positioning specifically comprises:

21) finally obtaining an image of a roughly positioned vehicle front windshield area according to the relative position relation; and preprocessing the obtained image of the front windshield area of the vehicle: the preprocessing comprises graying, edge detection by using a Canny operator and projection calculation in the transverse and longitudinal directions of the image;

22) finding the left, right, upper and lower boundaries of a front windshield area according to the statistical transverse and longitudinal projection histograms, performing straight line fitting on an edge image of the vehicle front windshield by adopting Hough transformation, calculating the rotation angle of the image, and rotating the original image of the vehicle front windshield area to obtain a horizontal image of the vehicle front windshield area.

6. The vehicle retrieval method according to claim 1, characterized in that: the step 3) of constructing the overcomplete dictionary of the vehicle-mounted ornament region image by adopting a K-singular value decomposition method specifically comprises the following steps:

31) initializing a dictionary D by using a training sample set, and solving an optimal coefficient vector of an input sample by using an orthogonal matching pursuit algorithm;

assuming that all sample sets of the dictionary to be trained are represented by Y, the dictionary D is solved as an unknown quantity,x represents the sparse vector to be solved, T₀Representing sparsity constraint parameters, the dictionary learning can be converted into the following formula to solve, namely

$\begin{array}{ccc}\underset{D,X}{\min }\left\{\right||Y-DX|{|}_F^2\}& subjectto& \&ForAll;i,\left|\right|x_i|{|}_0\&le;T_0\end{array}$

32) Updating the dictionary, keeping the other columns of the dictionary D unchanged, and updating the new column D_kNew column d_kBy singular value decomposition, and the coefficients of the new columnAssuming that sparse vector X and dictionary D are fixed, for a certain column D of the dictionary_kCoefficients corresponding to the column, penalty term for the target signal being

$\left|\right|Y-DX|{|}_F^2=||Y-\underset{j=1}{\overset{K}{\&Sigma;}}{d}_j{x}_T^j|{|}_F^2=\left|\right|Y-\underset{j\&NotEqual;k}{\&Sigma;}{d}_j{x}_T^j-d_k{x}_T^k|{|}_F^2=||E_k-d_k{x}_T^k|{|}_F^2$

Wherein, the matrix E_kIndicates to remove d_kError of all samples after vector; definition of ω_kColumn d for successfully storing dictionary updates_kOf (2), i.e. coefficient vectorsIs not zero, i.e.

${\mathrm{\&omega;}}_k=\left\{i\right|1\&le;i\&le;K,x_T^k\left(i\right)\&NotEqual;0\}$

Define the matrix omega_k＝N×|ω_kI, orderThe penalty term for the target signal is expressed as

$\left|\right|E_k{\mathrm{\&Omega;}}_k-d_k{x}_T^k{\mathrm{\&Omega;}}_k|{|}_2^2=||E_R^k-d_k{x}_R^k|{|}_2^2$

At this time, directly face toPerforming singular value decomposition to obtain:wherein the first column of U isV multiplied by Δ (1,1) is

33) And step 32) is repeated, and iteration is carried out until the maximum iteration times defined by initialization is reached, so that the final dictionary is obtained.

7. The vehicle retrieval method according to claim 1, characterized in that: the process of solving the sparse vector of the vehicle-mounted ornament region image to be inquired by adopting the sparsity self-adaptive matching tracking algorithm in the step 4) is specifically as follows:

in the algorithm, L represents the initialization sparsity, r_tRepresenting the signal residual, t the number of iterations, Λ_tSet of indices representing t iterations, α_jColumn j, A, representing dictionary A_tRepresenting a collection Λ of pass indexes_tA set of columns selected from dictionary A, the number of columns being L_t，θ_tIs L_t× 1 column vectors, the input of the algorithm is signal y, the size of step S;

a) initialization signal residual r₀＝y，Λ₀Is empty set, L is S, t is 1;

b) let h be < r_t-1A >, and calculates the first L maximum elements in the vector h, and sets S with these values in the column number of dictionary A_k；

c) Let C_k＝Λ_t-1∪S_k，A_t＝{α_jJ is the set C_kAll subscript sets in;

d) finding y as A_tθ_tIs least squares solution of, i.e.

e) FromThe L term with the maximum absolute value obtained by the calculation is recorded asCorresponding to A_tL in (1) is listed as A_tLL in corresponding A is listed as Λ_tL；

f) Residual update

g) Such as r_newGo to the next step if 0 r_new||₂≥||r_t-1||₂Updating the step length to L + S and returning to b) to continue iteration;

h) to obtain a final least squares solutionAt Λ_tLWith a non-zero term having a value obtained in the last iterationNamely the obtained sparse vector.

8. The vehicle retrieval method according to claim 1, characterized in that: in the step 5), the established overcomplete dictionary and the established sparse vector are divided into k types, namely

A＝[A₁,A₂,…,A_j,…,…,A_n]＝[D₁,D₂,…,D_k]

Reconstruction of image C with segmented elements_k＝D_kx₀。

9. The vehicle retrieval method according to claim 1, characterized in that: in the step 6), the reconstructed image C is calculated according to the Euclidean distance calculation formula_kWith the image S to be queried_kDistance between, two M × N-dimensional images C_kAnd S_kIs calculated as follows

$ED(C_k,S_k)={\&lsqb;\underset{d=1}{\overset{MN}{\&Sigma;}}{(C_k^d-S_k^d)}^2\&rsqb;}^{1/2}$

Wherein,andrespectively represent images C_kAnd figuresImage S_kThe d-th pixel point of (1).

From the calculated reconstructed image C_kUpdate sparse vectors with Euclidean distance from the image to be queried, i.e.

${x_0}^{\&prime;}=\&lsqb;\frac{\left|x_{0,1}\right|}{{\mathrm{ED}}_1};\frac{\left|x_{0,2}\right|}{{\mathrm{ED}}_2};\mathrm{...};\frac{\left|x_{0,k}\right|\&rsqb;}{{\mathrm{ED}}_k}$

And sequencing according to the updated sparse vectors, and setting a threshold value to obtain a retrieval result.