Frequency distribution alignment-based spacecraft cross-domain attitude estimation method and device

Abstract

The invention provides a spacecraft cross-domain attitude estimation method and device based on frequency distribution alignment, and the method comprises the steps: dividing the frequency of source domain data and the frequency of target domain data into a plurality of frequency groups based on a first model; based on the plurality of frequency groups, performing local distribution alignment on the source domain data and the target domain data, and based on the data subjected to local distribution alignment, obtaining pseudo target domain data; performing supervised training fine tuning on the first model based on the pseudo target domain data to obtain a second model; obtaining spacecraft attitude estimation results in different illumination scenes based on target domain data and the second model; wherein the first model is a deep learning model obtained by training according to a source domain data set. Therefore, the generalization ability and the estimation accuracy of the model in a target domain are improved, and the attitude estimation of the spacecraft can be more accurate and stable especially in complex scenes such as illumination variation and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of spacecraft attitude estimation, and in particular to a spacecraft cross-domain attitude estimation method and device based on frequency distribution alignment. Background Art

[0002] Monocular six degrees of freedom (6-DoF) attitude estimation is crucial for spacecraft to achieve precise navigation and positioning. It helps to obtain the position and rotation information of the spacecraft in real time, so that the spacecraft can autonomously adjust its attitude in different space environments and improve the efficiency of mission execution.

[0003] The current method for monocular 6-DoF attitude estimation of spacecraft is mainly to directly migrate the existing deep learning model technology from synthetic simulation data sets (source domain) to real scenes with different lighting (target domain). However, the deep learning model obtained in this way has low accuracy in estimating the monocular 6-DoF attitude of spacecraft under different lighting scenes. Summary of the invention

[0004] The present invention provides a method and device for spacecraft cross-domain attitude estimation based on frequency distribution alignment, which is used to solve the defect that the accuracy of deep learning models in the prior art for estimating the monocular 6-DoF attitude of spacecraft in different lighting scenes is not high, and to improve the accuracy of deep learning models in estimating the monocular 6-DoF attitude of spacecraft in different lighting scenes.

[0005] The present invention provides a method for spacecraft cross-domain attitude estimation based on frequency distribution alignment, comprising the following steps: Based on the first model, the frequencies of the source domain data and the frequencies of the target domain data are divided into a plurality of frequency groups; Based on the multiple frequency groups, locally aligning the source domain data with the target domain data, and obtaining pseudo target domain data based on the locally aligned data; Based on the pseudo target domain data, fine-tuning the first model through supervised training to obtain a second model; Based on the target domain data and the second model, obtaining spacecraft attitude estimation results under different lighting scenarios; Among them, the first model is a deep learning model trained according to the source domain dataset.

[0006] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, based on a first model, the frequency of source domain data and the frequency of target domain data are divided into multiple frequency groups, specifically including: Based on the first model, determining a low-frequency partition value and a high-frequency partition value of the entire frequency band of the source domain data and the target domain data; Based on the low frequency division value and the high frequency division value, the entire frequency band is divided into a low frequency group, a medium frequency group and a high frequency group; Based on the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group, a plurality of frequency groups of source domain data and a plurality of frequency groups of target domain data are determined.

[0007] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, based on the low-frequency frequency group, the medium-frequency frequency group and the high-frequency frequency group, multiple frequency groups of source domain data and multiple frequency groups of target domain data are determined, specifically including: For each frequency group among the low frequency group, the medium frequency group and the high frequency group, based on information represented by each frequency group, determining a plurality of frequency subgroups of each frequency group; Based on the multiple frequency subgroups respectively corresponding to the low-frequency frequency group, the medium-frequency frequency group and the high-frequency frequency group, multiple frequency groups of source domain data and multiple frequency groups of target domain data are determined.

[0008] According to a method for spacecraft cross-domain attitude estimation based on frequency distribution alignment provided by the present invention, based on the information represented by each frequency group, multiple frequency subgroups of each frequency group are determined, specifically including: For the low-frequency frequency group, with the goal of making the distribution of multiple frequency subgroups of the source domain data close to that of multiple frequency subgroups of the target domain data, determine multiple frequency subgroups of the low-frequency frequency group; For the intermediate frequency group, with the goal of maintaining the distribution of multiple frequency subgroups of source domain data and multiple frequency subgroups of target domain data within a preset range, determine multiple frequency subgroups of the intermediate frequency group; For the high frequency group, a plurality of frequency subgroups of the high frequency group are determined with the goal that the distribution of the plurality of frequency subgroups of the source domain data and the plurality of frequency subgroups of the target domain data are not close.

[0009] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, based on a first model, a low-frequency division value and a high-frequency division value of the entire frequency band of source domain data and target domain data are determined, specifically including: The source domain satellite image and the target domain satellite image are converted into the frequency domain space by discrete Fourier transform; Selecting a plurality of different groups of initial low-frequency division values ​​and initial high-frequency division values, each group of different initial low-frequency division values ​​and initial high-frequency division values ​​corresponds to three different frequency groups; By inverse discrete Fourier transform, the frequency sum of each frequency group is converted into an image space to obtain an image corresponding to each frequency group in a source domain and an image corresponding to each frequency group in a target domain; Based on a histogram alignment function, aligning the images of the source domain corresponding to the respective frequency groups with the images of the target domain corresponding to the respective frequency groups to obtain aligned images of the respective frequency groups; Based on the aligned images of the frequency groups and the loss function value of the first model, the low-frequency division value and the high-frequency division value of the entire frequency band of the source domain data and the target domain data are determined.

[0010] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, pseudo target domain data is obtained based on local distribution aligned data, specifically including: Dividing a plurality of frequency groups in a plurality of different ways, and obtaining a plurality of groups of different locally distributed aligned data corresponding to the plurality of different ways; Based on the pixel distribution metric function, the overall optimal locally distributed aligned data is selected from the multiple groups of different distributed aligned data as the pseudo target domain data.

[0011] The present invention also provides a spacecraft cross-domain attitude estimation device based on frequency distribution alignment, comprising the following modules: A division module, used for dividing the frequency of the source domain data and the frequency of the target domain data into a plurality of frequency groups based on the first model; An alignment module, configured to perform local distribution alignment on the source domain data and the target domain data based on the multiple frequency groups, and obtain pseudo target domain data based on the locally distributed aligned data; A fine-tuning module, configured to perform supervised training fine-tuning on the first model based on the pseudo target domain data to obtain a second model; An estimation module, used for obtaining spacecraft attitude estimation results under different lighting scenarios based on target domain data and the second model; Among them, the first model is a deep learning model trained according to the source domain dataset.

[0012] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a spacecraft cross-domain attitude estimation method based on frequency distribution alignment as described in any one of the above-mentioned methods is implemented.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for estimating a spacecraft's cross-domain attitude based on frequency distribution alignment as described in any one of the above is implemented.

[0014] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-described methods for spacecraft cross-domain attitude estimation based on frequency distribution alignment.

[0015] The method and device for spacecraft cross-domain attitude estimation based on frequency distribution alignment provided by the present invention divide the frequencies of data in the source domain and the target domain into multiple frequency groups based on a first model, and perform local distribution alignment on each frequency group to obtain pseudo target domain data, and obtain a second model after fine-tuning the first model through supervised training based on the pseudo target domain data. Then, based on the second model and the target domain data, the spacecraft attitude estimation results under different lighting scenarios are obtained, thereby improving the generalization ability and estimation accuracy of the model in the target domain, especially in complex scenarios such as lighting changes, so that the attitude estimation of the spacecraft can be more accurate and stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A schematic flow chart of the spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention.

[0018] Figure 2 A schematic diagram of a flow chart of an embodiment of the present invention.

[0019] Figure 3 A schematic diagram of a process for determining low and high frequency division values ​​through a pre-training model provided by the present invention.

[0020] Figure 4 A schematic diagram of the statistical distribution of the pseudo target domain, source domain, and target domain data generated by the present invention.

[0021] Figure 5 This is a schematic diagram of the visualization of satellite images in different domains provided by the present invention.

[0022] Figure 6 A schematic structural diagram of a spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention.

[0023] Figure 7 This is a schematic structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] Figure 1 is a flow chart of a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, such as Figure 1 As shown, the method comprises the following steps: Step 100: Based on the first model, the frequencies of the source domain data and the target domain data are divided into a plurality of frequency groups.

[0026] Step 101: Based on multiple frequency groups, locally align source domain data with target domain data, and obtain pseudo target domain data based on the locally aligned data.

[0027] Step 102: Based on the pseudo target domain data, supervised training and fine-tuning are performed on the first model to obtain a second model.

[0028] Step 103: Based on the target domain data and the second model, the spacecraft attitude estimation results under different lighting scenes are obtained.

[0029] Among them, the first model is a deep learning model trained according to the source domain dataset.

[0030] Specifically, a deep learning model can be trained in advance according to a source domain dataset to obtain a first model that can perform spacecraft attitude estimation based on the source domain dataset.

[0031] In an embodiment of the present invention, the first model can be used to divide the frequency of source domain data and the frequency of target domain data into multiple frequency groups, wherein the source domain data refers to simulated synthetic data, and the target domain data refers to real data. This process can be achieved by converting the image data of the source domain and the target domain into the frequency domain through Fourier transform. Different frequency bands can represent different information of the image data. There are differences in the distribution of the images of the source domain and the target domain in different frequency bands. Therefore, multiple frequency groups can be divided according to the first model.

[0032] It should be noted that the “plurality” in each embodiment of the present invention refers to “two or more”, and this concept remains consistent throughout the text and will not be repeated later.

[0033] For each divided frequency group, the source domain data and the target domain data can be locally aligned. The process of local distribution alignment can use histogram alignment or other statistical methods, which is not limited in the embodiment of the present invention. The local distribution alignment process is to maintain the invariant features of the target in different domains, migrate the style of the source domain data and the style of the target domain data, and make the style of the source domain data closer to the target domain data. Based on the data after local distribution alignment, pseudo target domain data with labels that are closer to the target domain data distribution in the real scene can be obtained.

[0034] After obtaining the pseudo target domain data, the first model can be fine-tuned through supervised training based on the labeled pseudo target domain data set. This process is to adjust the parameters of the deep learning model to better adapt it to the pseudo target domain data, thereby improving the performance of the deep learning model in the target domain. After fine-tuning, the second model can be obtained.

[0035] After obtaining the second model, the second model can be self-trained on unlabeled target domain data to improve the accuracy of the model's six-degree-of-freedom attitude estimation in the target domain, and finally obtain the spacecraft attitude estimation results under different lighting scenarios.

[0036] The spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention divides the frequencies of data in the source domain and the target domain into multiple frequency groups based on a first model, and performs local distribution alignment on each frequency group to obtain pseudo target domain data, and obtains a second model after fine-tuning the first model through supervised training based on the pseudo target domain data, and then obtains spacecraft attitude estimation results under different lighting scenarios based on the second model and the target domain data, thereby improving the generalization ability and estimation accuracy of the model in the target domain, especially in complex scenarios such as lighting changes, so that the attitude estimation of the spacecraft can be more accurate and stable.

[0037] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, based on a first model, the frequency of source domain data and the frequency of target domain data are divided into multiple frequency groups, specifically including: Based on the first model, determining a low-frequency partition value and a high-frequency partition value of the entire frequency band of the source domain data and the target domain data; Based on the low-frequency division value and the high-frequency division value, the entire frequency band is divided into a low-frequency frequency group, a medium-frequency frequency group and a high-frequency frequency group; Based on the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group, a plurality of frequency groups of the source domain data and a plurality of frequency groups of the target domain data are determined.

[0038] Specifically, in the process of dividing the frequency of the source domain data and the frequency of the target data into multiple frequency groups according to the first model, the low-frequency division value and the high-frequency division value of the entire frequency band of the source domain data and the target domain data can be first determined according to the first model, thereby dividing the entire frequency band of the source domain data and the target domain data into low-frequency frequency groups, medium-frequency frequency groups and high-frequency frequency groups.

[0039] For example, determine the low-frequency band split value and the division value of the high frequency band , thereby dividing the entire frequency band of the source domain data and the target domain data into low ( ),middle( ),high( The low-frequency group represents domain-invariant features; the medium-frequency group represents undetermined features; and the high-frequency group represents variable features.

[0040] Then, multiple frequency groups of source domain data and multiple frequency groups of target domain data may be determined based on the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group.

[0041] In some implementations, the frequencies of the source domain data and the target domain data may be directly divided into a low-frequency frequency group, a medium-frequency frequency group, and a high-frequency frequency group.

[0042] According to a method for spacecraft cross-domain attitude estimation based on frequency distribution alignment provided by the present invention, multiple frequency groups of source domain data and multiple frequency groups of target domain data are determined based on low-frequency frequency groups, medium-frequency frequency groups and high-frequency frequency groups, specifically including: For each frequency group in the low frequency frequency group, the medium frequency frequency group, and the high frequency frequency group, based on information represented by each frequency group, determining a plurality of frequency subgroups of each frequency group; Based on the multiple frequency subgroups corresponding to the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group, multiple frequency groups of the source domain data and multiple frequency groups of the target domain data are determined.

[0043] Specifically, in some implementations, in order to achieve finer-grained local frequency band distribution alignment, for each frequency group in the low-frequency frequency group, the mid-frequency frequency group, and the high-frequency frequency group, multiple frequency subgroups of each frequency group can be determined based on the information represented by each frequency group.

[0044] In a preferred embodiment, for each frequency group, at most three frequency subgroups are determined.

[0045] For example, determine the low-frequency band split value and the division value of the high frequency band , dividing the entire frequency band into low ( ),middle( ),high( The low frequency band represents the domain-invariant features; the medium frequency band represents the undetermined features; and the high frequency band represents the variable features. Then, these three frequency bands are refined. Randomly select up to 2 frequency band division values ​​so that each frequency band group Can be divided into up to 3 local frequency subgroups , ,, The randomly selected partition value set can be repeated multiple times, and finally the optimal partition value set is determined based on the different information represented by each frequency band. , , .

[0046] According to this method, the low-frequency frequency group, the medium-frequency frequency group and the high-frequency frequency group each correspond to multiple frequency subgroups. By further subdividing each frequency group into multiple frequency subgroups, and then dividing the source domain data and the target domain data according to the multiple frequency subgroups of the low-frequency frequency group, the multiple frequency subgroups of the medium-frequency frequency group and the multiple frequency subgroups of the high-frequency group, the information distribution in each frequency segment can be analyzed more finely, which helps the model to more deeply understand and utilize the subtle feature differences in the data.

[0047] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, based on the information represented by each frequency group, multiple frequency subgroups of each frequency group are determined, specifically including: For the low-frequency frequency group, a plurality of frequency subgroups of the low-frequency frequency group are determined with the goal of making the distribution of the plurality of frequency subgroups of the source domain data close to the distribution of the plurality of frequency subgroups of the target domain data; For the intermediate frequency group, with the goal of maintaining the distribution of the multiple frequency subgroups of the source domain data and the multiple frequency subgroups of the target domain data within a preset range, determine the multiple frequency subgroups of the intermediate frequency group; For the high frequency group, a plurality of frequency subgroups of the high frequency group are determined with the goal that the distribution of the plurality of frequency subgroups of the source domain data and the plurality of frequency subgroups of the target domain data are not close.

[0048] Specifically, in an embodiment of the present invention, the low-frequency frequency group represents domain-invariant features, which usually contain the global structure and main shape information of the image. This information should be as consistent as possible between the source domain and the target domain to ensure that the model can recognize the same objects or features between different domains.

[0049] By making the frequency subgroup distributions of the low-frequency frequency groups in the source and target domains close, these key global features can be better aligned, resulting in more stable performance in cross-domain tasks. This helps reduce domain shift and enables the model to maintain good performance in different environments.

[0050] In some embodiments, for For all local frequency subgroups in the source domain, the distribution of each local frequency subgroup in the target domain should be very close to maintain the invariance of the target between different domains. Optimal local partition set , the following relationship can be satisfied:

[0051] in, represents a function that measures the similarity of two pixel distributions. In the embodiment of the present invention, Wasserstein distance can be used as the distribution similarity function. , is the threshold hyperparameter of the distribution information.

[0052] The intermediate frequency group represents undetermined features, which usually contain edges, contours, and some detail features. These features may have certain differences between the source domain and the target domain, but these differences should not exceed a reasonable range. The strategy of staying within the preset range can ensure that the model can adapt to certain changes without losing important detail information.

[0053] By allowing a certain range of distribution differences, the model is able to adapt to different domains while maintaining sensitivity to important features. This balance helps improve the model's generalization ability in complex environments.

[0054] In some embodiments, for For all local frequency subgroups in the source domain and the target domain, the distribution of each local frequency subgroup should be kept within a certain range, which helps to prevent the introduction of significant noise that may interfere with the target semantic information. Optimal local partition set , the following relationship can be satisfied:

[0055] in, represents a function that measures the similarity of two pixel distributions. In the embodiment of the present invention, Wasserstein distance can be used as the distribution similarity function. , is the threshold hyperparameter of the distribution information.

[0056] The high-frequency frequency group represents variable features, which include noise, small details, and fast-changing features in the image. These features tend to be highly domain-specific and differ significantly between the source and target domains. By making the distribution of high-frequency subgroups not close, the unique features in the source and target domains can be effectively distinguished.

[0057] By emphasizing the differences in high-frequency features between the source and target domains, the model can better capture the specificity between domains and reduce the impact of noise and unnecessary interference information on the model. This helps improve the accuracy and robustness of the model in the target domain, especially when facing high-complexity or high-noise data.

[0058] In some embodiments, for For all local frequency subgroups in , the distribution of each local frequency subgroup in the source domain and the target domain should be very different, which is used to achieve the style transfer between the source domain and the target domain. Optimal local partition set , the following relationship can be satisfied:

[0059] in, represents a function that measures the similarity of two pixel distributions. In the embodiment of the present invention, Wasserstein distance can be used as the distribution similarity function. , is the threshold hyperparameter of the distribution information.

[0060] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, based on a first model, a low-frequency division value and a high-frequency division value of the entire frequency band of source domain data and target domain data are determined, specifically including: The source domain satellite image and the target domain satellite image are converted into the frequency domain space by discrete Fourier transform; Selecting a plurality of different groups of initial low-frequency division values ​​and initial high-frequency division values, each group of different initial low-frequency division values ​​and initial high-frequency division values ​​corresponds to three different frequency groups; By using inverse discrete Fourier transform, the frequency sum of each frequency group is converted to the image space to obtain the image of each frequency group in the source domain and the image of each frequency group in the target domain; Based on the histogram alignment function, the images corresponding to each frequency group in the source domain and the images corresponding to each frequency group in the target domain are aligned to obtain the aligned images of each frequency group; Based on the aligned images of each frequency group and the loss function value of the first model, the low-frequency division value and the high-frequency division value of the entire frequency band of the source domain data and the target domain data are determined.

[0061] Specifically, in the process of determining the low-frequency division value and the high-frequency division value through the first model, the source domain satellite image and the target domain satellite image can first be converted into the frequency domain space through discrete Fourier transform, and the information in the image can be decomposed into different frequency components to facilitate independent analysis of each frequency band.

[0062] Then, multiple groups of different initial low-frequency division values ​​and initial high-frequency division values ​​are selected, each group of different initial low-frequency division values ​​and initial high-frequency division values ​​corresponds to three different frequency groups, and different initial division value combinations can produce different frequency group divisions. The embodiment of the present invention aims to explore the optimal division method.

[0063] After determining the three different frequency groups corresponding to each set of different initial low-frequency division values ​​and initial high-frequency division values, the frequency sum of each frequency group can be converted back to the image space through inverse discrete Fourier transform, thereby obtaining images corresponding to different frequency groups in the source domain and the target domain.

[0064] Then, the histogram alignment function can be used to align the images corresponding to each frequency group in the source domain and the target domain to obtain the aligned images of each frequency group. The loss function value of the first model is calculated based on the aligned image data. By analyzing the losses under different division values, the low-frequency division values ​​and high-frequency division values ​​that can minimize the losses are selected to determine the optimal low-frequency division values ​​and high-frequency division values.

[0065] For example, using the first model , consider synthetic data from a pair of satellites With real data , the optimal low- and high-frequency division values ​​are determined by using the loss value of the synthetic data after alignment of different frequency bands. and high frequency partition value The calculation solution is as follows:

[0066] in, is the loss function of satellite attitude estimation , and are the rotation matrix and translation labels, respectively. and are the predicted values ​​of the rotation matrix and translation, respectively. is the rotation error, is the translation error, which is defined as follows:

[0067] According to a spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the present invention, pseudo target domain data is obtained based on local distribution aligned data, specifically including: Divide multiple frequency groups in multiple different ways, and obtain multiple groups of different local distribution aligned data corresponding to the multiple different ways; Based on the pixel distribution metric function, the overall best locally distributed aligned data is selected from multiple groups of different distributed aligned data as the pseudo target domain data.

[0068] Specifically, the embodiment of the present invention has explained the process of dividing multiple frequency groups based on the first model to obtain a group of locally distributed aligned data. In order to overcome the problem that the local optimal solution is not necessarily the global optimal solution, multiple frequency groups can also be divided in other ways to obtain multiple groups of different locally distributed aligned data corresponding to other methods, such as multiple groups of different locally distributed aligned data can be obtained by random division.

[0069] After generating multiple sets of different distribution alignment data, the pixel distribution metric function can be used to evaluate the alignment effect of each set of data. The pixel distribution metric function is used to measure the similarity or consistency of the source domain and target domain data after alignment. By calculating these metrics, it can be determined which distribution-aligned datasets maintain the best consistency between the source domain and the target domain.

[0070] Finally, the selected overall optimal alignment data can be used as pseudo target domain data. The pseudo target domain data will be used as target domain data in subsequent model training or fine-tuning to further optimize the cross-domain performance of the model.

[0071] For example, repeat the random division into 3 frequency groups , and then divide the local optimal partition set into three frequency groups, so as to obtain the local optimal partition set Not unique, using pixel distribution metric function , to select the overall optimal distribution alignment data , as synthetic data Pseudo target domain data :

[0072] Different local distribution alignment data are generated through multiple frequency group division methods, and the optimal alignment result is selected through the pixel distribution metric function as pseudo target domain data, which can help the model learn the target domain features more accurately and improve the performance in cross-domain tasks. This strategy ensures that the model can find the optimal solution among multiple possible alignment methods, thereby maintaining consistent performance across different domains.

[0073] The method provided in this application is further explained below through embodiments in specific application scenarios.

[0074] Figure 2 A schematic diagram of the process of an embodiment provided by the present invention, such as Figure 2 As shown, it includes steps A10-A50: Step A10: In this embodiment, discrete Fourier transform Satellite images on the source and target domains All are converted to the frequency domain space, using the model trained on the simulated synthetic dataset (source domain) To determine the low frequency division value and the division value of the high frequency band , thus dividing the entire frequency band into low ( ),middle( ),high( ) frequency groups, representing domain-invariant, undetermined, and variable features, respectively.

[0075] In step A10, in order to better determine the different information represented by each frequency band of the image, this embodiment uses a model pre-trained on the synthetic data set , to determine the low and high frequency division values ​​of the image frequency. Figure 3 A schematic diagram of a process for determining low and high frequency partition values ​​by a pre-training model provided by the present invention, such as Figure 3 As shown, in some preferred embodiments, the low-frequency and high-frequency division method using the pre-trained model includes steps B10-B30: Step B10, in this embodiment, discrete Fourier transform Satellite images on the source and target domains Convert to frequency domain , Represents the total frequency. Select multiple groups of different initial low-frequency division values and high frequency partition value Each different set of initial low-frequency and high-frequency division values ​​can divide the entire frequency band into three frequency groups. For example: low-frequency group , IF Group , high frequency group , and remember , , Respectively represent the low-frequency division values ​​in this group With high frequency division value The frequency sum of each frequency group below.

[0076] Step B20, using inverse discrete Fourier transform The frequency sum of each frequency group in step B10 Convert to image space , and then use a histogram alignment function , which makes the distribution of the image corresponding to the source domain frequency band close to the distribution of the image corresponding to the target domain frequency band. Align the distribution of the source domain and the target domain corresponding frequency group images to obtain the aligned image .

[0077] Step B30, in order to better adaptively select the division values ​​of low and high frequency bands, with the help of the network model pre-trained on the synthetic data set , consider synthetic data from a pair of satellites With real data , the optimal low-frequency and high-frequency division values ​​are determined by using the loss value of the synthetic data after alignment of different frequency bands, and the entire frequency band of the image can be divided into a domain-invariant low-frequency group, an undetermined intermediate frequency group, and a domain-variable high-frequency group. and high frequency partition value The calculation solution is as follows:

[0078] in, is the loss function of satellite attitude estimation , and is the rotation matrix and the translation label, and is the predicted value of the rotation matrix and translation, are the rotation error and Translation error, which is defined as follows:

[0079] Step A20: In order to achieve finer-grained local frequency band distribution alignment, this embodiment refines the three groups of frequency bands. , Randomly select up to 2 frequency band division values ​​so that each frequency band group Can be divided into up to 3 local frequency subgroups , , , repeat the randomly selected partition value set many times, and finally determine the optimal partition value set by representing different information in each frequency band , , .

[0080] In step A20, in order to achieve finer-grained local frequency band distribution alignment, the three frequency band groups divided in step A10 are aligned. Perform local distribution alignment. Specifically, for each frequency group In this embodiment, at most two partition values ​​are randomly selected in this frequency group, so that each frequency group is divided into at most three local frequency subgroups. All the local frequency subgroups divided are recorded as The specific operation steps include steps C10-C30: Step C10, for The goal of this embodiment is to make the distribution of each local frequency subgroup of the synthetic data and the real data very close to each other, so as to maintain the invariance of the target between different domains. Optimal local partition set , the following relationship should be satisfied:

[0081] in, represents a function that measures the similarity of two pixel distributions. In the work of this embodiment, this embodiment uses Wasserstein distance as the distribution similarity function , is the threshold hyperparameter of the distribution information.

[0082] Step C20, Located between low frequency and high frequency, it contains ambiguous information. The goal of this embodiment is to keep the distribution of each local frequency subgroup of the synthetic data and the real data within a certain range. This helps to prevent the introduction of significant noise that may interfere with the target semantic information, so for Optimal local partition set , the following relationship should be satisfied:

[0083] is the threshold hyperparameter of the distribution information.

[0084] Step C30, by contrast, the high frequency band Mainly captures information about environmental noise and lighting. The goal of this embodiment is to make the distribution of each local frequency subgroup of the synthetic data and the real data very dissimilar, so as to achieve the style migration between the synthetic data and the real data. Optimal local partition set , the following relationship should be satisfied:

[0085] is the threshold hyperparameter of the distribution information.

[0086] Step A30, based on the optimal partition value set for each segment determined in step A20 , , , using the histogram distribution alignment function , the synthetic data With real data Each segment of the local distribution is aligned to generate the data after the local distribution alignment of each segment .

[0087] In step A30, based on the local optimal partition set obtained in step A20 This embodiment can maintain the invariant features of the target in different domains, migrate the style of the synthetic data with the style of the real data, and obtain the data after the local distribution of each segment is aligned. :

[0088] Here and .

[0089] Step A40, since each frequency band group in step A30 , the optimal partition value set determined This embodiment uses the pixel distribution metric function , to select the overall optimal distribution alignment data , as synthetic data Pseudo target domain data .

[0090] In step A40, in order to overcome the problem that the local optimum is not necessarily the global optimum solution, the three frequency groups can be randomly divided repeatedly in step A20. , so that the local optimal partition set obtained in step A30 is Not unique, using pixel distribution metric function , to select the overall optimal distribution alignment data , as synthetic data Pseudo target domain data :

[0091] Step A50: simulate the model trained on the synthetic dataset (source domain) The model is obtained by fine-tuning supervised training on the pseudo target domain data generated in step A40 , the fine-tuned model is used to fine-tune the real (target domain) data to obtain the model's estimation of the satellite attitude of the real data in the lighting scenario.

[0092] In step A50, the simulated synthetic dataset (source domain) On the trained model , is the label in the synthetic dataset. The pseudo target domain data generated in step A40 Fine-tune the model , using the fine-tuned model on real (target domain) data Perform self-training on it to get the converged model , and finally the model estimates the real data satellite attitude in this lighting scene .

[0093] The spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided in this embodiment is guided by dividing the entire frequency band into three parts: unchanged, uncertain, and variable. The synthetic data is aligned with the real data in terms of distribution, so that the style of the synthetic data is closer to the real data, and pseudo-target domain data is generated. The model pre-trained on the source domain is fine-tuned on the pseudo-target domain data set with labels, and finally the fine-tuned model is self-trained on the unlabeled real data to improve the accuracy of the model's six-degree-of-freedom attitude estimation on the real data.

[0094] In order to evaluate the effect of the spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided in this embodiment, the method of this embodiment is used to perform attitude estimation on light box and fluorescent satellite data, and the proposed method of this embodiment is compared with several recent state-of-the-art methods, including SPNv2, TangoUnchained and VPU.

[0095] Specifically, SPNv2 is a multi-scale, multi-task convolutional neural network (CNN) that utilizes a shared multi-scale feature encoder and multiple prediction heads to enhance the generalization ability of the model. TangoUnchained predicts the 2D bounding box around the target spacecraft and crops the image, adversarially regresses the 2D position, and iteratively solves the pose. VPU integrates 3D structure into the spacecraft pose estimation pipeline, enhancing robustness to significant illumination displacements between different domains.

[0096] Table 1 is a comparison table of various methods and this method. Figure 4 A schematic diagram of the statistical distribution of the pseudo target domain, source domain, and target domain data generated by the present invention. Figure 5 This is a schematic diagram of the visualization of satellite images in different domains provided by the present invention. Figure 4-5As shown in the figure, these experimental results clearly show that the algorithm of this embodiment has achieved advanced results in terms of the overall attitude estimation error of both the light box task and the fluorescent lamp task, and it is observed that the method proposed in this embodiment has achieved significant improvement in the accuracy of attitude estimation. This verifies the effectiveness of the frequency distribution alignment technology introduced in the method provided by this embodiment.

[0097] Table 1 Comparison results between various methods and this method

[0098] The spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention is described below. The spacecraft cross-domain attitude estimation device based on frequency distribution alignment described below and the spacecraft cross-domain attitude estimation method based on frequency distribution alignment described above can be referenced to each other.

[0099] Figure 6 A schematic diagram of the structure of a spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention is shown in FIG. Figure 6 As shown, the device includes the following modules: A division module 600, configured to divide the frequencies of the source domain data and the target domain data into a plurality of frequency groups based on the first model; An alignment module 610 is used to perform local distribution alignment on the source domain data and the target domain data based on multiple frequency groups, and obtain pseudo target domain data based on the locally distributed aligned data; A fine-tuning module 620 is used to perform supervised training fine-tuning on the first model based on pseudo target domain data to obtain a second model; An estimation module 630 is used to obtain spacecraft attitude estimation results under different lighting scenarios based on the target domain data and the second model; Among them, the first model is a deep learning model trained according to the source domain dataset.

[0100] According to a spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention, based on a first model, the frequency of source domain data and the frequency of target domain data are divided into multiple frequency groups, specifically including: Based on the first model, determining a low-frequency partition value and a high-frequency partition value of the entire frequency band of the source domain data and the target domain data; Based on the low-frequency division value and the high-frequency division value, the entire frequency band is divided into a low-frequency frequency group, a medium-frequency frequency group and a high-frequency frequency group; Based on the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group, a plurality of frequency groups of the source domain data and a plurality of frequency groups of the target domain data are determined.

[0101] According to a spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention, multiple frequency groups of source domain data and multiple frequency groups of target domain data are determined based on low-frequency frequency groups, medium-frequency frequency groups and high-frequency frequency groups, specifically including: For each frequency group in the low frequency frequency group, the medium frequency frequency group, and the high frequency frequency group, based on information represented by each frequency group, determining a plurality of frequency subgroups of each frequency group; Based on the multiple frequency subgroups corresponding to the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group, multiple frequency groups of the source domain data and multiple frequency groups of the target domain data are determined.

[0102] According to a spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention, based on the information represented by each frequency group, multiple frequency subgroups of each frequency group are determined, specifically including: For the low-frequency frequency group, a plurality of frequency subgroups of the low-frequency frequency group are determined with the goal of making the distribution of the plurality of frequency subgroups of the source domain data close to the distribution of the plurality of frequency subgroups of the target domain data; For the intermediate frequency group, with the goal of maintaining the distribution of the multiple frequency subgroups of the source domain data and the multiple frequency subgroups of the target domain data within a preset range, determine the multiple frequency subgroups of the intermediate frequency group; For the high frequency group, a plurality of frequency subgroups of the high frequency group are determined with the goal that the distribution of the plurality of frequency subgroups of the source domain data and the plurality of frequency subgroups of the target domain data are not close.

[0103] According to a spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention, based on a first model, a low-frequency division value and a high-frequency division value of the entire frequency band of source domain data and target domain data are determined, specifically including: The source domain satellite image and the target domain satellite image are converted into the frequency domain space by discrete Fourier transform; Selecting a plurality of different groups of initial low-frequency division values ​​and initial high-frequency division values, each group of different initial low-frequency division values ​​and initial high-frequency division values ​​corresponds to three different frequency groups; By using inverse discrete Fourier transform, the frequency sum of each frequency group is converted to the image space to obtain the image of each frequency group in the source domain and the image of each frequency group in the target domain; Based on the histogram alignment function, the images corresponding to each frequency group in the source domain and the images corresponding to each frequency group in the target domain are aligned to obtain the aligned images of each frequency group; Based on the aligned images of each frequency group and the loss function value of the first model, the low-frequency division value and the high-frequency division value of the entire frequency band of the source domain data and the target domain data are determined.

[0104] According to a spacecraft cross-domain attitude estimation device based on frequency distribution alignment provided by the present invention, pseudo target domain data is obtained based on local distribution aligned data, specifically including: Divide multiple frequency groups in multiple different ways, and obtain multiple groups of different local distribution aligned data corresponding to the multiple different ways; Based on the pixel distribution metric function, the overall best locally distributed aligned data is selected from multiple groups of different distributed aligned data as the pseudo target domain data.

[0105] Figure 7 A schematic diagram of the structure of an electronic device provided by the present invention, such as Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the spacecraft cross-domain attitude estimation method based on frequency distribution alignment, and the method includes: Based on the first model, the frequencies of the source domain data and the frequencies of the target domain data are divided into a plurality of frequency groups; Based on multiple frequency groups, the source domain data and the target domain data are locally distributed aligned, and pseudo target domain data is obtained based on the locally distributed aligned data; Based on the pseudo target domain data, supervised training and fine-tuning are performed on the first model to obtain the second model; Based on the target domain data and the second model, the spacecraft attitude estimation results under different lighting scenarios are obtained; Among them, the first model is a deep learning model trained according to the source domain dataset.

[0106] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0107] On the other hand, the present invention further provides a computer program product, the computer program product comprising a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can execute the spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the above methods, the method comprising: Based on the first model, the frequencies of the source domain data and the frequencies of the target domain data are divided into a plurality of frequency groups; Based on multiple frequency groups, the source domain data and the target domain data are locally distributed aligned, and pseudo target domain data is obtained based on the locally distributed aligned data; Based on the pseudo target domain data, supervised training and fine-tuning are performed on the first model to obtain the second model; Based on the target domain data and the second model, the spacecraft attitude estimation results under different lighting scenarios are obtained; Among them, the first model is a deep learning model trained according to the source domain dataset.

[0108] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the spacecraft cross-domain attitude estimation method based on frequency distribution alignment provided by the above methods, the method comprising: Based on the first model, the frequencies of the source domain data and the frequencies of the target domain data are divided into a plurality of frequency groups; Based on multiple frequency groups, the source domain data and the target domain data are locally distributed aligned, and pseudo target domain data is obtained based on the locally distributed aligned data; Based on the pseudo target domain data, supervised training and fine-tuning are performed on the first model to obtain the second model; Based on the target domain data and the second model, the spacecraft attitude estimation results under different lighting scenarios are obtained; Among them, the first model is a deep learning model trained according to the source domain dataset.

[0109] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0110] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A spacecraft cross-domain attitude estimation method based on frequency distribution alignment, characterized in that: include: Based on the first model, the frequencies of the source domain data and the frequencies of the target domain data are divided into a plurality of frequency groups; Based on the multiple frequency groups, locally aligning the source domain data with the target domain data, and obtaining pseudo target domain data based on the locally aligned data; Based on the pseudo target domain data, fine-tuning the first model through supervised training to obtain a second model; Based on the target domain data and the second model, obtaining spacecraft attitude estimation results under different lighting scenarios; Among them, the first model is a deep learning model trained according to the source domain dataset.

2. The spacecraft cross-domain attitude estimation method based on frequency distribution alignment according to claim 1 is characterized in that: Based on the first model, the frequencies of the source domain data and the target domain data are divided into multiple frequency groups, specifically including: Based on the first model, determining a low-frequency partition value and a high-frequency partition value of the entire frequency band of the source domain data and the target domain data; Based on the low frequency division value and the high frequency division value, the entire frequency band is divided into a low frequency group, a medium frequency group and a high frequency group; Based on the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group, a plurality of frequency groups of source domain data and a plurality of frequency groups of target domain data are determined.

3. The spacecraft cross-domain attitude estimation method based on frequency distribution alignment according to claim 2 is characterized in that: Determining a plurality of frequency groups of source domain data and a plurality of frequency groups of target domain data based on the low-frequency frequency group, the medium-frequency frequency group, and the high-frequency frequency group specifically includes: For each frequency group among the low frequency group, the medium frequency group and the high frequency group, based on information represented by each frequency group, determining a plurality of frequency subgroups of each frequency group; Based on the multiple frequency subgroups respectively corresponding to the low-frequency frequency group, the medium-frequency frequency group and the high-frequency frequency group, multiple frequency groups of source domain data and multiple frequency groups of target domain data are determined.

4. The spacecraft cross-domain attitude estimation method based on frequency distribution alignment according to claim 3 is characterized in that: Determining a plurality of frequency subgroups of each frequency group based on the information represented by each frequency group specifically includes: For the low-frequency frequency group, with the goal of making the distribution of multiple frequency subgroups of the source domain data close to that of multiple frequency subgroups of the target domain data, determine multiple frequency subgroups of the low-frequency frequency group; For the intermediate frequency group, with the goal of maintaining the distribution of multiple frequency subgroups of source domain data and multiple frequency subgroups of target domain data within a preset range, determine multiple frequency subgroups of the intermediate frequency group; For the high frequency group, a plurality of frequency subgroups of the high frequency group are determined with the goal that the distribution of the plurality of frequency subgroups of the source domain data and the plurality of frequency subgroups of the target domain data are not close.

5. The spacecraft cross-domain attitude estimation method based on frequency distribution alignment according to claim 2 is characterized in that: Based on the first model, determining the low-frequency division value and the high-frequency division value of the entire frequency band of the source domain data and the target domain data specifically includes: The source domain satellite image and the target domain satellite image are converted into the frequency domain space by discrete Fourier transform; Selecting a plurality of different groups of initial low-frequency division values ​​and initial high-frequency division values, each group of different initial low-frequency division values ​​and initial high-frequency division values ​​corresponds to three different frequency groups; By inverse discrete Fourier transform, the frequency sum of each frequency group is converted into an image space to obtain an image corresponding to each frequency group in a source domain and an image corresponding to each frequency group in a target domain; Based on a histogram alignment function, aligning the images of the source domain corresponding to the respective frequency groups with the images of the target domain corresponding to the respective frequency groups to obtain aligned images of the respective frequency groups; Based on the aligned images of the frequency groups and the loss function value of the first model, the low-frequency division value and the high-frequency division value of the entire frequency band of the source domain data and the target domain data are determined.

6. The spacecraft cross-domain attitude estimation method based on frequency distribution alignment according to any one of claims 1 to 5, characterized in that: Based on the data of local distribution alignment, the pseudo target domain data is obtained, including: Dividing a plurality of frequency groups in a plurality of different ways, and obtaining a plurality of groups of different locally distributed aligned data corresponding to the plurality of different ways; Based on the pixel distribution metric function, the overall optimal locally distributed aligned data is selected from the multiple groups of different distributed aligned data as the pseudo target domain data.

7. A spacecraft cross-domain attitude estimation device based on frequency distribution alignment, characterized in that: include: A division module, used for dividing the frequency of the source domain data and the frequency of the target domain data into a plurality of frequency groups based on the first model; An alignment module, configured to perform local distribution alignment on the source domain data and the target domain data based on the multiple frequency groups, and obtain pseudo target domain data based on the locally distributed aligned data; A fine-tuning module, configured to perform supervised training fine-tuning on the first model based on the pseudo target domain data to obtain a second model; An estimation module, used for obtaining spacecraft attitude estimation results under different lighting scenarios based on target domain data and the second model; Among them, the first model is a deep learning model trained according to the source domain dataset.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for spacecraft cross-domain attitude estimation based on frequency distribution alignment as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for spacecraft cross-domain attitude estimation based on frequency distribution alignment as described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for spacecraft cross-domain attitude estimation based on frequency distribution alignment as described in any one of claims 1 to 6 is implemented.

Images