WO2021046681A1 - Complex scenario-oriented multi-source target tracking method - Google Patents

Abstract

A complex scenario-oriented multi-source target tracking method, which relates to the field of artificial intelligence technology, and comprises the following steps: forming an initial deep learning network, setting a weight of the initial deep learning network, and performing steepest gradient dimensionality reduction on the weight of the initial deep learning network, so as to obtain an initial feature matrix of a multi-source target; performing simplified sparse representation and sparse processing on the initial feature matrix of the multi-source target, so as to obtain a sparsified target feature matrix; and classifying the sparsified target feature matrix, and establishing and forming feature matrix representations of different feature targets. In said method, a multi-source target feature matrix library established by means of a deep learning network has a good modeling characteristic and a data statistical structure, and a multi-source target tracker model can be generated for label-free training data.

Description

Multi-source target tracking method for complex scenes Technical field

The invention relates to the technical field of data processing, in particular to a deep learning-oriented multi-source target tracking method for complex scenes.

Background technique

The multi-source target tracking method for complex scenes can provide accurate target data characteristics, and realize the tracking of multi-source targets by constructing a deep learning processing architecture based on feature matrix learning. Specifically, a template is formed by training feature matrices of a large number of typical simulation scenarios, and an autonomous, controllable and addable template model is obtained. This model also has the feature matrix enhancement function of the scene, which can realize the enhancement of special features such as shadow and texture of the scene.

As a signal transformation method, feature matrix learning can approximate high-dimensional spatial features with low-dimensional space vectors based on the basis vectors of the complete feature matrix. However, it is difficult to achieve the goal of reducing the dimensionality of the feature matrix through the feature matrix learning itself to update the feature matrix atoms, and a large amount of calculation will be formed when the feature selection of the image target is performed. At the same time, one of the core problems of feature matrix learning is sparse representation. In order to improve the sparse representation of the feature matrix learning, the design of the feature matrix usually needs to update the fixed feature matrix basis vector. Whether a complete feature matrix can be designed determines whether the real signal can be represented more closely. Therefore, there is currently a need for a tracking processing method that can reduce the dimension of the feature matrix and more closely represent the real target.

Summary of the invention

The purpose of the embodiments of the present invention is to provide a deep learning-based multi-source target tracking method for complex scenes, aiming to solve the technical problem that the prior art is difficult to reduce the dimensionality of the feature matrix and cannot more closely represent the real target.

To achieve the foregoing objective, the embodiments of the present invention provide the following technical solutions:

A multi-source target tracking method for complex scenes includes the following steps:

S1: Form an initial deep learning network for a multi-source target, set the weights of the initial deep learning network, and perform the steepest gradient dimensionality reduction process on the weights of the initial deep learning network to obtain a preliminary feature matrix library of the multi-source target ；

S2: Perform simplified sparse representation and sparse processing on the preliminary features of the multi-source target to obtain a sparse feature matrix;

S3: Perform multi-source target classification on the sparse feature matrix, and establish feature representations of different feature targets;

S4: Minimize the sparse feature matrix;

S5: Establish the generative model and reconstruction model of feature matrix learning;

S6: Form effective multi-source tracking of the target through the generation model and the reconstruction model.

Preferably, the step S1 specifically includes: establishing a deep learning network, forming the initial feature matrix from the features of the multi-source target through the deep learning network, and the initial feature matrix has the function of adaptive change and improvement. The expression of the initial feature matrix is:

Among them, R represents the multi-source target feature matrix, k represents the number of targets, R _k represents the feature of each target, s represents the weight matrix of the deep learning network, and Y _k represents the tracked multi-source target.

Preferably, the step S2 includes: performing an effective feature matrix update iteration through feature decomposition, and completing the generation of sparse features through the preprocessing of the feature matrix.

Further preferably, the step S2 includes: the optimization problem of formula (1) is transformed into the optimal value problem of formula (2), which is obtained by the following formula:

Where R represents the initial feature matrix of the multi-source target, and s represents the weight matrix of the deep learning network.

Preferably, the step S3 further includes the step of updating the feature matrix, wherein the feature matrix R of formula (1) is minimized to

The range of the characteristic matrix R is selected to correspond to the zero position of the k column of the matrix s.

Preferably, the step S4 includes: the image feature R _opt ={r _i |i=1, 2, K n} obtained through the step S1 is used as the input of the deep learning network, and d′ _{i is} used as the corresponding weight Structure features, calculate the reconstruction error and average error of the multi-source target,

And the feature reconstruction is expressed as:

T'={v _i |e _i <η, v _i εV} (5).

Preferably, the step S5 includes: calculating the reconstruction error by formula (7):

Preferably, the multi-source target feature data is one or more of image data, radar data, communication data and position data.

Preferably, in the step S5, the generative model of the feature matrix learning is a top-down generative model.

Preferably, in the step S5, the reconstruction model of the feature matrix learning is a bottom-up reconstruction model.

Compared with the prior art, the beneficial effects of the present invention are:

(1) A multi-source target tracking method oriented to complex scenes based on deep learning. This method first builds a multi-source complete feature matrix library of video image features of a large number of different tracking targets through feature matrix learning. Specifically, the feature matrix is compressed and learned, and non-zero columns of the original video are selected to correspond to the sparse feature matrix atoms to form a complete feature matrix library based on deep learning, which is suitable for deep learning hierarchical architecture. According to the MMSE criterion, a top-down generative model and a bottom-up reconstruction model are established. At the same time, a multi-layer feedforward network of discriminative training based on this criterion is adopted for feature reconstruction.

(2) The method of the present invention can be applied to target features of different complex environments, such as various feature models such as wilderness scenes, urban scenes, open space scenes, etc., to establish a sparse representation of feature matrix learning, and at the same time form a complete feature matrix. At the same time, the present invention establishes non-zero coefficients of constraint weights to obtain sparse representations closer to real different scenes.

(3) The present invention adopts a typical multi-layer deep learning network. The basic structure is a memory, storage and knowledge network, which is fully linked by multiple networks and serves as a feature transfer channel between layers. At the same time, each layer is also It is used to train the feature structure of the next composition. The deep learning network of the present invention has the ability to complete complex data modeling, including a top-down generation model and a bottom-up discrimination model. The deep learning network also has the data training performance of weakly supervised learning. Therefore, the multi-source target feature matrix library established by the present invention through the deep learning network has good modeling characteristics and data statistical structure, and can generate a multi-source target tracker model for unlabeled training data.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

Fig. 1 is a flowchart of a deep learning-oriented multi-source target tracking method for complex scenes according to an embodiment of the present invention.

detailed description

In order to make the technical problems, technical solutions, and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, but not to limit the present invention.

The embodiment of the present invention provides a complex scene-oriented multi-source target tracking method based on deep learning, which can be used in technical fields such as video recognition and tracking.

The method of the present invention first establishes a complete feature matrix library of multi-source video images of massive different tracking targets through feature matrix learning. The feature matrix learning of the present invention, as a signal transformation method, can approximate high-dimensional space features through low-dimensional space vectors based on the matrix basis vectors of the over-complete feature matrix. The embodiment of the present invention establishes a sparse representation of the feature matrix learning for various feature models of target features in different complex environments, such as a wilderness scene, an urban scene, and an open space scene, and at the same time forms a complete feature matrix. At the same time, the constraint weights established by the method of the present invention are non-zero coefficients to obtain a sparse representation that is closer to the real different scenes. Then complete the multi-source target feature matrix optimization approximation in different feature scenes.

Referring to Fig. 1, Fig. 1 is a flowchart of a deep learning-oriented multi-source target tracking method for complex scenes according to an embodiment of the present invention. The tracking method includes the following steps:

S1: Provide multi-source target data, such as video data, to form an initial deep learning network, set the weight of the initial deep learning network, reduce the weight of the initial deep learning network to the steepest gradient, and form the initial of the multi-source target Feature matrix

S2: Simplify the sparse representation of the features of the multi-source target, and sparse the multi-source target through the deep learning network to obtain the sparse feature matrix;

S3: Through the multi-source target classification of the sparse feature matrix, a feature matrix representation of different feature targets is established. The target system can include: target image features, target radar features, target communication features, and target location features;

S4: Minimize the sparse feature matrix, update the initial feature matrix to obtain the minimized sparse feature matrix, which is the output of the memory system knowledge network in the figure;

S5: Establish a top-down generative model and a bottom-up reconstruction model through the deep learning network, where the generative model belongs to the generative template model of feature matrix learning, and the reconstructed model belongs to the reconstruction template of feature matrix learning model;

S6: The effective multi-source target tracking of the target is formed by generating the template model and reconstructing the template model.

Specifically, the aforementioned multi-source target feature data may be one or more of image data, radar data, communication data, and position data.

Take the ground-to-air multi-source target tracking for complex environments as an example: a ground-to-air communication scenario is established, and the ground station achieves target tracking of 10 UAVs through different data link characteristics. Among them, the radar reflection cross-sections of different UAVs are different. Through different UAV radar reflection cross-sections and acquired image characteristics, combined with the multi-source characteristics of the communication carrier frequency, a feature matrix network is formed.

In this specific embodiment, step S2 includes: obtaining the UAV radar reflection cross-section of the multi-source target, and the acquired image features, combining the characteristics of the communication carrier frequency to sparse through the deep learning network to obtain a sparse target feature matrix ; Establish target mapping through feature matrix.

Specifically, step S5 includes: establishing a top-down generative model through the deep learning network. The generative model includes the UAV radar reflection cross-section, and the acquired image features, combined with the joint feature model of the communication carrier frequency, and the generative model The model belongs to the generation template model of feature matrix learning; through the deep learning network, bottom-up reconstruction model, the model belongs to the reconstruction template model of feature matrix learning.

Further, step S1 specifically includes: establishing a deep learning network, and forming the target features of the multi-source target into an initial feature matrix through the deep learning network. The initial feature matrix can be based on the feature requirements of the multi-source target, that is, the UAV radar reflection The cross-section, the characteristics of the acquired image and the characteristics of the communication carrier frequency, as well as the subsequent iterations of its own, form adaptive changes and improvements. The expression of the initial feature matrix is:

Among them, k represents the target number of multi-source targets, R _k represents each target feature, and s represents the weight matrix of the deep learning network. Y _k represents the tracked multi-source target.

In other embodiments, the initial feature matrix can be adaptively changed and improved according to the feature requirements of the multi-source target, such as texture, contour feature, etc., and subsequently based on the iterative initialization of the feature matrix.

Further, the specific steps in step S2 include: substituting formula (1) through the feature decomposition of the current multi-source target, that is, forming two stages of iterative update through the initial feature matrix. This step increases the amount of calculation by decomposing each node, and in the decomposition process, the solution of the problem is transformed into finding _{a sparse code that minimizes R k} s (the iterative update of the feature matrix). To address this problem, and to reduce the risk of selecting redundant feature matrices, the present invention simplifies the sparse representation step, which is determined by updating the matrix R column. If R is less than the determined threshold at this time, then the k rows of the feature matrix D Can be treated as a zero vector. The target x will be updated under the joint support of the feature matrix D and the matrix coefficient y.

Among them, the core idea of step S2 is to perform an effective feature matrix update iteration through the feature decomposition method, and complete the generation of sparse features through the preprocessing of the feature matrix.

One of the improvement points of the feature matrix learning of the present invention is sparse representation. In order to improve the sparse representation of feature matrix learning, it is usually necessary to update the fixed feature matrix basis vector when designing the feature matrix. Therefore, whether a complete feature matrix can be designed determines whether it can more closely represent the real signal. The basic structure of the multi-layer deep learning network used in the embodiment of the present invention is a memory, storage, and knowledge network. Specifically, multiple networks are fully linked, thereby acting as a feature transfer channel between layers. At the same time, each layer is also It is used to train the feature structure of the next composition. Deep learning networks have been applied in different fields and can include many different forms of data generation models at the same time. The deep learning network of the present invention can complete complex data modeling, including a top-down generation model and a bottom-up discrimination model. This shows that the deep learning network of the present invention establishes data training performance with weak supervision through a multi-source target matrix library. The multi-source target feature matrix library established by the invention through the deep learning network has good modeling characteristics and data statistical structure, and can generate a multi-source target tracker model for unlabeled training data.

Further, step S2 also includes: under the joint decision of the initial feature matrix R of the multi-source target and the weight matrix s of the deep learning network, the optimization problem of formula (1) can be equivalent to the optimization of formula (2) Value problem, as shown in the following formula:

It also includes the feature matrix of the most recent common sparse representation. The core point of solving this problem is to determine the hard threshold judgment of the column of s in order to retain the amplitude judgment threshold in each column. For example, in similarity analysis, a simple soft threshold function is used to solve formula (1). If the sparse constraint is convex and relaxed, it will be difficult to solve it by calculating a simpler soft threshold method. Therefore, the method of the present invention adopts a certain hard decision threshold to implement.

Further, the specific steps in step S4 include: a feature matrix update step, which minimizes the feature matrix R of formula (1).

Here, the update range of the feature matrix R is determined by column selection of the weight matrix s. The range of the matrix R is selected to correspond to the zero position of the k column of the weight matrix s. This step uses only limited prior information of s instead of the complete matrix, which reduces the calculation amount of feature matrix update, and effectively supports the learning of the feature matrix update step with limited calculation amount.

Further, the specific steps in step S4 include: through step S1, the obtained image feature R _opt ={r _i |i=1, 2, K n} is used as the input of the deep learning network, and d′ _{i is taken} as the corresponding Reconstruct the distance feature, calculate the reconstruction error of the multi-source target according to the MMSE (Minimum Mean Squared Error) criterion, and calculate the average error by the following formula

Among them, e _i represents the error value of each target.

In the feature learning iteration, the feature reconstruction can be expressed as:

T'={v _i |e _i <η, v _i εV} (5).

Wherein, v _i represents reconstructed feature multiple source object, V represents acquired target feature, η represents a set of threshold characteristics. Further, the step of "building a top-down generative model" in step S5 includes: a feature learning process terminated by the method of obtaining a parameter. The error between the average reconstruction error of the current iteration and the last one that will stop the iteration, and the iteration is completed by the error. In the iterative feature learning process, since the reconstruction weight matrix of the eigenvalues is more reliable, suppose it is the reconstruction weight matrix of M, so that the multi-source feature of I on the test data set,

X is the extracted multiple features, reconstructed features

Can be written as:

From the characteristic matrix, calculate the reconstruction error:

The method proposed in the present invention learns the UAV radar reflection cross section, the acquired image characteristics and the characteristics of the communication carrier frequency by compressing the feature matrix, and obtains the weakly supervised learning initialization weights of the deep learning model, and through the reconstruction and judgment between layers And feature selection, adjust the weight of the network, and finally obtain the simplified weight of the system to obtain the tracking of multi-source feature targets.

The deep learning network proposed by the present invention is a hierarchical architecture. First, the learning feature matrix is compressed by selecting non-zero columns, and then a preliminary feature matrix is formed. By establishing a hierarchical deep learning architecture, it is possible to effectively track multi-source targets in complex scenarios. The verification of scene data shows that the target tracking efficiency of the multi-source target tracking method of the present invention surpasses the one-time algorithm tracking of a single dimension. In the specific operation, for the low-altitude multi-sort UAV tracking scene, high-order modulation for QAM, the radar reflection cross section is set to 1, which can complete the multi-source target tracking efficiency of 99%.

The above descriptions are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modification, equivalent replacement and improvement made within the spirit and principle of the present invention shall be included in the protection of the present invention. Within range.

Claims (10)

1. A multi-source target tracking method for complex scenes is characterized in that it comprises the following steps:

   S1: For multi-source target feature data, an initial deep learning network is formed, the weights of the initial deep learning network are set, and the weights of the initial deep learning network are reduced by the steepest gradient, and a multi-source target is formed Initial feature matrix;

   S2: Simplify and sparse the initial feature matrix of the multi-source target and

   Sparse processing to obtain a sparse target feature matrix;

   S3: Classify the sparse target feature matrix, and establish feature matrix representations of different feature targets;

   S4: Minimize the sparse target feature matrix;

   S5: Establish the generative model and reconstruction model of feature matrix learning;

   S6: Form effective multi-source target tracking of the target through the generation model and the reconstruction model.
2. The multi-source target tracking method according to claim 1, wherein the step S1 comprises:

   A deep learning network is established, and the features of multi-source targets are formed into the initial feature matrix through the deep learning network. The initial feature matrix has the function of adaptive change and improvement. The expression of the initial feature matrix is:

   
3. The multi-source target tracking method according to claim 1, wherein the step S2 comprises:

   The feature decomposition method is used to perform effective feature matrix update iterations, and the feature matrix preprocessing is used to complete the generation of sparse features.
4. The multi-source target tracking method according to claim 2, wherein the step S2 comprises: the optimization problem of formula (1) is transformed into the optimal value problem of formula (2), which is obtained by the following formula:

   

   Where R represents the initial feature matrix of the multi-source target, and s represents the weight matrix of the deep learning network.
5. The multi-source target tracking method according to claim 2, wherein the step S3 further includes the step of updating the feature matrix, wherein the feature matrix R of formula (1) is minimized to

   

   The range of the characteristic matrix R is selected to correspond to the zero position of the k column of the matrix s.
6. The multi-source target tracking method according to claim 1, wherein the step S4 comprises: the image feature R _opt ={r _i |i=1, 2, K n} obtained through the step S1 as the depth Learn the input of the network, use d' _i as the corresponding reconstruction feature, calculate the reconstruction error and average error of the multi-source target,

   And the feature reconstruction is expressed as:

   T'={v _i |e _i ＜η,v _i ∈V} (5),

   Where v _i represents reconstructed feature multiple source object, V represents acquired target feature, η represents a set of threshold characteristics.
7. The multi-source target tracking method according to claim 1, wherein the step S5 comprises: calculating the reconstruction error by formula (7):

   
8. The multi-source target tracking method according to claim 1, wherein the multi-source target feature data is one or more of image data, radar data, communication data, and position data.
9. The multi-source target tracking method according to claim 1, wherein in step S5, the generative model is a top-down generative model.
10. The multi-source target tracking method according to claim 1, wherein in step S5, the reconstruction model is a bottom-up reconstruction model.

Images