CN107154024A - Dimension self-adaption method for tracking target based on depth characteristic core correlation filter - Google Patents

Abstract

The invention discloses a kind of dimension self-adaption method for tracking target based on depth characteristic core correlation filter.This method comprises the following steps：The convolutional neural networks of pre-training completion are input an image into, depth convolution feature is extracted；Target tracking, using the model of training, estimates position and the yardstick of target；According to the target location of current detection and yardstick, core correlation filter is trained；Using the model update method of adaptive high confidence level, core correlation filter is updated.The present invention is extracted depth convolution feature, using adaptive scale method of estimation and the model modification strategy of adaptive high confidence level, improve the robustness of target following of the target in complex scene and cosmetic variation, target scale change can efficiently and accurately be handled, in addition, due to the model modification strategy using adaptive high confidence level, model following drift has been reduced as far as.

Description

Scale Adaptive Target Tracking Method Based on Deep Feature Kernel Correlation Filter

technical field

The invention relates to the technical field of computer vision, in particular to a scale-adaptive target tracking method based on a deep feature kernel correlation filter.

Background technique

In recent years, with the emergence of large-scale labeled data sets and the improvement of computer computing power, deep learning methods, especially convolutional neural networks, have been successfully applied to computer vision fields such as image classification, target detection, target recognition, and semantic segmentation. Thanks to the powerful object representation ability of convolutional neural network. Different from traditional image features, deep convolutional features are learned from a large number of thousands of categories of image data, so high-level convolutional features represent the semantic features of the target and are suitable for image classification problems. Due to the low resolution of high-level convolutional features, it is not suitable for determining the position of the target, and due to the lack of training data, it is difficult to train a deep model in the first few frames of tracking.

Recently, the discriminative tracking method based on correlation filter has aroused the interest of many researchers because of its efficient and accurate tracking effect. The tracking method based on the correlation filter trains a correlation filter online by regressing the input features to the Gaussian distribution of the target, and finds the peak of the corresponding spectrum output by the correlation filter in the subsequent tracking process to determine the position of the target. The correlation filter cleverly applies the fast Fourier transform in the operation, which reduces the computational complexity and greatly improves the tracking speed. However, the kernel correlation filter algorithm uses the traditional gradient direction histogram feature, and the tracking algorithm is prone to drift when the appearance representation of the target changes; in addition, the algorithm cannot estimate the scale change of the target, and uses simple linear interpolation to update the model, so When an error occurs in target tracking, the tracking algorithm drifts due to the update mechanism.

Contents of the invention

The purpose of the present invention is to provide a scale-adaptive target tracking method based on the depth feature kernel correlation filter, thereby improving the robustness of target tracking in complex scenes and appearance changes, and efficiently and accurately processing target scale changes, as much as possible Potentially reduce model tracking drift due to false detections.

The technical solution to realize the object of the present invention is: a scale-adaptive target tracking method based on depth feature kernel correlation filter, comprising the following steps:

Step 1, input the initial position p ₀ and scale s ₀ of the target, and set the window size to 2.0 times the initial bounding box of the target;

Step 2, according to the target position p _t-1 in the t-1th frame, obtain the target area x _t-1 , whose size is the window size;

Step 3, extract the deep convolution feature of the target area x _t-1 , and perform fast Fourier transform to obtain the feature map Wherein ^ represents the discrete Fourier transform;

Step 4, according to the feature map Computing Kernel Autocorrelation

Step 5, training location and scale correlation filters;

Step 6, according to the position p _t-1 of the target in the t-1th frame, obtain the candidate area z _t of the target in the t-th frame, and the size is the window size;

Step 7, extract the depth convolution feature of the candidate area z _t , and perform fast Fourier transform to obtain the feature map Wherein ^ represents the discrete Fourier transform;

Step 8, according to the feature map of the previous frame of the target Computing Kernel Cross-Correlation

Step 9, detecting the position corresponding to the maximum value in the output spectrum of the position filter and the scale filter respectively, to determine the position p _t and scale _st of the target in the current frame;

In step 10, an adaptive model updating strategy is used to update the kernel correlation filter.

Further, the extraction method of the depth convolution feature described in step 3 and step 7 is as follows:

(3.1) Preprocessing, scaling the window area I to the input size 224×224 specified by the convolutional neural network;

(3.2) Feature extraction, extracting the feature maps of the 3rd, 4th and 5th convolutional layers of the convolutional neural network;

(3.3) Bilinear interpolation, upsampling the extracted 3-layer convolution features to the same size.

Further, the calculation kernel autocorrelation described in step 4 Cross-correlate with the calculation kernel described in step 8 The specific method is as follows:

(4.1) Using the Gaussian kernel, the formula is as follows:

Among them, k(x, x′) represents the Gaussian kernel calculated by two feature maps x and x′, exp(.) represents the e exponential function, σ is the standard deviation of the Gaussian function, and the value of σ is 0.5, ||.| | ² means the 2 normal form of the vector or matrix;

(4.2) Calculate the kernel correlation, the formula is as follows:

Among them, k ^xx' represents the kernel correlation of feature maps x and x', exp(.) represents the e exponential function, σ is the standard deviation of the Gaussian function, and the value of σ is 0.5, and ||.|| ² represents the vector or matrix 2 paradigms, Represents the inverse transform of the discrete Fourier transform, * represents the complex conjugate, ^ represents the discrete Fourier transform, and ⊙ represents the multiplication of the corresponding elements of the two matrices.

Further, the training location and scale correlation filters described in step 5 are as follows:

According to the deep convolution features extracted in step 3, a kernel correlation filter is trained for each layer of feature maps, and the following formula is used to train the model:

in, Indicates that the feature map is convolved according to the depth of the l layer The obtained correlation filter model, Represents a feature map Kernel autocorrelation of , ^ means discrete Fourier transform, λ is a regularization parameter to prevent overfitting of the trained model, and the value of λ is 0.001.

Further, in step 9, respectively detect the position corresponding to the maximum value in the output spectrum of the position filter and the scale filter to determine the position p _t and scale _st of the target in the current frame, as follows:

(9.1) From the tth frame image, select the candidate area z _t,trans with the position p _t-1 and the window size;

(9.2) Extract the 3-layer deep convolution feature map of the candidate area z _{t, trans}

(9.3) Calculate the position filter correlation output confidence map f _t,trans ^(l) for the l-th layer feature map:

Among them, f _t ^(l) represents the output corresponding map of the position filter of the l-th layer feature map, Represents a feature map with The nuclear cross-correlation of is the position filter trained and updated in the previous frame, Represents the inverse transform of the discrete Fourier transform, ^ represents the discrete Fourier transform, Indicates that the corresponding elements of the two matrices are multiplied;

(9.4) Estimating the target position from coarse to fine from the output corresponding map f _t ⁽³⁾ , the target position p _t ^(l) of the first layer is the position corresponding to the maximum value of the output corresponding map f _t,trans ^(l) ;

(9.5) From the t-th frame image, extract the candidate region z _t,sacle for scale estimation with the position p _t and scale s _t-1 , and construct a scale pyramid;

(9.6) Extract the gradient direction histogram feature of the candidate area Calculate scale filter correlation output confidence map f _t,sacle ;

(9.7) The target scale s _t detected in the tth frame is the scale corresponding to the maximum value of the output corresponding map f _t,sacle .

Further, as described in step 10, an adaptive model update strategy is used to update the kernel correlation filter, specifically as follows:

(10.1) After estimating the target position and scale, calculate two tracking result confidence metrics, one of which is the peak value of the corresponding output spectrum:

f _max = maxf

Wherein, f is the corresponding map output by the kernel correlation filter and the candidate region, and f _max is the peak value of the map;

The other is the average peak correlation energy APCE:

Among them, f _max and f _min are the maximum value and minimum value of the output spectrum f respectively, mean(.) represents the mean function, and f _{i, j} represent the value of the i-th row and the j-th column of the output spectrum f;

(10.2) If f _max and APCE are both greater than f _max and the historical average value of APCE, the model is updated, otherwise, no update is performed; for each deep convolutional layer, the linear interpolation method is used to update, the formula is as follows:

in, with respectively represent the feature map and correlation filter of the previous frame of the first layer, η is the learning rate, the larger η is, the faster the model is updated, and the value of η is 0.02.

Compared with the prior art, the present invention has the following significant advantages: (1) using deep convolution features, which are learned from a large number of thousands of categories of image data, and have powerful target representation capabilities, making the algorithm It has strong robustness when the appearance of the target changes and external factors such as illumination changes; (2) adopts the adaptive scale estimation method, which is similar to the position estimation method, trains a separate scale filter, and cleverly uses the fast Fu Lie transform, which achieves high efficiency and accurate scale estimation, can be combined into any discriminative tracking algorithm framework; (3) adopts an adaptive high-confidence model update strategy, when an error occurs in the target tracking phase, the confidence of target detection At this time, the model should not be updated, and this strategy effectively reduces the risk of tracking algorithm drift.

Description of drawings

FIG. 1 is a flow chart of the scale-adaptive target tracking method based on the deep feature kernel correlation filter of the present invention.

Figure 2 is a schematic diagram of extracting deep convolution features.

Fig. 3 is a schematic diagram of target position and scale estimation, where (a) is a schematic diagram of target position estimation from coarse to fine, and (b) is a schematic diagram of adaptive target scale estimation.

Fig. 4 is a schematic diagram of an adaptive high-confidence model update.

Fig. 5 is the evaluation result diagram of the present invention on the standard visual tracking data set, wherein (a) is the accuracy drawing of the OTB50 data set, (b) is the accuracy drawing of the OTB50 data set, (c) is the accurate drawing of the OTB100 data set Degree drawing, (d) is the correct rate drawing of the OTB100 dataset.

Fig. 6 is the actual video target tracking result diagram of the present invention, wherein (a) is the Human test video result diagram on the OTB100 dataset, (b) is the Walking test video result diagram on the OTB100 dataset, and (c) is the Tiger test video diagram on the OTB50 dataset Test video result graph, (d) is the Dog test video result graph on the OTB50 dataset.

detailed description

In order to further understand and understand the structural features of the present invention and the achieved effects, a detailed description is provided in conjunction with preferred embodiments and accompanying drawings, as follows:

In conjunction with Fig. 1, the scale adaptive target tracking method based on the depth feature kernel correlation filter of the present invention includes the following steps:

Step 1, input the initial position p ₀ and scale s ₀ of the target, and set the window size to 2.0 times the initial bounding box of the target;

Step 2, according to the target position p _t-1 in the t-1th frame, obtain the target area x _t-1 , whose size is the window size;

Step 3, extract the deep convolution feature of the target area x _t-1 , and perform fast Fourier transform to obtain the feature map Wherein ^ represents the discrete Fourier transform;

Step 4, according to the feature map Computing Kernel Autocorrelation

Step 5, training location and scale correlation filters;

Step 6, according to the position p _t-1 of the target in the t-1th frame, obtain the candidate area z _t of the target in the t-th frame, and the size is the window size;

Step 7, extract the depth convolution feature of the candidate area z _t , and perform fast Fourier transform to obtain the feature map Wherein ^ represents the discrete Fourier transform;

Step 8, according to the feature map of the previous frame of the target Computing Kernel Cross-Correlation

Step 9, detecting the position corresponding to the maximum value in the output spectrum of the position filter and the scale filter respectively, to determine the position p _t and scale _st of the target in the current frame;

In step 10, an adaptive model updating strategy is used to update the kernel correlation filter.

As a specific example, the extraction method of the depth convolution feature described in step 3 and step 7 is as follows:

(3.1) Preprocessing, scaling the window area I to the input size 224×224 specified by the convolutional neural network;

(3.2) Feature extraction, extracting the feature maps of the 3rd, 4th and 5th convolutional layers of the convolutional neural network;

(3.3) Bilinear interpolation, upsampling the extracted 3-layer convolution features to the same size.

As a specific example, the calculation kernel autocorrelation described in step 4 Cross-correlate with the calculation kernel described in step 8 The specific method is as follows:

(4.1) Using the Gaussian kernel, the formula is as follows:

Among them, k(x, x′) represents the Gaussian kernel calculated by two feature maps x and x′, exp(.) represents the e exponential function, σ is the standard deviation of the Gaussian function, and the value of σ is 0.5, ||.| | ² means the 2 normal form of the vector or matrix;

(4.2) Calculate the kernel correlation, the formula is as follows:

Among them, k ^xx' represents the kernel correlation of feature maps x and x', exp(.) represents the e exponential function, σ is the standard deviation of the Gaussian function, and the value of σ is 0.5, and ||.|| ² represents the vector or matrix 2 paradigms, Represents the inverse transform of the discrete Fourier transform, * represents the complex conjugate, ^ represents the discrete Fourier transform, Represents the multiplication of corresponding elements of two matrices.

As a specific example, the training position and scale correlation filter described in step 5 is as follows:

According to the deep convolution features extracted in step 3, a kernel correlation filter is trained for each layer of feature maps, and the following formula is used to train the model:

in, Indicates that the feature map is convolved according to the depth of the l layer The obtained correlation filter model, Represents a feature map Kernel autocorrelation of , ^ means discrete Fourier transform, λ is a regularization parameter to prevent overfitting of the trained model, and the value of λ is 0.001.

As a specific example, in step 9, respectively detect the position corresponding to the maximum value in the output spectrum of the position filter and the scale filter to determine the position p _t and scale _st of the target in the current frame, as follows:

(9.1) From the tth frame image, select the candidate area z _t,trans with the position p _t-1 and the window size;

(9.2) Extract the 3-layer deep convolution feature map of the candidate area z _{t, trans}

(9.3) Calculate the position filter correlation output confidence map f _t,trans ^(l) for the l-th layer feature map:

Among them, f _t ^(l) represents the output corresponding map of the position filter of the l-th layer feature map, Represents a feature map with The nuclear cross-correlation of is the position filter trained and updated in the previous frame, Represents the inverse transform of the discrete Fourier transform, ^ represents the discrete Fourier transform, Indicates that the corresponding elements of the two matrices are multiplied;

(9.4) Estimating the target position from coarse to fine from the output corresponding map f _t ⁽³⁾ , the target position p _t ^(l) of the first layer is the position corresponding to the maximum value of the output corresponding map f _t,trans ^(l) ;

(9.5) From the t-th frame image, extract the candidate region z _t,sacle for scale estimation with the position p _t and scale s _t-1 , and construct a scale pyramid;

(9.6) Extract the gradient direction histogram feature of the candidate area Calculate scale filter correlation output confidence map f _t,sacle ;

(9.7) The target scale s _t detected in the tth frame is the scale corresponding to the maximum value of the output corresponding map f _t,sacle .

As a specific example, in step 10, an adaptive model update strategy is used to update the kernel correlation filter, specifically as follows:

(10.1) After estimating the target position and scale, calculate two tracking result confidence metrics, one of which is the peak value of the corresponding output spectrum:

f _max = maxf

Wherein, f is the corresponding map output by the kernel correlation filter and the candidate region, and f _max is the peak value of the map;

The other is the average peak-to-correlation energy (APCE):

Among them, f _max and f _min are the maximum value and minimum value of the output spectrum f respectively, mean(.) represents the mean function, and f _{i, j} represent the value of the i-th row and the j-th column of the output spectrum f;

(10.2) If f _max and APCE are both greater than f _max and the historical average value of APCE, the model is updated, otherwise, no update is performed; for each deep convolutional layer, the linear interpolation method is used to update, the formula is as follows:

in, with respectively represent the feature map and correlation filter of the previous frame of the first layer, η is the learning rate, the larger η is, the faster the model is updated, and the value of η is 0.02.

The invention solves the defect of tracking failure caused by object deformation, illumination change, object rotation, scale change and other appearance changes, as well as object occlusion. With the help of the powerful target representation ability of deep convolution features, the robustness of target tracking in complex scenes and appearance changes is improved; in addition, the present invention can efficiently and accurately deal with target scale changes. Finally, due to the use of adaptive high Confidence-based model update strategy that reduces model tracking drift due to false detections.

The present invention will be described in further detail below in conjunction with specific embodiments.

Example 1

The present invention is based on the scale adaptive target tracking method of the deep feature kernel correlation filter, the method is mainly divided into four steps, the first step is to extract the depth convolution feature; the second step is to train the kernel correlation filter; the third step is to estimate the target at The position and scale of the current frame; the fourth step is to adopt an adaptive high-confidence model update strategy.

Step 1, input the initial position p ₀ and scale s ₀ of the target, and set the window size to 2.0 times the initial bounding box of the target;

Step 2, according to the target position p _t-1 of the t-1th frame, obtain the target area x _t-1 , whose size is the window size;

Step 3, extract the deep convolution feature of the target area x _t-1 , fast Fourier transform, and obtain the feature map Wherein ^ represents the discrete Fourier transform;

Step 4, according to the feature map Computing Kernel Autocorrelation

Step 5, training location and scale correlation filters;

Step 6, according to the position p _t-1 of the target in the t-1th frame, obtain the candidate area z _t of the target in the t-th frame, and the size is the window size;

Step 7, extract the depth convolution feature of the candidate area z _t , and perform fast Fourier transform to obtain the feature map Wherein ^ represents the discrete Fourier transform;

Step 8, according to the feature map of the previous frame of the target Computing Kernel Cross-Correlation

Step 9, respectively detecting the position corresponding to the maximum value of the output corresponding map of the position filter and the scale filter to determine the position p _t and scale _st of the target in the current frame;

In step 10, an adaptive high-confidence model update strategy is used to update the kernel correlation filter.

As shown in Figure 2, a schematic diagram of extracting deep convolutional features of the scale-adaptive target tracking method based on deep feature kernel correlation filters is given. The present invention is based on the scale-adaptive target tracking method of the deep feature kernel correlation filter, which is characterized in that the deep convolution feature is extracted, and the deep convolution feature is learned from a large number of thousands of categories of image data, and has a strong target representation ability , which makes the algorithm more robust when the appearance of the target changes and external factors such as illumination changes. The specific steps are as follows:

(3.1) Preprocessing, scaling the window area I to the input size 224×224 specified by the convolutional neural network;

(3.2) Feature extraction, extracting the feature maps of the 3rd, 4th and 5th convolutional layers of the convolutional neural network;

(3.3) Bilinear interpolation, upsampling the extracted 3-layer convolution features to the same size.

As shown in Figure 3, a schematic diagram of the target position and scale estimation of the scale-adaptive target tracking method based on the deep feature kernel correlation filter is given, where Figure 3(a) is a schematic diagram of target position estimation from coarse to fine, and Figure 3 (b) is a schematic diagram of adaptive target scale estimation. The scale-adaptive target tracking method based on the deep feature kernel correlation filter of the present invention is characterized by a coarse-to-fine layered position estimation method and an adaptive target scale estimation method. The model size of the target in the traditional kernel correlation filtering method is fixed, which cannot deal with the change of the target scale, so it is easy to cause tracking failure. The present invention proposes an adaptive scale estimation method, the main idea of which is to train an independent scale filter and estimate the scale corresponding to the maximum correlation response of the scale filter. This method cleverly uses fast Fourier transform, which is simple and efficient , which can be integrated into the traditional discriminative target tracking method, the specific steps are as follows:

(9.1) From the tth frame image, select the candidate area z _t,trans with the position p _t-1 and the window size;

(9.2) Extract the 3-layer deep convolution feature map of the candidate area z _{t, trans}

(9.3) Calculate the position filter correlation output confidence map f _t,trans ^(l) for the l-th layer feature map:

Among them, f _t ^(l) represents the output corresponding map of the position filter of the l-th layer feature map, Represents a feature map with The nuclear cross-correlation of is the position filter trained and updated in the previous frame, Represents the inverse transform of the discrete Fourier transform, ^ represents the discrete Fourier transform, Represents the multiplication of corresponding elements of two matrices.

(9.4) Estimating the target position from coarse to fine from the output corresponding map f _t ⁽³⁾ , the target position p _t ^(l) of the first layer is the position corresponding to the maximum value of the output corresponding map f _t,trans ^(l) ;

(9.5) From the t-th frame image, extract the candidate region z _t,sacle for scale estimation with the position p _t and scale s _t-1 , and construct a scale pyramid;

(9.6) Extract the gradient direction histogram feature of the candidate area Calculate scale filter correlation output confidence map f _t,sacle ;

(9.7) The target scale s _t detected in the tth frame is the scale corresponding to the maximum value of the output corresponding map f _t,sacle .

As shown in Figure 4, a schematic diagram of the adaptive high-confidence model update of the scale-adaptive target tracking method based on the deep feature kernel correlation filter is given. The scale adaptive target tracking method based on the depth feature kernel correlation filter of the present invention is characterized by an adaptive high-confidence model update strategy. The traditional nuclear correlation filter model update method uses a simple linear interpolation method without tracking error detection. When the target detection error occurs during tracking, the model update will pollute the nuclear correlation filter and cause the tracking algorithm to drift. The present invention proposes two detection confidence criteria. When it is determined that the target is tracked correctly, the model is updated; otherwise, the model is not updated. The specific steps are as follows:

(10.1) After estimating the target position and scale, calculate two tracking result confidence metrics, one of which is the peak value of the corresponding output spectrum:

f _max = maxf

Among them, f is the corresponding map output by the kernel correlation filter and the candidate region, and f _max is the peak value of the map. The other is the average peak-to-correlation energy (APCE):

Among them, f _max and f _min are the maximum and minimum values of the output spectrum f, and mean(.) represents the mean function. f _{i, j} represents the value of row i and column j of the output spectrum f.

(10.2) If both f _max and APCE are greater than f _max and APCE historical average, the model is updated, otherwise, no update is performed. For each deep convolutional layer, it is updated using the linear interpolation method, the formula is as follows:

in, with Respectively represent the feature map and the correlation filter of the previous frame of the first layer, η is the learning rate, the larger the η, the faster the model is updated, and the value of η is 0.02 in the invention.

As shown in Figure 5, it shows the evaluation results of the present invention on the standard visual tracking data sets OTB50 and OTB100, wherein (a) is the accuracy drawing of the OTB50 data set, (b) is the correct rate drawing of the OTB50 data set, ( c) is the accuracy plot of the OTB100 dataset, and (d) is the accuracy plot of the OTB100 dataset. The OTB50 dataset has 50 video sequences with a total of 29,000 frames, while the OTB100 dataset has 100 video sequences with a total of 58,897 frames, each of which has a target label. There are two main evaluation indicators: accuracy and success rate. In the accuracy plots (a) and (c), the accuracy is defined as the percentage of the number of frames whose distance between the algorithm detection position and the target calibration position does not exceed 20 pixels to the total number of evaluation frames; in the success rate plot (b) In (d) and (d), the overlap rate refers to the percentage of the overlapping area (intersection operation) between the algorithm detection target bounding box and the target calibration bounding box to the total area (combination operation) and the number of frames exceeding 50% of the total evaluation Percentage of frames. It can be seen from the evaluation results that, compared with the classic target tracking algorithms, the present invention (fSCF2) performs well in the target tracking task.

As shown in Figure 6, it shows the comparison between the present invention and some excellent algorithms in recent years in the actual video target tracking results, where (a) is the Human test video result graph on the OTB100 dataset, (b) is the OTB100 dataset The results of the Walking test video, (c) is the result of the Tiger test video on the OTB50 dataset, and (d) is the result of the Dog test video on the OTB50 dataset. Overall, compared with the classic target tracking algorithm, the present invention (fSCF2) has the best tracking effect. Due to the use of deep convolution features, it has strong target representation ability, adaptive scale estimation mechanism and adaptive high confidence The model updating strategy enables the present invention to track the target accurately even under adverse conditions such as target occlusion, scale change, target deformation and fast movement of the target.

Claims (6)

1. A scale self-adaptive target tracking method based on a depth feature kernel correlation filter is characterized by comprising the following steps:

step 1, inputting an initial position p of a target₀Sum scale s₀Setting the window size to be 2.0 times of the target initial bounding box;

step 2, according to the target position p of the t-1 th frame_t-1Obtaining a target area x_t-1The size is the window size;

step 3, extracting a target area x_t-1Is characterized by deep convolution ofPerforming fast Fourier transform to obtain characteristic spectrumWherein ^ represents a discrete Fourier transform;

step 4, according to the characteristic mapCompute kernel autocorrelation

Step 5, training a position and scale correlation filter;

step 6, according to the position p of the t-1 frame target_t-1Obtaining a candidate region z of the target in the t-th frame_tThe size is the window size;

step 7, extracting candidate region z_tAnd performing fast Fourier transform to obtain a feature mapWherein ^ represents a discrete Fourier transform;

step 8, according to the characteristic map of the previous frame of the targetComputing kernel cross-correlation

Step 9, respectively detecting the positions corresponding to the maximum values in the output maps of the position filter and the scale filter to determine the position p of the target in the current frame_tSum scale s_t；

And step 10, updating the kernel correlation filter by adopting a self-adaptive model updating strategy.

2. The method for tracking scale-adaptive targets based on the depth feature kernel correlation filter according to claim 1, wherein the method for extracting depth convolution features in steps 3 and 7 specifically comprises the following steps:

(3.1) preprocessing, scaling the window area I to the input size 224 × 224 specified by the convolutional neural network;

(3.2) extracting features, namely extracting feature maps of the 3 rd, 4 th and 5 th convolutional layers of the convolutional neural network;

and (3.3) carrying out bilinear interpolation, and upsampling the extracted 3-layer convolution characteristics to the same size.

3. The method for scale-adaptive target tracking based on depth feature kernel correlation filter according to claim 1, wherein the step 4 is to calculate kernel autocorrelationAnd the step 8 of calculating the kernel cross correlationThe specific method comprises the following steps:

(4.1) using a Gaussian kernel, the formula is as follows:

<mrow> <mi>k</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <msup> <mi>x</mi> <mo>&amp;prime;</mo> </msup> <mo>)</mo> </mrow> <mo>=</mo> <mi>exp</mi> <mrow> <mo>(</mo> <mo>-</mo> <mfrac> <mn>1</mn> <msup> <mi>&amp;sigma;</mi> <mn>2</mn> </msup> </mfrac> <mo>|</mo> <mo>|</mo> <mi>x</mi> <mo>-</mo> <msup> <mi>x</mi> <mo>&amp;prime;</mo> </msup> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> <mo>)</mo> </mrow> </mrow>

wherein k (x, x ') represents a Gaussian kernel calculated by two feature maps x and x', exp (eta) represents an e index function, sigma is a standard deviation of the Gaussian function, the value of sigma is 0.5, | | |²2-normal form representing a vector or matrix;

(4.2) calculating the kernel correlation, the formula is as follows:

wherein k is^xx′Representing the nuclear correlation of the characteristic maps x and x', exp (eta) representing an e index function, sigma is the standard deviation of a Gaussian function, the value of sigma is 0.5, | | |. | tory²A 2-normal form representing a vector or matrix,represents the inverse of the discrete fourier transform, denotes the complex conjugate, denotes the discrete fourier transform, and ⊙ denotes the multiplication of the corresponding elements of the two matrices.

4. The method for tracking a scale-adaptive target based on a depth feature kernel correlation filter according to claim 1, wherein the training position and scale correlation filter of step 5 is specifically as follows:

respectively training 1 kernel correlation filter for each layer of feature spectrum according to the deep convolution features extracted in the step 3, and training a model by adopting the following formula:

<mrow> <msup> <mover> <mi>&amp;alpha;</mi> <mo>^</mo> </mover> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> <mo>=</mo> <mfrac> <mover> <mi>y</mi> <mo>^</mo> </mover> <mrow> <msup> <mover> <mi>k</mi> <mo>^</mo> </mover> <mrow> <msup> <mi>xx</mi> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> </mrow> </msup> <mo>+</mo> <mi>&amp;lambda;</mi> </mrow> </mfrac> </mrow>

wherein,representing a feature map convolved according to the l-th layer depthThe obtained correlation filter model is used for calculating the correlation filter model,representation feature mapRepresents discrete Fourier transform, lambda is a regularization parameter to prevent overfitting of the trained model, and lambda is 0.001.

5. The method according to claim 1, wherein the step 9 detects the position corresponding to the maximum value in the output maps of the position filter and the scale filter to determine the position p of the target in the current frame_tSum scale s_tThe method comprises the following steps:

(9.1) from the t-th frame image, with position p_t-1And selecting candidate area z by window size_t,trans；

(9.2) extracting candidate region z_t,trans3 layers depth convolution characteristic map

(9.3) calculating a position filter correlation output confidence coefficient map f for the first layer characteristic map_t,trans ^(l)：

Wherein f is_t ^（l)The output corresponding map of the position filter representing the characteristic map of the l-th layer,representation feature mapAndthe cross-correlation of the kernels of (a),is the position filter trained and updated from the previous frame,represents the inverse of the discrete Fourier transform, represents the discrete Fourier transform, ⊙ represents the multiplication of the corresponding elements of the two matrices;

(9.4) outputting the corresponding map f_t ⁽³⁾Starting to estimate the target position from coarse to fine, the target position p of the l-th layer_t ^(l)For outputting the corresponding map f_t,trans ^(l)The position corresponding to the maximum value;

(9.5) from the t-th frame image, with position p_tSum scale s_t-1Extracting candidate region z of scale estimate_t,sacleConstructing a scale pyramid;

(9.6) extracting gradient direction histogram feature of candidate regionComputing a scale filter correlation output confidence map f_t,sacle；

(9.7) target dimension s detected at the t-th frame_tFor outputting the corresponding map f_t,sacleThe maximum value corresponds to the scale.

6. The method for scale-adaptive target tracking based on the depth feature kernel correlation filter according to claim 1, wherein the step 10 updates the kernel correlation filter by using an adaptive model update strategy, specifically as follows:

(10.1) after completing the estimation of the target position and scale, calculating two tracking result confidence measure criteria, one of which is the peak value of the relevant corresponding output map:

f_max＝maxf

wherein f is a corresponding map output by the correlation between the kernel correlation filter and the candidate region, f_maxIs the peak of the map;

the other is the average peak correlation energy APCE:

<mrow> <mi>A</mi> <mi>P</mi> <mi>C</mi> <mi>E</mi> <mo>=</mo> <mfrac> <mrow> <mo>|</mo> <msub> <mi>f</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>f</mi> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> <msup> <mo>|</mo> <mn>2</mn> </msup> </mrow> <mrow> <mi>m</mi> <mi>e</mi> <mi>a</mi> <mi>n</mi> <mrow> <mo>(</mo> <msub> <mi>&amp;Sigma;</mi> <mrow> <mi>i</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <msup> <mrow> <mo>(</mo> <msub> <mi>f</mi> <mrow> <mi>i</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>f</mi> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow>

wherein f is_maxAnd f_minRespectively, of the output map fLarge and minimum values, mean () representing the averaging function, f_i,jThe ith row and jth column of values representing the output map f;

(10.2) if f_maxAnd APCE are both greater than f_maxAnd APCE historical average value, updating the model, otherwise, not updating; for each depth convolution layer, updating using a linear interpolation method, the formula is as follows:

<mrow> <msup> <msub> <mover> <mi>&amp;alpha;</mi> <mo>^</mo> </mover> <mi>t</mi> </msub> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> <mo>=</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <mi>&amp;eta;</mi> <mo>)</mo> </mrow> <msup> <msub> <mover> <mi>&amp;alpha;</mi> <mo>^</mo> </mover> <mrow> <mi>t</mi> <mo>-</mo> <mn>1</mn> </mrow> </msub> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> <mo>+</mo> <mi>&amp;eta;</mi> <msup> <msub> <mover> <mi>&amp;alpha;</mi> <mo>^</mo> </mover> <mi>t</mi> </msub> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> </mrow>2

<mrow> <msup> <msub> <mover> <mi>x</mi> <mo>^</mo> </mover> <mi>t</mi> </msub> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> <mo>=</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <mi>&amp;eta;</mi> <mo>)</mo> </mrow> <msup> <msub> <mover> <mi>x</mi> <mo>^</mo> </mover> <mrow> <mi>t</mi> <mo>-</mo> <mn>1</mn> </mrow> </msub> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> <mo>+</mo> <mi>&amp;eta;</mi> <msup> <msub> <mover> <mi>x</mi> <mo>^</mo> </mover> <mi>t</mi> </msub> <mrow> <mo>(</mo> <mi>l</mi> <mo>)</mo> </mrow> </msup> </mrow>

wherein,andthe feature map and the related filter of the previous frame of the ith layer are respectively represented, η is the learning rate, the larger η is, the faster model is updated, and η takes the value of 0.02.