Drone target tracking method integrating global tracker and local tracker

Abstract

The present disclosure provides a UAV target tracking method integrating a global tracker and a local tracker, comprising: inputting a current frame image acquired by a UAV into a global backbone network in the global tracker, obtaining a feature map corresponding to a target by using the global backbone network; processing the feature map by using a similarity measurement module in the global tracker, and outputting a similarity response map; processing the similarity response map by using a classification branch and a regression branch in the global tracker, respectively, so as to determine a global tracking result; inputting the current frame image into a correlation filtering module in the local tracker, obtaining a histogram of oriented gradients feature and a color feature of a target candidate region, matching a filtering template with the histogram of oriented gradients feature and the color feature by using the correlation filtering module to obtain a correlation relationship, so as to determine a local tracking result; inputting the global tracking result and the local tracking result into an ensemble learning module, and outputting a final tracking result of the target.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle target tracking, and more particularly to an unmanned aerial vehicle target tracking method, system and device integrating a global tracker and a local tracker. BACKGROUND

[0002] With the rapid improvement of computer operation capability, the fields of computer vision, artificial intelligence and machine perception are developing rapidly, and target tracking, as one of the basic and core research topics, is applied to intelligent driving, medical diagnosis, robot vision and other life services.

[0003] At present, existing target tracking methods include a target tracking method based on correlation filtering and a target tracking method based on a twin neural network. These two types of methods have different features: the tracker based on correlation filtering focuses on the region where the target is located and uses the information of historical frames for tracking; the method based on a twin neural network focuses on the whole image range for searching. However, these two types of methods have some deficiencies: the target tracking method based on correlation filtering can explore the time domain context information similar to the target historical appearance and positioning, but the local tracker will be misled by similar interference, so that the re-detector is not activated when the target is lost; the target tracking method based on a twin neural network tracks the target in the whole image instead of performing local tracking based on the tracking results of previous frames, but such global tracking cannot effectively utilize the time domain context. SUMMARY

[0004] To solve the problems in the prior art and other problems, the present disclosure provides a UAV target tracking method and system integrating a global tracker and a local tracker, a current frame image acquired by a UAV is input into a global backbone network in the global tracker, image features of the current frame image are extracted using the global backbone network, a feature map corresponding to a target is obtained, the feature map is processed using a similarity measurement module in the global tracker, a similarity response map is output, a classification branch and a regression branch in the global tracker are used to process the similarity response map respectively, a classification score map and distances between each regression point and a prediction box are obtained, so as to determine a global tracking result according to the classification score map and the distances between each regression point and the prediction box, the current frame image is input into a correlation filtering module in the local tracker, histogram of oriented gradient features and color features of a target candidate region are acquired, the correlation filtering module is used to match a filtering template with the histogram of oriented gradient features and the color features to obtain a correlation relationship, so as to determine a local tracking result according to the correlation relationship, the global tracking result and the local tracking result are input into an ensemble learning module, and a final tracking result of the target is output, the global tracker is used to find a region where the target may be located, the local tracker is used to search and track in the region where the target may be located, and the ensemble learning module is used to integrate the global tracking result and the local tracking result, so that accurate tracking in real time is achieved.

[0005] The embodiment of the present disclosure provides a UAV target tracking method integrating a global tracker and a local tracker, which comprises the following steps: inputting a current frame image acquired by a UAV into a global backbone network in the global tracker, extracting image features of the current frame image using the global backbone network, and obtaining a feature map corresponding to a target; processing the feature map using a similarity measurement module in the global tracker, and outputting a similarity response map; processing the similarity response map using a classification branch and a regression branch in the global tracker respectively, obtaining a classification score map and distances between each regression point and a prediction box, and determining a global tracking result according to the classification score map and the distances between each regression point and the prediction box; inputting the current frame image into a correlation filtering module in the local tracker, acquiring histogram of oriented gradient features and color features of a target candidate region, matching a filtering template with the histogram of oriented gradient features and the color features using the correlation filtering module to obtain a correlation relationship, and determining a local tracking result according to the correlation relationship; and inputting the global tracking result and the local tracking result into an ensemble learning module, and outputting a final tracking result of the target.

[0006] According to some embodiments of the present disclosure, the image feature of the current frame image is extracted by using the global backbone network, including: performing convolution operation on the current frame image to obtain a first feature map; performing periodic normalization operation, nonlinear activation operation and convolution operation on the first feature map to obtain the feature map corresponding to the target.

[0007] According to some embodiments of the present disclosure, the direction gradient histogram feature of the target candidate region is obtained by inputting the current frame image into the correlation filter module in the local tracker, including: using correction operation to standardize the color space of the current frame image to adjust the contrast of the current frame image; calculating the gradient of each pixel in the current frame image to capture the contour information; dividing the current frame image into a plurality of units; counting the direction gradient histogram of each unit, wherein the abscissa of the direction gradient histogram is the angle of the gradient, and the ordinate of the direction gradient histogram is the cumulative of the gradient amplitude; forming a descriptor for each unit by using the direction gradient histogram of each unit; forming a module by using a plurality of units, and the descriptor of the module is formed by concatenating the descriptors of the units in the module; concatenating the direction gradient histogram features of all modules in the current frame image to obtain the direction gradient histogram feature of the target candidate region.

[0008] According to some embodiments of the present disclosure, the color feature of the target candidate region is obtained by inputting the current frame image into the correlation filter module in the local tracker, including: refining M colors of a color system into N colors; defining the N colors as N color containers, and subtracting 1 / N of the N color containers respectively, so that the color becomes an N-1 dimensional subspace; projecting the color onto the standard orthogonal basis of the N-1 dimensional subspace to obtain the color feature of the target candidate region.

[0009] According to some embodiments of the present disclosure, the feature map is processed by using the similarity measurement module in the global tracker to output the similarity response map, including: performing global average pooling operation and convolution operation on the feature map by using attention mechanism to obtain an adjusted feature map; performing channel dimension reduction processing on the adjusted feature map to filter the background interference in the adjusted feature map; matching the feature map output from the reduced channel with the feature map to obtain the similarity response map.

[0010] According to some embodiments of the present disclosure, the global tracking result is determined according to the classification score map and the distance between each regression point and the prediction box, including: obtaining the regression point corresponding to the coordinates of the highest point in the classification score map to obtain the distance between the regression point and the prediction box; determining the global tracking result according to the distance between the regression point and the prediction box.

[0011] According to some embodiments of the present disclosure, the UAV target tracking method further comprises: in the case that the target moves or is occluded, re-searching by using the global tracker and finding the local tracker matched with the global tracker to re-determine the local tracking result of the target.

[0012] Another aspect of the embodiments of the present disclosure provides a UAV target tracking system integrating a global tracker and a local tracker, comprising: a global tracker comprising: a global backbone network configured to extract image features of a current frame image acquired by a UAV to obtain a feature map corresponding to a target; a similarity measurement module configured to process the feature map to output a similarity response map; a classification branch and a regression branch configured to process the similarity response map respectively to obtain a classification score map and distances between each regression point and a prediction box, so as to determine a global tracking result according to the classification score map and the distances between each regression point and the prediction box; a local tracker comprising: a correlation filtering module configured to process the current frame image to obtain a histogram of oriented gradient feature and a color feature of a target candidate region, match a filtering template with the histogram of oriented gradient feature and the color feature to obtain a correlation relationship, so as to determine a local tracking result according to the correlation relationship; and an integrated learning module configured to output a final tracking result of the target according to the global tracking result and the local tracking result.

[0013] According to some embodiments of the present disclosure, the global backbone network is configured to perform convolution operation and multiple periodic normalization operation, nonlinear activation operation and convolution operation on the current frame image to obtain the feature map corresponding to the target; and the similarity measurement module is configured to perform channel dimension reduction processing on the feature map, and match the feature map output from the reduced channel with the feature map to output the similarity response map.

[0014] Another aspect of the embodiments of the present disclosure provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the methods when executing the computer program. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of a UAV target tracking method integrating a global tracker and a local tracker according to an illustrative embodiment of the present disclosure;

[0016] Figure 2 is Figure 1 a structure diagram of a global backbone network of a global tracker of an illustrative embodiment shown in FIG. 4;

[0017] Figure 3 is Figure 1 a structure diagram of a similarity measure module of a global tracker of the exemplary embodiment shown. DETAILED DESCRIPTION

[0018] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. It is to be understood, however, the embodiments are merely exemplary and do not limit the scope of the present application. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present application. However, it would be apparent to those skilled in the art that the present application can be practiced without these specific details. In other instances, well-known structures and techniques have not been shown in detail in order not to obscure aspects of the present application.

[0019] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the term "includes" and tautological equivalents thereof, means that the named feature, step, operation, and / or component is present, but not excluding the presence or addition of one or more other features, steps, operations, or components.

[0020] All terms used herein including technical and scientific terms have the same meanings as commonly understood by one of ordinary skill in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning that is consistent with the context of this specification, and not be interpreted in an idealized or overly formal way.

[0021] In the case where expressions such as "at least one of A, B, and C" are used, it generally should be interpreted to include any of one, two, or all of the items listed after the conjunction, unless otherwise specifically stated. In the case where expressions such as "at least one of A, B, or C" are used, it generally should be interpreted to include any of one, two, or all of the items listed after the conjunction, unless otherwise specifically stated.

[0022] Those skilled in the art will appreciate that embodiments of the disclosure can be provided as methods, systems, or computer program products. Accordingly, the disclosure can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the disclosure can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and the like) embodying computer-readable program code.

[0023] The disclosure is described in reference to the flowcharts and / or block diagrams of the methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in the flowcharts and / or block diagrams.

[0024] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in the flowcharts and / or block diagrams.

[0025] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to generate a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in the flowcharts and / or block diagrams.

[0026] In order to make the purposes, technical solutions and advantages of the disclosure clearer, the disclosure is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0027] In the technical solutions of the disclosure, the collection, storage, use, processing, transmission, provision, disclosure and application of the user's personal information involved in the technical solutions comply with the relevant legal regulations, necessary security measures are taken, and do not violate public order and good customs.

[0028] In the technical solutions of the present disclosure, the acquisition, collection, storage, use, processing, transmission, provision, disclosure and application of data comply with relevant laws and regulations, necessary security measures are taken, and the public order and good customs are not violated.

[0029] Figure 1 is a flowchart of a UAV target tracking method integrating a global tracker and a local tracker according to an illustrative embodiment of the present disclosure.

[0030] As shown in Figure 2 , the UAV target tracking method integrating a global tracker and a local tracker comprises: inputting a current frame image acquired by a UAV into a global backbone network in a global tracker, extracting image features of the current frame image by using the global backbone network to obtain a feature map corresponding to a target, processing the feature map by using a similarity measurement module in the global tracker to output a similarity response map, processing the similarity response map by using a classification branch and a regression branch in the global tracker to obtain a classification score map and distances between each regression point and a prediction box, so as to determine a global tracking result according to the classification score map and the distances between each regression point and the prediction box, inputting the current frame image into a correlation filter module in a local tracker to acquire a histogram of oriented gradients feature and a color feature of a target candidate region, matching a filter template with the histogram of oriented gradients feature and the color feature by using the correlation filter module to obtain an association relationship, so as to determine a local tracking result according to the association relationship, and inputting the global tracking result and the local tracking result into an ensemble learning module to output a final tracking result of the target.

[0031] The UAV target tracking method integrating a global tracker and a local tracker provided by the embodiments of the present disclosure finds the area where the target may be located by using the global tracker, searches for the target in the area where the target may be located by using the local tracker, and integrates the global tracking result and the local tracking result by using the ensemble learning module, thereby realizing real-time accurate tracking.

[0032] Figure 1 is a structure diagram of a global backbone network of a global tracker of an illustrative embodiment as shown in Figure 2 .

[0033] According to the embodiments of the present disclosure, the extraction of the image features of the current frame image by using the global backbone network comprises: performing a convolution operation on the current frame image to obtain a first feature map, performing a plurality of periodic normalization operations, nonlinear activation operations and convolution operations on the first feature map to obtain the feature map corresponding to the target.

[0034] According to the embodiments of the present disclosure, as shown in Figure 3As shown, the current frame image is first subjected to a convolution operation to obtain a first feature map, the first feature map is subjected to two periodic pooling operations, normalization operations, nonlinear activation operations and convolution operations, and then subjected to two normalization operations, nonlinear activation operations and convolution operations to obtain a feature map corresponding to the target.

[0035] According to an embodiment of the present disclosure, the role of the pooling layer is to maintain the features of the current frame image while reducing the amount of calculation, prevent network overfitting, and improve the generalization ability.

[0036] According to an embodiment of the present disclosure, the global backbone network removes the padding operation in the convolution operation and removes the fully connected layer on the basis of the original network structure. For target tracking, the center cropping process needs to be performed on the current frame image during the training process, which may cause the problem that most of the training samples are concentrated in the middle of the image. In the traditional convolution operation, the padding operation is added to expand the receptive field, but in target tracking, this may cause inaccurate positioning of the target and the problem of target positioning drift. Therefore, canceling the padding operation can keep the target in the center and ensure the translation invariance of the feature map, so as to further train and optimize.

[0037] According to an embodiment of the present disclosure, the role of the normalization operation is to normalize the same feature of different images, which can accelerate the training and convergence of the network, control the gradient during network training, prevent problems such as gradient disappearance or gradient explosion, and prevent network overfitting.

[0038] According to an embodiment of the present disclosure, the role of the nonlinear activation operation is to strengthen the expression ability of the network and increase the nonlinear factor. If the nonlinear factor is not increased, no matter how many layers the neural network has, it is linear, which may cause a significant decrease in the expression ability of the neural network.

[0039] According to an embodiment of the present disclosure, the step of the global backbone network includes but is not limited to being set to 8. The step is composed of the convolution layer and the pooling layer in the global backbone network. The step of the global backbone network is set to 8, which is a comprehensive consideration of the effects of accuracy and timeliness, and can better preserve the image features and the information of the feature map.

[0040] Figure 1 is Figure 3 A structure diagram of a similarity measurement module of a global tracker of the illustrative embodiment.

[0041] According to an embodiment of the present disclosure, processing the feature map output by the similarity measurement module in the global tracker to obtain a similarity response map comprises: performing a global average pooling operation and a convolution operation on the feature map by using an attention mechanism to obtain an adjusted feature map, performing channel dimension reduction processing on the adjusted feature map to filter background interference in the adjusted feature map, and matching the feature map output from the channel after the dimension reduction with the feature map to obtain the similarity response map.

[0042] According to an embodiment of the present disclosure, as shown in ​ , an input feature map is input, and after the input, there are two branches, the upper half is mainly a channel dimension reduction processing, which is to reduce the subsequent calculation amount, and the lower half is a process of learning a response degree weight between channels, and then the learned weight is multiplied with the image after the dimension reduction to obtain an enhanced output feature map.

[0043] According to an embodiment of the present disclosure, a traditional similarity measurement module includes two operations, first, performing channel dimension reduction processing on an input feature map, and then matching the feature map after the channel dimension reduction processing with the input feature map to output a final similarity response map, and the similarity response map is very sensitive to channel information, and direct channel dimension reduction processing can cause a large amount of information contained in the channel to be lost, thus causing the generated similarity response map to be difficult to focus on a target region. Therefore, the present disclosure adds an attention mechanism, which is a parallel branch added to the original channel dimension reduction process, first performing a global average pooling operation on the feature map to compress the spatial features of the feature map, then learning the importance of different channels by two 1x1 convolution operations to obtain an adjusted feature map, and finally outputting the adjusted feature map to the channel after the dimension reduction, which is similar to a reweighting adjustment process, so that the channels related to the target can be preserved and strengthened in the dimension reduction process, and the irrelevant channels are reduced in weight, and the finally generated similarity response map can better filter the background interference in the adjusted feature map and focus more on the target region.

[0044] According to an embodiment of the present disclosure, determining the global tracking result according to the classification score map and the distance between each regression point and the prediction box comprises: obtaining a regression point corresponding to the coordinates of the point with the highest score in the classification score map, obtaining the distance between the regression point and the prediction box, and determining the global tracking result according to the distance between the regression point and the prediction box.

[0045] According to an embodiment of the present disclosure, the classification branch is to decode features of the similarity response map to classify into two categories of target and background, and the regression branch is to decode features of the similarity response map to obtain distances between each point in the similarity response map and the prediction box, to obtain a regression point on the regression branch corresponding to coordinates of a point with the highest score in the classification score map in the classification branch, to obtain distances between the regression point and the prediction box, and to determine the global tracking result according to the distances between the regression point and the prediction box.

[0046] In an illustrative embodiment, the generated classification score map is a 25x25 matrix, each point in which has a corresponding classification score, i.e., there are 625 classification scores, and the coordinates of a point with the highest classification score are found, such as (3, 4), and the regression branch is found according to the coordinates, the regression branch outputs a 25x25x4 matrix, the regression point corresponding to the (3, 4) point in the regression branch is found, and there are four regression points corresponding to the (3, 4) point, the distances between the four regression points and the prediction box are obtained, and the global tracking result is determined according to the distances.

[0047] According to an embodiment of the present disclosure, the direction gradient histogram features of the target candidate region are obtained by inputting the current frame image into a correlation filtering module in the local tracker, including: performing color space normalization on the current frame image by using a correction operation to adjust the contrast of the current frame image, calculating the gradient of each pixel in the current frame image to capture contour information, dividing the current frame image into a plurality of units, counting the direction gradient histogram of each unit, the abscissa of the direction gradient histogram being the angle of the gradient, and the ordinate of the direction gradient histogram being the cumulative of the gradient amplitude, forming a descriptor for each unit by using the direction gradient histogram of each unit, forming a module by using a plurality of units, and concatenating the descriptors of the units in the module to form a descriptor of the module, and concatenating the direction gradient histogram features of all the modules in the current frame image to obtain the direction gradient histogram features of the target candidate region.

[0048] According to an embodiment of the present disclosure, the color space of the current frame image is normalized by using a correction operation to adjust the contrast of the current frame image, so as to reduce the influence of local shadows and light in the current frame image, and the correction operation can also suppress noise, and the correction operation includes but is not limited to a Gamma correction method.

[0049] According to an embodiment of the present disclosure, the gradient of each pixel in the current frame image is calculated to capture contour information, and further to weaken the interference of light.

[0050] According to an embodiment of the present disclosure, the color feature of the target candidate region is obtained by inputting a current frame image into a correlation filtering module in a local tracker, including: refining M colors of a color system into N colors, defining the N colors as N color containers, subtracting 1 / N of the N color containers respectively, the color will become an N-1 dimensional subspace, projecting the color onto a standard orthogonal basis of the N-1 dimensional subspace, and obtaining the color feature of the target candidate region.

[0051] According to an embodiment of the present disclosure, the color feature (CN feature) is used to track the target according to color information, and is greatly affected by factors such as illumination change. In the case that the CN feature is sensitive to illumination change, a light change feature description operator is trained to improve the ability of the CN feature to cope with obvious illumination change and other factors that will affect the color information of the tracking target in the tracking process. At the same time, a light invariant feature description operator is also trained according to the tracking process without obvious illumination change, and the corresponding operator is selected according to the tracking situation, so as to improve the ability of the present disclosure to cope with complex and variable tracking environment.

[0052] In an illustrative embodiment, the color attribute in computer vision is generally divided into red, green and blue, and the CN feature is subdivided and mapped into 11-dimensional color containers, which are black, blue, brown, gray, green, orange, pink, purple, red, white and yellow, and then 1 / 11 of each color is extracted, and the 11-dimensional mapping is changed into a 10-dimensional subspace, so that the expression ability of the CN feature is stronger, the accuracy of the present disclosure method is improved, and the PCA linear dimension reduction technology can also be used to reduce the originally over-fitted 10-dimensional feature to three dimensions, which guarantees the tracking accuracy of the present disclosure method while reducing the computational complexity and guaranteeing the real-time requirement.

[0053] Specifically, the extracted CN feature is first processed by the PCA dimension reduction method to reduce the D1-dimensional CN feature to D2-dimensional. The PCA method selects useful information while reducing dimension and eliminates interference information. The present disclosure selects a CN feature dimension reduction mapping most suitable for each frame by minimizing a loss function:

[0054]

[0055] wherein, is a current frame influence factor, is a smoothing influence factor of the i-th frame, is a total influence factor, and p is a current frame weight coefficient, i is an i-th frame image weight coefficient.

[0056]

[0057] wherein, B pis a mapping matrix of the current frame of D1xD2, [M, N] is the size of the search area of the current frame, (m, n) is the target position determined in the last frame, is a feature map of D1 dimension of the current frame.

[0058] After PCA dimension reduction processing, a dimension reduction mapping matrix B is obtained p The feature dimension of the feature map is reduced to D2 dimension.

[0059] According to an embodiment of the present disclosure, the integrated learning module is used to determine the final tracking result according to the global tracking result and the local tracking result, which can be understood as determining the final tracking result by weighting the two. The specific weights of the two modules are dynamic during the training process, and a trained result is finally obtained through the training set and the validation set. In the actual inference prediction stage, the respective weights of the two modules are the trained result, which is a fixed value. Each time tracking is performed, the global tracker and the local tracker output a result respectively, and the final tracking result is obtained by adding the corresponding weights.

[0060] According to an embodiment of the present disclosure, the unmanned aerial vehicle target tracking method further comprises: in the case that the target moves or is occluded, re-searching by using the global tracker, and finding a local tracker matched with the global tracker to re-determine the local tracking result of the target.

[0061] According to an embodiment of the present disclosure, when the target is tracked, the parameters of the tracker in the region where the target is located are constantly updated, in the case that the target moves or is occluded, re-searching by using the global tracker, and finding a local tracker matched with the global tracker to re-determine the local tracking result of the target, so that real-time accurate tracking can be realized.

[0062] According to an embodiment of the present disclosure, the unmanned aerial vehicle target tracking system integrating the global tracker and the local tracker comprises: the global tracker comprises: a global backbone network, used to extract image features of a current frame image acquired by an unmanned aerial vehicle, to obtain a feature map corresponding to a target, a similarity measurement module, used to process the feature map to output a similarity response map, a classification branch and a regression branch, used to process the similarity response map respectively to obtain a classification score map and distances between each regression point and a prediction box, so as to determine a global tracking result according to the classification score map and the distances between each regression point and the prediction box, the local tracker comprises: a correlation filtering module, used to process the current frame image to obtain a histogram of oriented gradients feature and a color feature of a target candidate region, and to match a filtering template with the histogram of oriented gradients feature and the color feature to obtain a correlation relationship, so as to determine a local tracking result according to the correlation relationship, and an integrated learning module, used to output a final tracking result of the target according to the global tracking result and the local tracking result.

[0063] According to an embodiment of the present disclosure, the global backbone network is configured to perform convolution operation on the current frame image and a plurality of periodic normalization operation, nonlinear activation operation and convolution operation, to obtain a feature map corresponding to the target, and the similarity measurement module is configured to perform channel dimension reduction processing on the feature map, and match the feature map output from the channel after the dimension reduction with the feature map, and output a similarity response map.

[0064] According to another aspect of the embodiments of the present disclosure, a computer device is provided, which comprises a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps of any of the above methods are implemented.

[0065] It should also be noted that the directional terms mentioned in the embodiments, such as "up", "down", "front", "back", "left", "right", etc., are only the directions of the drawings, and are not intended to limit the protection scope of the present disclosure. Throughout the drawings, the same elements are represented by the same or similar reference numerals. When it may cause confusion to the understanding of the present disclosure, the conventional structures or configurations will be omitted.

[0066] The above describes the embodiments of the present disclosure. However, these embodiments are only for illustrative purposes, and are not intended to limit the scope of the present disclosure. Although each embodiment is described above separately, this does not mean that the measures in each embodiment cannot be used advantageously in combination. The scope of the present disclosure is defined by the appended claims and their equivalents. Without departing from the scope of the present disclosure, those skilled in the art can make various substitutions and modifications, which should all fall within the scope of the present disclosure.

Claims

1. A UAV target tracking method integrating a global tracker and a local tracker, characterized in that, The method comprises the following steps: inputting a current frame image acquired by a UAV into a global backbone network in a global tracker, extracting image features of the current frame image by using the global backbone network, and obtaining a feature map corresponding to a target; processing the feature map by using a similarity measurement module in the global tracker, and outputting a similarity response map; processing the similarity response map by using a classification branch and a regression branch in the global tracker, obtaining a classification score map and distances between each regression point and a prediction box, and determining a global tracking result according to the classification score map and the distances between each regression point and the prediction box; inputting the current frame image into a correlation filter module in a local tracker, acquiring a histogram of oriented gradient feature and a color feature of a target candidate region, matching a filter template with the histogram of oriented gradient feature and the color feature by using the correlation filter module to obtain an association relationship, and determining a local tracking result according to the association relationship, wherein the inputting the current frame image into the correlation filter module in the local tracker to acquire the histogram of oriented gradient feature of the target candidate region comprises the following steps: performing a normalization operation on the current frame image in a color space to adjust the contrast of the current frame image by using a correction operation; calculating the gradient of each pixel in the current frame image to capture contour information; dividing the current frame image into a plurality of units; counting the histogram of oriented gradient of each unit, wherein the abscissa of the histogram of oriented gradient is the angle of the gradient, and the ordinate of the histogram of oriented gradient is the cumulative gradient amplitude; forming a descriptor of each unit by using the histogram of oriented gradient of each unit; concatenating the descriptors of a plurality of units to form a module, and concatenating the descriptors of each unit in the module to form the descriptor of the module; concatenating the histogram of oriented gradient features of all modules in the current frame image to obtain the histogram of oriented gradient feature of the target candidate region; inputting the global tracking result and the local tracking result into an integrated learning module, and outputting a final tracking result of the target. 2.The UAV target tracking method of claim 1, wherein, The method comprises the following steps: performing a convolution operation on the current frame image to obtain a first feature map; performing a plurality of periodic normalization operations, nonlinear activation operations and convolution operations on the first feature map to obtain the feature map corresponding to the target. 3.The UAV target tracking method of claim 1, wherein, The method comprises the following steps: refining M colors of a color system into N colors; defining the N colors as N color containers, respectively subtracting 1 / N of the N color containers, and the color will become an N-1-dimensional subspace; projecting the color onto a standard orthogonal basis of the N-1-dimensional subspace to obtain the color feature of the target candidate region. 4.The UAV target tracking method of claim 1, wherein, The method comprises the following steps: performing a global average pooling operation and a convolution operation on the feature map by using an attention mechanism to obtain an adjusted feature map; perform channel dimension reduction processing on the adjusted feature map to filter background interference in the adjusted feature map; match the feature map output from the dimension-reduced channel with the feature map to obtain the similarity response map. 5.The UAV target tracking method of claim 1, wherein, The global tracking result is determined according to the classification score map and the distance between each regression point and the prediction box. A regression point corresponding to the coordinates of the highest scoring point in the classification score map is obtained, and the distance between the regression point and the prediction box is obtained. The global tracking result is determined according to the distance between the regression point and the prediction box. 6.The UAV target tracking method of claim 1, wherein, The UAV target tracking method further comprises: In the case where the target moves or is occluded, the global tracker is used to search again, and the local tracker matching the global tracker is found to re-determine the local tracking result of the target.

7. An unmanned aerial vehicle target tracking system integrating a global tracker and a local tracker, characterized in that, The UAV target tracking system for performing the integrated global tracker and local tracker of any one of claims 1-6, the UAV target tracking system comprising: The global tracker comprises: The global backbone network is configured to extract image features of a current frame image obtained by a UAV to obtain a feature map corresponding to a target. The similarity measurement module is configured to process the feature map to output a similarity response map. The classification branch and the regression branch are configured to process the similarity response map respectively to obtain a classification score map and a distance between each regression point and a prediction box, so as to determine a global tracking result according to the classification score map and the distance between each regression point and the prediction box. The local tracker comprises: The correlation filtering module is configured to process the current frame image to obtain a histogram of oriented gradients feature and a color feature of a target candidate region, match a filtering template with the histogram of oriented gradients feature and the color feature to obtain a correlation relationship, and determine a local tracking result according to the correlation relationship. The integrated learning module is configured to output a final tracking result of the target according to the global tracking result and the local tracking result.

8. The UAV target tracking system of claim 7, wherein wherein The global backbone network is configured to perform convolution operation, multiple periodic normalization operation, nonlinear activation operation and convolution operation on the current frame image to obtain a feature map corresponding to a target. The similarity measurement module is configured to perform channel dimension reduction processing on the feature map, match the feature map output from the dimension-reduced channel with the feature map, and output a similarity response map. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor implements the steps of the method of any one of claims 1-6 when executing the computer program.

Images