A polarization orientation method based on the antisymmetry of atmospheric polarization patterns

Abstract

This invention discloses a polarization orientation method based on the antisymmetry of the atmospheric polarization pattern. First, a UNet convolutional neural network is used to perform binary segmentation on the polarization image of the sky acquired by a polarization camera to obtain a de-obstructed polarization image of the sky. The separated sky is then divided into multiple circular rings centered at the zenith, and the corresponding positive and negative feature points within each ring are obtained. The antisymmetry of the atmospheric polarization pattern is used to fit the solar meridian, ultimately completing the calculation of the heading angle. This invention solves the problem of traditional polarization orientation methods failing due to complex weather conditions, especially large-scale obstructions, and effectively improves the robustness of polarization orientation methods under complex weather conditions.

Description

Technical Field

[0001] The invention relates to modern information image processing and bionic navigation, and in particular to a polarization orientation method based on the antisymmetry of atmospheric polarization patterns. Background Art

[0002] Polarization orientation is a navigation method developed based on the foraging behavior of insects. Biological research has shown that insects such as sand ants, monarch butterflies, and bees can autonomously orient themselves by sensing the atmospheric polarization pattern in the sky. Atmospheric polarization pattern refers to the specific polarization state distribution in the sky formed by polarized light generated by the scattering of atmospheric particles during the transmission of sunlight. Vehicle orientation can be achieved by extracting the solar meridian information from the atmospheric polarization pattern. Compared with traditional satellite navigation and inertial navigation, the bionic polarization orientation method is not subject to electromagnetic interference and has no cumulative error. It has broad application prospects in the field of autonomous navigation and has therefore been widely studied.

[0003] However, while atmospheric polarization-based navigation offers high accuracy, long flight time, and strong autonomy, it can also produce varying degrees of distortion depending on weather conditions. In adverse conditions, such as thick cloud cover or obstructions, polarization information in parts of the imaging field of view becomes invalid, severely reducing polarization-based navigation accuracy and even rendering the compass unusable. This high dependence on weather conditions limits the application of atmospheric polarization-based navigation and navigation. Summary of the Invention

[0004] Purpose of the invention: In order to solve the problem that the navigation orientation method based on the atmospheric polarization pattern in the existing technology fails to orient in complex weather conditions, this paper provides a polarization orientation method based on the antisymmetry of the atmospheric polarization pattern.

[0005] Technical solution: A polarization orientation method based on the antisymmetry of atmospheric polarization patterns, comprising the following steps:

[0006] Step 1: Use a polarization camera to obtain an original sky polarization image, use the UNet convolutional neural network to perform binary segmentation on the sky polarization image to obtain a de-obstructed sky polarization image, and calculate the polarization angle of each pixel in the image;

[0007] Step 2: In the de-obstructed polarized sky image, take the zenith as the center of the circle and randomly set the axis so that the axis passes through the circle point;

[0008] Step 3: Using the axis as a reference, divide the sky into eight regions at 45° intervals, named 1 to 8. Regions 1, 4, 5, and 8 are designated as valid regions, while regions 2, 3, 6, and 7 are designated as interference regions. Separate the polarization image of the sky into multiple equal-width circular rings. For each circular ring, calculate the following parameters based on the polarization angle of each pixel:

[0009] The number of positive feature points P_num, negative feature points N_num and total feature points T_num in the effective area of ​​the ring, the number of positive feature points P_score in areas 1 and 5, the number of negative feature points N_score in areas 4 and 8 and the total number of feature points I_num in the interference area;

[0010] Calculate the δ value of each ring area, and calculate the total value δ from the δ value of each ring area T ;

[0011]

[0012] Among them, w i is the weight corresponding to each ring;

[0013] Step 4: Rotate the axis by an angle θ in a fixed direction, repeat step 3, and calculate the total value;

[0014] Step 5: Repeat step 4 until the axis rotates one circle, compare all the total values, and take the axis position when the total value is the smallest as the position of the solar meridian;

[0015] Step 6: Use the carrier's position and time information to calculate the sun's position and the absolute heading angle.

[0016] Furthermore, in step 1, the method for obtaining the de-obstructed sky polarization image is as follows:

[0017] Use the dataset to train the UNet convolutional neural network and use the trained UNet network to obtain the mask;

[0018] The original sky polarization image is preprocessed using an image opening operation to remove abnormal small areas in the original sky polarization image to obtain a preprocessed sky polarization image.

[0019] The mask is multiplied with the preprocessed sky polarization image to obtain the de-occluded sky polarization image.

[0020] Furthermore, in step 1, the polarization angle of each pixel in the image is calculated using the Stokes formula, specifically:

[0021]

[0022] Where Aop is the polarization angle of each pixel in the image, I(0), I(45), I(90), and I(135) are the intensity values ​​of the polarization directions of the pixel at 0°, 45°, 90°, and 135°, respectively.

[0023] Furthermore, in step 3, the total feature points include positive feature points and negative feature points, and the feature points are obtained by:

[0024] Set a threshold α to expand the polarization angle range and perform benchmark transformation. Based on the Aop value obtained in step 1, select pixels whose polarization angle absolute value is greater than α as feature points. Pixels with positive polarization angles are positive feature points, and pixels with negative polarization angles are negative feature points.

[0025] Furthermore, in step 1, the network structure of the UNet convolutional neural network is a U-shaped structure with four layers of downsampling and four layers of upsampling.

[0026] Furthermore, in step 6, the solar azimuth angle under the load system is defined Solar azimuth under the navigation system Calculate the absolute heading angle

[0027] θ s =arcsin(sinεsinβ+cosεCosβcos t)

[0028]

[0029] Among them, θ s is the solar altitude angle, ε is the latitude of the observation point, β is the solar declination angle, and t is the solar hour angle.

[0030] Beneficial effect: Compared with the existing polarization light orientation method, the polarization orientation method provided by the present invention is based on the antisymmetry of the atmospheric polarization mode. First, the polarization image is binary segmented through the UNet convolutional neural network to separate and remove the thick cloud / building occlusion area in the image, thereby improving the ability to resist weather interference during atmospheric polarization light orientation; then the solar meridian is fitted by utilizing the antisymmetric distribution characteristic of the polarization angle distribution along the solar meridian under the atmospheric polarization mode, thereby realizing accurate orientation of the polarization light system under complex weather conditions, so that even in complex weather, atmospheric polarization light orientation can still be effectively realized, thereby alleviating the problem of significant decrease in orientation accuracy or even failure due to weather influence during atmospheric polarization light orientation, breaking through the limitation of atmospheric polarization light orientation being highly dependent on weather conditions, and effectively improving the robustness of the polarization orientation method under complex weather conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1is a flow chart of the polarization orientation method based on the antisymmetry of the atmospheric polarization pattern of the present invention;

[0032] Figure 2 It is the UNet neural network structure diagram used in the present invention;

[0033] Figure 3 The eight sky areas are divided clockwise along the axis in this embodiment;

[0034] Figure 4 is the polarization diagram of the effective sky area and the interference area divided in this embodiment;

[0035] Figure 5 Schematic diagram of sky ring segmentation statistics used in this embodiment;

[0036] Figure 6 This is a comparison of the meridian fitting results of this embodiment under cloudy weather;

[0037] Figure 7 This is a comparison of the meridian fitting results of this embodiment under cloudy and obstructed weather conditions;

[0038] Figure 8 This is a comparison of the meridian fitting results of this embodiment under extreme shading weather conditions. DETAILED DESCRIPTION

[0039] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0040] A polarization orientation method based on the antisymmetry of atmospheric polarization patterns, such as Figure 1 As shown, the following steps are included:

[0041] Step 1: Use a polarization camera to obtain an original sky polarization image. Use the UNet convolutional neural network to perform binary segmentation on the sky polarization image to separate thick clouds, occlusions, and the sky, and obtain a de-occluded sky polarization image. The specific method is as follows:

[0042] like Figure 2 As shown in the figure, a UNet convolutional neural network with a four-layer downsampling and four-layer upsampling U-shaped structure was selected. Using polarized images with occluders as samples, the sky and occluders were manually segmented and labeled. 80% of the samples and labels were used as the training set, and 20% was used as the validation set. Together, the training and validation sets comprised the dataset. The UNet convolutional neural network was trained using this dataset to separate the sky from thick clouds and occluders in the image, generating a mask.

[0043] The original sky polarization image is preprocessed using an image opening operation to remove abnormal small areas in the original sky polarization image to obtain a preprocessed sky polarization image.

[0044] Multiply the mask with the pre-processed sky polarization image to obtain the de-occluded sky polarization image;

[0045] The AOP angle image is calculated using the Stokes formula, specifically:

[0046]

[0047] Where Aop is the polarization angle of each pixel in the image, and I(0), I(45), I(90), and I(135) are the intensity values ​​of the polarization directions of 0°, 45°, 90°, and 135°, respectively.

[0048] The calculation process of the polarization angle is as follows:

[0049] The polarization state of light is described by the Mueller matrix, which is expressed as:

[0050]

[0051] Among them, S out is the Stokes vector of the outgoing light, S in is the Stokes vector of the incident light, M is the Mueller matrix, and ψ is the angle between the transmission axis of the polarized light and the reference direction.

[0052] Then using the above Mueller matrix we can get:

[0053]

[0054] Where I(ψ) is the intensity of the outgoing light when the polarized light transmission axis and the reference direction are ψ. In the process of solving I0, Q0, and U0, it is only necessary to know two or more ψ angles and their corresponding I(ψ). Assuming the ψ values ​​are 0°, 45°, 90°, and 135°, we can obtain:

[0055]

[0056] Calculate the AOP angle image using images I0, Q0, and U0:

[0057]

[0058] Where I(0), I(45), I(90), and I(135) are the intensity values ​​in the polarization directions of 0°, 45°, 90°, and 135°.

[0059] Step 2: In the de-obstructed polarized sky image, take the zenith as the center of the circle and randomly set the axis so that the axis passes through the circle point;

[0060] Step 3: If Figure 3As shown in the figure, the sky is divided into 8 areas with the axis as the reference and each area is divided into 45° intervals, and named as areas 1 to 8. Among them, areas 1, 4, 5, and 8 are set as effective areas, and areas 2, 3, 6, and 7 are set as interference areas. Figure 4 As shown; the sky polarization image is separated into n rings with a width of r, as shown Figure 3 ; For each ring, the following parameters are counted:

[0061] The number of positive feature points P_num, negative feature points N_num and total feature points T_num in the effective area of ​​the ring, the number of positive feature points P_score in areas 1 and 5, the number of negative feature points N_score in areas 4 and 8 and the total number of feature points I_num in the interference area;

[0062] Calculate the δ value of each ring area. The calculation formula is as follows, and calculate the total value δ from the δ value of each ring area. T ;

[0063]

[0064] Among them, w i is the weight corresponding to each ring, w i From the calculation formula, it can be seen that when the difference in the number of positive and negative feature points in a circle is too large, the weight of the circle will be reduced. By adaptively adjusting the weight, the error caused by the occlusion can be controlled. Figure 5 It is a statistical diagram of segmented rings.

[0065] Among them, the total feature points include positive feature points and negative feature points, and the method for obtaining feature points is:

[0066] A threshold α is set to expand the polarization angle range and perform benchmark transformation, and pixels with polarization angles greater than α are selected as feature points. Feature points with positive polarization angles are positive feature points, and feature points with negative polarization angles are negative feature points.

[0067] Step 4: Rotate the axis by an angle θ in a fixed direction, repeat step 3, and calculate the total value.

[0068] Step 5: Repeat step 4 until the axis rotates one full revolution. Compare all the total values ​​and take the axis position with the minimum total value as the solar meridian position. This method uses the antisymmetry of the atmospheric polarization pattern to fit the solar meridian and determine the solar meridian position.

[0069] Step 6: Use the carrier's position and time information to calculate the absolute azimuth of the sun and the absolute heading angle.

[0070] Specifically, the absolute azimuth of the sun can be calculated based on the latitude and longitude coordinates of the carrier's location and the current time, that is, the azimuth of the sun under the navigation system The solar azimuth angle under the load system is calculated using the slope k of the solar meridian obtained by fitting. Due to the solar azimuth angle in the navigation system The north direction is used as the reference, so use The angle between the carrier and the true north direction, that is, the absolute heading angle, can be obtained.

[0071] in, The calculation process is:

[0072] θ s =arcsin(sinεsinβ+cosεcosβcost)

[0073]

[0074] Among them, θ s is the solar altitude angle, is the solar azimuth angle under the navigation system, ε is the latitude of the observation point, β is the solar declination angle, and t is the solar hour angle.

[0075] The following comparative test verifies the effect of this method. The experimental equipment mainly includes a homemade bionic polarization compass, a satellite / inertial combined navigation benchmark and an unmanned vehicle. The bionic polarization compass consists of a PHX050S-PC polarization camera, an NVIDIA TX2 processor and a wide-angle lens. The polarization camera is equipped with an IMX250ZR CMOS photosensitive chip, which is covered with a pixelated micro-polarization array. The four adjacent pixels will output light intensities in different polarization directions (0°, 45°, 90°, 135°). Therefore, polarization images in different directions can be directly obtained through the camera, and the heading angle can be calculated. The satellite / inertial combined navigation benchmark is SPAN-KVH1750, and its orientation accuracy is 0.035°. As Figure 6 Comparison of meridian fitting results between two existing technologies and the method of this patent under cloudy weather, such as Figure 7 Comparison of meridian fitting results between two existing technologies and the method of this patent under cloudy and blocked weather conditions, such as Figure 8 Comparison of meridian fitting results for two existing techniques and the patented method under extreme obstruction weather conditions. Each figure includes images captured in four different scenarios, each with fitting results from three different methods. The green line represents the least squares method, the blue line represents the Ransac algorithm, and the white line represents the fitting result of the present invention. The ideal solar meridian should be located at the red-blue dividing line. The accuracy of the invention's fitting can be determined by the deviation angle between the fitted solar meridian and the dividing line.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A polarization orientation method based on the antisymmetry of atmospheric polarization mode, characterized in that: The following steps are involved: Step 1: Use a polarization camera to obtain an original sky polarization image, use the UNet convolutional neural network to perform binary segmentation on the sky polarization image to obtain a de-obstructed sky polarization image, and calculate the polarization angle of each pixel in the image; Step 2: In the de-obstructed polarized sky image, take the zenith as the center of the circle and randomly set the axis so that the axis passes through the circle point; Step 3: Using the axis as a reference, divide the sky into eight regions at 45° intervals, named 1 to 8. Regions 1, 4, 5, and 8 are designated as valid regions, while regions 2, 3, 6, and 7 are designated as interference regions. Separate the polarization image of the sky into multiple equal-width circular rings. For each circular ring, calculate the following parameters based on the polarization angle of each pixel: The number of positive feature points P_num, negative feature points N_num and total feature points T_num in the effective area of ​​the ring, the number of positive feature points P_score in areas 1 and 5, the number of negative feature points N_score in areas 4 and 8 and the total number of feature points I_num in the interference area; Calculate the δ value of each ring area, and calculate the total value δ from the δ value of each ring area T ; Among them, w i is the weight corresponding to each ring; Step 4: Rotate the axis by an angle θ in a fixed direction, repeat step 3, and calculate the total value; Step 5: Repeat step 4 until the axis rotates one circle, compare all the total values, and take the axis position when the total value is the smallest as the position of the solar meridian; Step 6: Use the carrier's position and time information to calculate the sun's position and the absolute heading angle.

2. The polarization orientation method based on the antisymmetry of atmospheric polarization mode according to claim 1, characterized in that: In step 1, the method for obtaining the polarized image of the sky after de-obstruction is: Use the dataset to train the UNet convolutional neural network and use the trained UNet network to obtain the mask; The original sky polarization image is preprocessed using an image opening operation to remove abnormal small areas in the original sky polarization image to obtain a preprocessed sky polarization image. The mask is multiplied with the preprocessed sky polarization image to obtain the de-occluded sky polarization image.

3. The polarization orientation method based on the antisymmetry of the atmospheric polarization mode according to claim 1 or 2, characterized in that: In step 1, the polarization angle of each pixel in the image is calculated using the Stokes formula, specifically: Where Aop is the polarization angle of each pixel in the image, I(0), I(45), I(90), and I(135) are the intensity values ​​of the polarization directions of the pixel at 0°, 45°, 90°, and 135°, respectively.

4. The polarization orientation method based on the antisymmetry of atmospheric polarization mode according to claim 3, characterized in that: In step 3, the total feature points include positive feature points and negative feature points. The method for obtaining feature points is: Set a threshold α to expand the polarization angle range and perform benchmark transformation. Based on the Aop value obtained in step 1, select pixels whose polarization angle absolute value is greater than α as feature points. Pixels with positive polarization angles are positive feature points, and pixels with negative polarization angles are negative feature points.

5. The polarization orientation method based on the antisymmetry of atmospheric polarization mode according to claim 2, characterized in that: In step 1, the network structure of the UNet convolutional neural network is a U-shaped structure with four layers of downsampling and four layers of upsampling.

6. The polarization orientation method based on the antisymmetry of atmospheric polarization mode according to claim 1 or 2, characterized in that: In step 6, define the solar azimuth under the load system Solar azimuth under the navigation system Calculate the absolute heading angle i s =arcsin(sineσinβ+cosεcosβcost) Among them, θ s is the solar altitude angle, ε is the latitude of the observation point, β is the solar declination angle, and t is the solar hour angle.

Images