Orientation method and device based on atmospheric polarization mode learning network

Abstract

The invention relates to the technical field of navigation, and provides an orientation method and device based on an atmospheric polarization mode learning network. By constructing a polarization mode learning network comprising two parallel feature extraction structures which are consistent in structure, feature extraction is performed on a polarization angle image and a polarization degree image by using the same feature extraction structure, and each branch can concentrate on processing polarization features of corresponding types while feature extraction logic is unified. A polarization mode learning network is trained through a data set comprising polarization image samples under multi-period and multi-weather conditions and corresponding target orientation parameters, data enhancement is carried out through polarization angle periodic processing, atmospheric attenuation simulation and regional disturbance in the training process, and a polarization mode is obtained through a data driving mode. A nonlinear mapping relation between a complex polarization mode and direction information is mined, so that the trained polarization mode learning network can output an accurate, stable and high-robustness orientation result under various meteorological conditions.

Description

Technical Field

[0001] This application relates to the field of navigation technology, and in particular to a orientation method and apparatus based on an atmospheric polarization mode learning network. Background Technology

[0002] In the natural environment, many organisms possess the ability to sense polarized light from the sky and use it for navigation. This phenomenon has spurred the research and development of biomimetic polarized light navigation technology. Traditional polarization navigation methods are mainly based on Rayleigh scattering models or the principle of atmospheric polarization pattern symmetry, achieving high accuracy under clear weather conditions. However, under non-ideal meteorological conditions such as cloud cover and haze, non-Rayleigh scattering significantly affects the polarization light distribution pattern, leading to a substantial decrease in the accuracy of traditional orientation methods or even complete failure. Related research indicates that atmospheric clouds cause disordered polarization information distribution patterns and attenuation of sunlight intensity, significantly reducing the accuracy of traditional orientation methods. Even attempts to directly observe the sun's position are often difficult to achieve due to cloud cover or interference from scattered light, posing a significant challenge to the autonomous navigation of unmanned platforms in environments with limited satellite signals. Summary of the Invention

[0003] Therefore, it is necessary to provide a orientation method and apparatus based on an atmospheric polarization mode learning network to address the aforementioned technical problems.

[0004] A orientation method based on an atmospheric polarization mode learning network includes the following steps: Acquire a polarization image of the sky region, the polarization image including a polarization angle image and a polarization degree image; A dataset containing polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions is constructed; a polarization pattern learning network is constructed, which includes two parallel feature extraction structures with consistent structures, which extract features from polarization angle images and polarization degree images respectively, and then fuse the extracted features. The polarization mode learning network was trained using the constructed dataset, and data augmentation was performed during the training process through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbation. Using the acquired polarization image information as input, the polarization pattern learning network, after training, outputs orientation parameters.

[0005] In one embodiment, the dataset is constructed according to the following steps: Polarization image samples were collected at different times and under different weather conditions in different geographical locations, and the target orientation parameters corresponding to the polarization image samples were recorded simultaneously. Based on atmospheric optical transmission theory and cloud scattering model, polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions are generated through high-fidelity numerical simulation. A dataset is constructed using the collected and generated polarization image samples and their corresponding target orientation parameters.

[0006] In one embodiment, the target orientation parameters include the solar position vector and the vehicle heading angle.

[0007] In one embodiment, the periodic polarization angle processing includes: periodically parallel shifting the polarization angle image, adding random angle perturbations, or adding Gaussian white noise.

[0008] In one embodiment, the atmospheric attenuation simulation includes: scaling the polarization image by light intensity, adjusting the contrast, or injecting simulated noise.

[0009] In one embodiment, the regional perturbation includes: randomly selecting regions of the polarized image and applying a preset level of Gaussian noise or salt-and-pepper noise.

[0010] In one embodiment, the loss function used during training is:

[0011] in, The loss function; The solar azimuth angle is calculated from the global information of the polarization image; For the first block The solar azimuth angle obtained by solving local information from a polarization image; k This represents the number of local information elements in the polarization image.

[0012] In one embodiment, using the acquired polarization image information as input, and outputting orientation parameters through a trained polarization pattern learning network, the method further includes: The acquired polarization image information is processed into a parallel tensor form.

[0013] In one embodiment, the feature extraction structure consists of a 5×5 convolutional layer, an average pooling layer, a Swish activation function layer, a 5×5 convolutional layer, an average pooling layer, a Swish activation function layer, and a local connection layer connected in sequence.

[0014] A direction-oriented device based on an atmospheric polarization mode learning network, comprising: The data acquisition module is used to acquire polarization images of the sky region, including polarization angle images and polarization degree images; The network construction module is used to construct a dataset containing polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions; and to construct a polarization pattern learning network, which includes two parallel feature extraction structures with consistent structures, to extract features from polarization angle images and polarization degree images respectively, and to fuse the extracted features. The network training module is used to train the polarization mode learning network with the constructed dataset. During the training process, data augmentation is performed through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbation. The localization and resolution module uses the acquired polarization image information as input and outputs orientation parameters through a trained polarization pattern learning network.

[0015] The aforementioned orientation method and apparatus based on an atmospheric polarization pattern learning network constructs two parallel polarization pattern learning networks with consistent feature extraction structures. These networks extract features from polarization angle and degree of polarization images using similar structures, allowing each branch to focus on processing its corresponding polarization feature while maintaining a unified feature extraction logic. The polarization pattern learning network is trained using a dataset containing polarization image samples and corresponding target orientation parameters across multiple time periods and meteorological conditions. During training, data augmentation is performed through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbations. This data-driven approach uncovers the nonlinear mapping relationship between complex polarization patterns and orientation information, enabling the trained polarization pattern learning network to output accurate, stable, and highly robust orientation results under various meteorological conditions.

[0016] This invention can significantly improve the robustness and practicality of polarization navigation systems in non-ideal environments, and provide reliable heading support for unmanned systems in environments where satellite signals are limited. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a orientation method based on an atmospheric polarization mode learning network in one embodiment. Figure 2 This is a schematic diagram of the structure of a polarization mode learning network in one embodiment; Figure 3 This is a structural block diagram of a orientation device based on an atmospheric polarization mode learning network in one embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The orientation method based on atmospheric polarization mode learning network provided in this application can be applied to polarization navigation systems.

[0020] In one embodiment, such as Figure 1 As shown, a orientation method based on an atmospheric polarization mode learning network is provided, including the following steps: Step S201: Obtain a polarization image of the sky region, the polarization image including a polarization angle image and a polarization degree image.

[0021] It should be noted that polarized light information can be acquired using polarization imaging equipment equipped with a fisheye lens or other wide-angle optical system. By analyzing and processing the acquired polarized light information, the angle of polarization (AOP) and degree of polarization (DOP) images can be obtained. The angle of polarization image characterizes the spatial distribution of the polarization direction of light rays at each pixel, while the degree of polarization image reflects the intensity distribution of light polarization. By using the angle of polarization and degree of polarization images as input to the network, the complete characteristics of the sky's polarization field can be described together.

[0022] Step S202: Construct a dataset containing polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions; construct a polarization pattern learning network, which includes two parallel feature extraction structures with consistent structures, to extract features from the polarization angle image and the polarization degree image respectively, and fuse the extracted features.

[0023] Understandably, the constructed dataset, which includes polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions, can simulate various weather conditions.

[0024] Step S203: The polarization mode learning network is trained using the constructed dataset. During the training process, data augmentation is performed through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbation.

[0025] Step S204: Using the acquired polarization image information as input, the polarization pattern learning network, after training, outputs orientation parameters.

[0026] In the aforementioned orientation method based on an atmospheric polarization pattern learning network, a polarization pattern learning network with two parallel, identical feature extraction structures is constructed. These networks extract features from polarization angle and degree of polarization images using similar structures, allowing each branch to focus on processing its corresponding polarization feature while maintaining a unified feature extraction logic. The polarization pattern learning network is trained using a dataset containing polarization image samples from multiple time periods and meteorological conditions, along with corresponding target orientation parameters. During training, data augmentation is performed through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbations. This data-driven approach uncovers the nonlinear mapping relationship between complex polarization patterns and orientation information, enabling the trained polarization pattern learning network to output accurate, stable, and robust orientation results under various meteorological conditions. By combining polarization angle and degree of polarization image information in the output of orientation parameters, the robustness, accuracy, and environmental adaptability of the positioning are improved. In one embodiment, the dataset is constructed according to the following steps: Polarization image samples were collected at different times and under different weather conditions in different geographical locations, and the target orientation parameters corresponding to the polarization image samples were recorded simultaneously. Based on atmospheric optical transmission theory and cloud scattering model, polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions are generated through high-fidelity numerical simulation. A dataset is constructed using the collected and generated polarization image samples and their corresponding target orientation parameters.

[0027] It should be noted that "multiple time periods" includes different seasons and specific times. "Multiple weather conditions" includes sunny days, thin clouds, thick clouds, overcast days, and smog.

[0028] In this embodiment, by constructing a dataset containing real environmental observation data and data generated based on cloud scattering model simulation, the complementary nature of the two types of data allows for a wider coverage of sky polarization patterns, resulting in a polarization pattern learning network trained using this dataset exhibiting higher accuracy and robustness.

[0029] In one embodiment, the target orientation parameters include the solar position vector and the vehicle heading angle.

[0030] In one embodiment, the periodic polarization angle processing includes: periodically shifting the polarization angle image, adding random angle perturbations, or adding Gaussian white noise.

[0031] Based on the unique periodicity of polarization angle images, data augmentation can be performed by periodic parallel shifting, adding random angle perturbations, or adding Gaussian white noise.

[0032] In one embodiment, the atmospheric attenuation simulation includes: scaling the polarization image by light intensity, adjusting the contrast, or injecting simulated noise.

[0033] By scaling the intensity of the polarization image, adjusting the contrast, or injecting simulated noise, the weakening of the polarization signal caused by cloud cover can be simulated.

[0034] In one embodiment, the regional perturbation includes: randomly selecting regions of the polarized image and applying a preset level of Gaussian noise or salt-and-pepper noise.

[0035] By randomly selecting regions of a polarized image and applying a preset level of Gaussian noise or salt-and-pepper noise, non-uniform atmospheric conditions such as local clouds and fog can be simulated.

[0036] During training, data augmentation is performed through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbations, thereby improving the adaptability of the polarization model learning network to polarization model changes under complex weather conditions.

[0037] In one embodiment, the loss function used during training is:

[0038] in, The loss function; The solar azimuth angle is calculated from the global information of the polarization image; For the first block The solar azimuth angle obtained by solving local information from a polarization image; k This represents the number of local information elements in the polarization image.

[0039] In this embodiment, the loss function uses different local information of the polarization image to constrain the same solar azimuth angle calculation result, which can significantly improve the accuracy and robustness of the network.

[0040] In one embodiment, step S204, using the acquired polarization image information as input, and outputting orientation parameters through a trained polarization mode learning network, further includes: The acquired polarization image information is processed into a parallel tensor form.

[0041] In this embodiment, the acquired polarization image information is processed into a parallel tensor form, and the information of polarization angle and polarization degree is used to help remove the ambiguity of traditional orientation based on polarization angle information, while enhancing the adaptability to different weather conditions at different times.

[0042] In one embodiment, reference Figure 2The feature extraction structure consists of a 5×5 convolutional layer, an average pooling layer, a Swish activation function layer, a 5×5 convolutional layer, an average pooling layer, a Swish activation function layer, and a local connection layer connected in sequence.

[0043] from Figure 2 As can be seen, the polarization angle image and polarization degree image are respectively input into two parallel feature extraction structures. First, local features are extracted through a 5×5 convolutional layer, then downsampled through an average pooling layer to simplify the features, and nonlinearity is introduced through a Swish activation function layer. After that, it goes through another 5×5 convolutional layer, average pooling layer, and Swish activation function layer to further extract depth features. Then, the features are initially integrated through a local connection layer. The features extracted by the two structures are concatenated and fed into a fully connected layer to calculate the orientation parameters.

[0044] It should be noted that the calculated orientation parameters include the solar position vector and the vehicle heading angle.

[0045] In this embodiment, local features are extracted through a 5×5 convolutional layer, which can extract texture and directional features while maintaining low parameter counts in polarization angle and degree images, thus simulating the "spatial perception" mechanism in biological systems. On the other hand, Swish is smoother and differentiable than ReLU during training, which is conducive to gradient propagation and adapts to the extraction of local gradient details during training of large input images.

[0046] To verify the effectiveness of this invention, tests were conducted on test sets under different meteorological conditions, using networks and datasets of different resolutions. The angular error of the orientation results was used as the evaluation index. Experiments were conducted using this invention to verify the accuracy advantage of this method under complex weather conditions.

[0047] The experimental results are shown in Table 1. The results show that the trained neural network can achieve an accuracy of 84.58% and 86.25% respectively on test sets with angular resolutions of 0.5° and 1° that simulate real-world environments. In particular, it demonstrates excellent generalization ability when faced with angular labels that change periodically.

[0048] Table 1 Experimental Results

[0049] It should be understood that, although Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are performed; they can be executed in other orders. Furthermore, Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0050] In one embodiment, such as Figure 3 As shown, a direction-oriented device based on an atmospheric polarization mode learning network is provided, comprising: The data acquisition module 901 is used to acquire polarization images of the sky region, the polarization images including polarization angle images and polarization degree images.

[0051] The network construction module 902 is used to construct a dataset containing polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions; and to construct a polarization pattern learning network, which includes two parallel feature extraction structures with consistent structures, to extract features from the polarization angle image and the polarization degree image respectively, and to fuse the extracted features.

[0052] The network training module 903 is used to train the polarization mode learning network with the constructed dataset, and to perform data augmentation during the training process through periodic polarization angle processing, atmospheric attenuation simulation and regional perturbation.

[0053] The localization and resolution module 904 is used as input to obtain polarization image information and outputs orientation parameters through a trained polarization pattern learning network.

[0054] Specific limitations regarding the orientation device based on atmospheric polarization mode learning networks can be found in the limitations of the orientation method based on atmospheric polarization mode learning networks mentioned above, and will not be repeated here. Each module in the aforementioned orientation device based on atmospheric polarization mode learning networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0055] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0056] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A orientation method based on an atmospheric polarization mode learning network, characterized in that, Includes the following steps: Acquire a polarization image of the sky region, the polarization image including a polarization angle image and a polarization degree image; A dataset containing polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions is constructed; a polarization pattern learning network is constructed, which includes two parallel feature extraction structures with consistent structures, which extract features from polarization angle images and polarization degree images respectively, and then fuse the extracted features. The polarization mode learning network was trained using the constructed dataset, and data augmentation was performed during the training process through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbation. Using the acquired polarization image information as input, the polarization pattern learning network, after training, outputs orientation parameters.

2. The orientation method based on an atmospheric polarization mode learning network according to claim 1, characterized in that, The dataset is constructed according to the following steps: Polarization image samples were collected at different times and under different weather conditions in different geographical locations, and the target orientation parameters corresponding to the polarization image samples were recorded simultaneously. Based on atmospheric optical transmission theory and cloud scattering model, polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions are generated through high-fidelity numerical simulation. A dataset is constructed using the collected and generated polarization image samples and their corresponding target orientation parameters.

3. The orientation method based on an atmospheric polarization mode learning network according to claim 2, characterized in that, The target orientation parameters include the solar position vector and the vehicle heading angle.

4. The orientation method based on an atmospheric polarization mode learning network according to claim 1, characterized in that, The periodic polarization angle processing includes: periodically parallel shifting the polarization angle image, adding random angle perturbation, or adding Gaussian white noise.

5. The orientation method based on an atmospheric polarization mode learning network according to claim 1, characterized in that, The atmospheric attenuation simulation includes: scaling the polarization image by light intensity, adjusting the contrast, or injecting simulated noise.

6. The orientation method based on an atmospheric polarization mode learning network according to claim 1, characterized in that, The regional perturbation includes: randomly selecting regions of the polarized image and applying a preset level of Gaussian noise or salt-and-pepper noise.

7. The orientation method based on an atmospheric polarization mode learning network according to claim 1, characterized in that, The loss function used during training is: in, The loss function; The solar azimuth angle is calculated from the global information of the polarization image; For the first block The solar azimuth angle obtained by solving local information from a polarization image; k This represents the number of local information elements in the polarization image.

8. The orientation method based on an atmospheric polarization mode learning network according to claim 1, characterized in that, Using the acquired polarization image information as input, the polarization pattern learning network, after training, outputs orientation parameters, which also include: The acquired polarization image information is processed into a parallel tensor form.

9. The orientation method based on an atmospheric polarization mode learning network according to claim 1, characterized in that, The feature extraction structure consists of a 5×5 convolutional layer, an average pooling layer, a Swish activation function layer, a 5×5 convolutional layer, an average pooling layer, a Swish activation function layer, and a local connection layer connected in sequence.

10. A direction-oriented device based on an atmospheric polarization mode learning network, characterized in that, include: The data acquisition module is used to acquire polarization images of the sky region, including polarization angle images and polarization degree images; The network construction module is used to construct a dataset containing polarization image samples and corresponding target orientation parameters under multiple time periods and meteorological conditions; and to construct a polarization pattern learning network, which includes two parallel feature extraction structures with consistent structures, to extract features from polarization angle images and polarization degree images respectively, and to fuse the extracted features. The network training module is used to train the polarization mode learning network with the constructed dataset. During the training process, data augmentation is performed through periodic polarization angle processing, atmospheric attenuation simulation, and regional perturbation. The localization and resolution module uses the acquired polarization image information as input and outputs orientation parameters through a trained polarization pattern learning network.

Images