Passive positioning method based on underwater polarized light

Abstract

The invention relates to a passive positioning method based on underwater polarized light. Comprising the following steps: collecting a voucher picture recording polarization mode information of underwater polarized light; preprocessing the voucher picture to generate a three-channel image integrated with an underwater polarization angle, a polarization degree and polarized light intensity; performing regression prediction by taking the three-channel image as input of a neural network, and outputting predicted sun position information; and estimating predicted geographical location information according to the predicted sun location information. According to the invention, underwater polarization information is processed through the neural network, the sun position is calculated, and then the geographic position is determined for high-precision positioning. According to the underwater positioning method, sensors such as an accelerometer and a gyroscope are avoided, so that the positioning error is reduced by adopting lower cost on the basis of ensuring the accuracy of underwater positioning.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of navigation and positioning technology, and in particular to a passive positioning method based on underwater polarized light. Background Art

[0002] With the development of science and technology and the progress of society, traditional methods such as observing the starry sky for navigation have gradually failed to meet the needs of people's daily production and life. Therefore, people began to develop navigation equipment. Nowadays, navigation has penetrated into every aspect of our lives: transportation, tourism, rescue, detection, marine operations, weather forecasting and many other fields rely on navigation and positioning technology.

[0003] According to the navigation environment, it can be divided into free space navigation and underwater navigation. Compared with free space navigation, underwater navigation is more difficult due to the skin effect of water on radio and the complexity of the underwater environment is much greater than that on land. Modern satellite systems, such as the US GPS, China's Beidou, Europe's Galileo, etc., can provide high-precision navigation for anyone, but the low visibility and high electromagnetic loss in the underwater environment greatly limit the accuracy of navigation. Finally, inertial navigation, or "dead reckoning", is a navigation based on the comprehensive observation of a person's internal state over time to determine the change of position, but the error of the inertial navigation system will accumulate over time. Due to the measurement errors of the sensors in the system (such as accelerometers and gyroscopes), the position and velocity calculation errors will gradually increase over time, and the maintenance cost of the system is high. Although it does not rely on external signals, the accuracy of the inertial navigation system will be affected by environmental factors such as temperature changes and vibrations, which may lead to increased errors.

[0004] Regarding the above technical solution, the inventors found that there are at least the following technical problems:

[0005] Positioning using sensors such as accelerometers and gyroscopes will inevitably result in measurement errors that accumulate over time, and the measurement errors will gradually increase, making it impossible to perform accurate positioning.

[0006] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0007] It should be noted that this section is intended to provide background or context for the technical solutions of the present invention stated in the claims. The description herein is not admitted to be prior art by inclusion in this section. Summary of the invention

[0008] The object of the present invention is to provide a passive positioning method based on underwater polarized light, thereby at least to a certain extent solving one or more problems caused by the limitations and defects of the related art.

[0009] The present invention provides a passive positioning method based on underwater polarized light, comprising:

[0010] Collect voucher images that record polarization pattern information of underwater polarized light;

[0011] Preprocessing the voucher image to generate a three-channel image integrating underwater polarization angle, polarization degree and polarization light intensity;

[0012] Using the three-channel image as input of a neural network for regression prediction, and outputting predicted sun position information;

[0013] The predicted geographical location information is estimated based on the predicted sun position information.

[0014] Optionally, the step of preprocessing the voucher image to generate a three-channel image integrating underwater polarization angle, polarization degree and polarized light intensity includes:

[0015] Extracting polarization mode information of the voucher image, and calculating characteristic values ​​of the underwater polarization angle, polarization degree and polarized light intensity according to the polarization mode information;

[0016] The characteristic values ​​of the underwater polarization angle, polarization degree and polarized light intensity are respectively taken as a channel and merged into the three-channel image.

[0017] Optionally, the step of using the three-channel image as an input of a neural network for regression prediction and outputting predicted sun position information includes:

[0018] The three-channel image is passed into the convolution layer for convolution operation; wherein the convolution operation of the convolution layer includes subjecting the result of the convolution kernel to normalization processing, maximum pooling processing, residual block fitting processing, average pooling processing, and then outputting the result through a fully connected layer.

[0019] Optionally, the step of using the three-channel image as an input of a neural network for regression prediction and outputting predicted sun position information includes:

[0020] The output result of the convolution operation of the convolution layer is subjected to global feature compression and channel weight learning through a compression excitation module, and a weighted feature map of the three channels is output.

[0021] Optionally, the step of performing global feature compression and channel weight learning on the output result of the convolution operation of the convolution layer through a compression excitation module and outputting a weighted feature map of the three channels includes:

[0022] The global feature compression compresses the spatial information of each channel into a single value through global average pooling.

[0023] Optionally, the step of performing global feature compression and channel weight learning on the output result of the convolution operation of the convolution layer through a compression excitation module and outputting a weighted feature map of the three channels includes:

[0024] The channel weight learning learns a nonlinear relationship through a ReLU activation function, captures the dependency between channels through the nonlinear relationship and outputs the weight of each channel;

[0025] The learned weight of each channel is assigned to the corresponding channel in the three-channel image to generate the three-channel weighted feature map.

[0026] Optionally, the step of estimating predicted geographical location information based on the predicted sun position information includes:

[0027] The predicted geographical location information is estimated based on the predicted sun position information by using a compression approximation algorithm.

[0028] Optionally, the step of using the three-channel image as an input of a neural network for regression prediction and outputting predicted sun position information further includes:

[0029] The predicted solar position information is subjected to error calculation according to the correct solar position information and the regression prediction process of the neural network is optimized.

[0030] Optionally, the step of performing error calculation on the predicted solar position information according to the correct solar position information and optimizing the regression prediction process of the neural network includes:

[0031] The mean square error function is used as the loss function. The gradient descent method is used to optimize the neural network with the purpose of minimizing the loss function, and the model parameters of the neural network are updated along the negative gradient direction of the loss function.

[0032] Optionally, the step of performing error calculation on the predicted solar position information according to the correct solar position information and optimizing the regression prediction process of the neural network includes:

[0033] The correct sun position information is calculated based on the determined geographic information, time information and environmental information of the observation point to obtain the sun position information of the observation point.

[0034] The technical solution provided by the present invention may include the following beneficial effects:

[0035] In the present invention, high-precision positioning is performed by processing underwater polarization information through a neural network and calculating the sun position to determine the geographical location, avoiding the use of sensors such as accelerometers and gyroscopes, thereby ensuring the accuracy of underwater positioning and reducing positioning errors at a lower cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present invention, and together with the specification are used to explain the principles of the present invention. Obviously, the accompanying drawings described below are only some embodiments of the present invention, and for those of ordinary skill in the art, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0037] Figure 1 A schematic flow chart showing a passive positioning method based on underwater polarized light in an exemplary embodiment of the present invention;

[0038] Figure 2 A schematic diagram showing an arrayed polarization image in a polarization sensor in an exemplary embodiment of the present invention;

[0039] Figure 3 A schematic diagram showing the structure of a compression excitation module in an exemplary embodiment of the present invention is shown;

[0040] Figure 4 An error diagram showing the error distribution of the predicted solar zenith angle and azimuth angle at different locations from 11:00 to 13:00 in an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0041] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the present invention will be more comprehensive and complete and fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0042] In addition, the accompanying drawings are only schematic illustrations of embodiments of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0043] The present invention provides a passive positioning method based on underwater polarized light, referring to Figure 1 As shown in , the following steps are included:

[0044] Step S100: collecting a voucher image recording polarization pattern information of underwater polarized light.

[0045] Step S200: pre-processing the voucher image to generate a three-channel image integrating underwater polarization angle, polarization degree and polarization light intensity.

[0046] Step S300: using the three-channel image as input of a neural network to perform regression prediction and output predicted sun position information.

[0047] Step S400: estimating predicted geographical location information according to the predicted sun position information.

[0048] It should be understood that the distribution of underwater polarization pattern information is strongly correlated with the position of the sun, and the position information of the sun at a certain moment is also correlated with geographic information, which makes it feasible to use polarization pattern information for geographic positioning. By using a polarization camera to record the voucher image of the polarization pattern information (pattern distribution) of underwater polarized light, based on the strong correlation between the distribution of polarization states and the position information of the sun, the compressed excitation residual neural network can be used to estimate the sun position information of the observation point, and then predict the observed geographic information (longitude and latitude). Extracting polarization information requires a focal plane polarization camera (DoFP).

[0049] It is also important to understand that reference Figure 2 As shown in the figure, the polarization image taken by the focal plane polarization camera is used as the evidence image. Each pixel of the polarization image is a super pixel point composed of 4 sub-pixels. The polarization mode information of 0°, 45°, 135°, and 90° can be extracted respectively, and the polarization information such as the underwater polarization angle (Angle of Polarization, AOP), degree of polarization (Degree of Polarization, DOP) and polarization intensity (Intensity, I) can be calculated based on this information. This information is used as a prediction basis for subsequent processing.

[0050] It is also necessary to understand that AOP, DOP and I are used at the same time to represent the polarization distribution as input, and the Squeeze and Excitation Residual Network (SE-ResNet) is used, so the network model of the standard residual neural network (ResNet) needs to be adjusted. The adjusted ResNet network can better process polarization images and extract the polarization distribution contained in the image.

[0051] The passive positioning method based on underwater polarized light is adopted to process underwater polarization information through neural network and calculate the sun position to determine the geographical location for high-precision positioning. It avoids the use of sensors such as accelerometers and gyroscopes, thereby ensuring the accuracy of underwater positioning and reducing positioning errors at a lower cost.

[0052] Next, we will refer to Figures 1 to 3 The various steps of the above-mentioned passive positioning method based on underwater polarized light in this exemplary embodiment are described in more detail.

[0053] In some embodiments, step S200 includes:

[0054] The polarization mode information of the voucher image is extracted, and the characteristic values ​​of the underwater polarization angle, polarization degree and polarized light intensity are calculated according to the polarization mode information.

[0055] The characteristic values ​​of the underwater polarization angle, polarization degree and polarized light intensity are respectively taken as a channel and merged into the three-channel image.

[0056] It is important to understand that reference Figure 2 As shown in the figure, the resolution of the polarization image is 2448×2048. Because of the edge occlusion and other problems that may exist during shooting, the 600×600 pixel part within the Snell window is cropped as the original image. The reduced sky seen by the observer underwater is called the Snell window. Then the AOP, DOP and light intensity I are calculated, and the data is standardized at the same time, and the range is scaled to between 0-1. Because the experiment uses a black and white camera, the size of the three images is (1×300×300). AOP is used as channel 1, DOP is used as channel 2, and I is used as channel 3. The three one-channel images are synthesized into a three-channel image of size (3×300×300) as input. The distribution of these three polarization characteristics is used to characterize the distribution of polarization modes, which can describe the polarization information at different positions in more detail, that is, "the polarization angle of the partial polarized light with a proportion of linear polarization light of DOP at this point is AOP and the light intensity is I".

[0057] In some embodiments, step S300 includes:

[0058] The three-channel image is passed into the convolution layer for convolution operation.

[0059] Among them, the convolution operation of the convolution layer includes processing the result of the convolution kernel through normalization, maximum pooling, fitting of the residual block, and average pooling in sequence, and then outputting the result through a fully connected layer.

[0060] It is important to understand that reference Figure 3As shown in , after data preprocessing, the three-channel image is passed to the convolution layer for convolution operation calculation, as shown in formula (1),

[0061]

[0062] Where f represents the input image, h represents the kernel used, and m and n represent the row and column indices, respectively.

[0063] The convolution kernel of the convolution layer is 7×7, and the change in image size caused by convolution is determined by formula (2).

[0064]

[0065] Among them, W 1 is the image size before convolution, W 2 Now is the image size after convolution, F is the receptive field size, p is the number of zero padding, and Str is the sliding step size. Finally, the results are processed by normalization, maximum pooling, fitting of residual blocks, average pooling, and finally output through the fully connected layer.

[0066] In some embodiments, step S300 includes:

[0067] Step S301: The output result of the convolution operation of the convolution layer is subjected to global feature compression and channel weight learning through a compression excitation module, and a weighted feature map of the three channels is output.

[0068] It is important to understand that reference Figure 3 As shown in , after the compression excitation module (SE module) is added to the last 1×1 convolutional layer of each layer of the block, the SE module accepts the feature map output by this block, and through global feature compression and channel weight learning, readjusts the weights of the three channels of AOP, DOP and I, and uses the weighted feature map as the final output of the block.

[0069] In some embodiments, step S301 includes:

[0070] The global feature compression compresses the spatial information of each channel into a single value through global average pooling.

[0071] It is important to understand that reference Figure 3As shown in , a compression excitation module is added after the convolution layer. Global feature compression (Squeeze): For the feature map of a given input polarization mode distribution image, the spatial information of each channel is compressed into a single value through global average pooling. The size of the input feature map is 300×300, and each channel obtains a global description after global average pooling. This scalar can be regarded as the global feature of the channel on the entire image. This operation can effectively capture the global information of each channel and remove the spatial dimension, thereby compressing the dimension of the input polarization image feature distribution from 3×300×300 to 3×1×1. This value represents the global information of the channel in the entire feature map, that is, the average response of the channel.

[0072] As shown in formula (3):

[0073]

[0074] Among them, X c represents the feature map of the cth channel in the input feature map, H and W represent the height and width of the polarization feature map respectively, and Z c The feature vector representing the global average pooling of the c-th channel.

[0075] In some embodiments, step S301 includes:

[0076] The channel weight learning learns a nonlinear relationship through a ReLU activation function, captures the dependency between channels through the nonlinear relationship and outputs the weight of each channel.

[0077] The learned weight of each channel is assigned to the corresponding channel in the three-channel image to generate the three-channel weighted feature map.

[0078] It is important to understand that reference Figure 3 As shown in, channel weight learning (Excitation): After feature compression, the SE module enters the channel weight learning stage. This stage uses a small fully connected neural network (FC layer), whose purpose is to learn the importance weight of each channel. First, the global features of each channel are reduced in dimension through a fully connected layer, and a small neural network with a linear rectification (Rectified Linear Unit, ReLU) activation function is used to learn nonlinear relationships. Then, the dimension of the original channel is restored through another fully connected layer, and the weights are mapped to between 0 and 1 using the Sigmoid function. The role of these two fully connected layers is to flexibly capture the complex dependencies between channels through nonlinear transformations and output the weights of each channel.

[0079] As shown in formula (4),

[0080] s=σ(U 2 RELU(U 1 ·z)) (4)

[0081] Among them, U 1 and U 2 is the weight matrix of the fully connected layer, s is the weight of each channel, σ is the Sigmoid function, and z represents the feature vector of average pooling.

[0082] After obtaining the channel weights, the SE module enters the channel recalibration phase. In this phase, the SE module redistributes the learned weights to the channels of the original feature map through channel-by-channel multiplication operations. This means that for each channel, the SE module will amplify or reduce the features of the channel according to its importance weight.

[0083] As shown in formula (5),

[0084]

[0085] in, is the recalibrated feature map, s c is the channel weight.

[0086] In some embodiments, step S400 includes:

[0087] The predicted geographical location information is estimated based on the predicted sun position information by using a compression approximation algorithm.

[0088] It is important to understand that when determining the time, geographic information and other basic information of the observation point that meets the conditions, the corresponding solar position information can be calculated. To establish the relationship between time, solar azimuth, solar zenith angle and the longitude and latitude of the location, if you try to build a deep learning network, then for a certain moment, you need to consider each corresponding geographic location between longitude 180°E and 180°W and latitude 90°S and 90°N, and the existing geographic positioning accuracy is in the "minute", that is, 1 / 3600, in 10 -4 The total amount of data at a single moment is 360×180×108=6.48×10 12 ,The amount of data is too large, which poses a great challenge to the model training in terms of hardware and software.,Therefore, a compression approximation algorithm is adopted to approximate the real geographic longitude and latitude information by compressing the space, and ensure the accuracy while reducing the amount of data.

[0089] ①First, obtain the time information t, and then generate two variables latitude1 and longitude1 by the computer. The value range of latitude1 is [-90,90], and the value range of longitude1 is [-180,180], and the interval is 1. That is, after obtaining the time information, the longitude and latitude are separated by 1°, and all the longitude and latitude are traversed. At this time, a set of solar zenith angle and solar azimuth angle information can be obtained.

[0090] ② According to formula (6), calculate the mean square error between the solar azimuth angle and solar zenith angle of each group and the predicted solar zenith angle and solar azimuth angle.

[0091]

[0092] Among them, azi pre , zen pre They are the solar azimuth and zenith angle predicted by the network, azi 0 , zen 0 The solar azimuth and zenith angles corresponding to each set of longitude and latitude are calculated with a latitude and longitude interval of 1°. 1 ,lon 1 ) can be roughly regarded as the longitude and latitude of the observation point.

[0093] ③Based on step ②, solve the longitude and latitude (lat 1 ,lon 1 ) is further subdivided near (lat 1 ±1,lon 1 ±1), repeat steps ① and ② with an interval of 0.1° to obtain a positioning value with an accuracy of 0.1° (lat 2 ,lon 2 ). The regeneration range is (lat 2 ±0.1,lon 2 ±0.1), with an interval of 0.01°, repeat steps ① and ② to obtain a positioning value with an accuracy of 0.01° (lat 3 ,lon 3 ), and approximate it in sequence, the accuracy can reach 0.0001, that is, the positioning information (lat, lon) at the 'minute' level.

[0094] ④ In practical applications, the accuracy can be further improved based on the solar information predicted at the same location and at different times. Take the (lat, lon) obtained in ③ as an example. Assume that there are n polarization images at the same location and at different times. After prediction, a series of data (lat 1 ,lon 1 ), (lat2 ,lon 2 ),.......(lat n ,lon n ), this set of data is recorded as data set P, where (lat 1 ,lon 1 ) represents the longitude and latitude predicted based on the first image, (lat 2 ,lon 2 ) is based on the longitude and latitude predicted by the second picture, and so on. Each picture is processed according to steps ① to ③, and finally the data set P is obtained. The RANSAC (Random Sample Consensus) iterative adaptive model fitting algorithm is used to find the inliers that meet the requirements, and the average value of the inliers is calculated as the centroid. The best model is shown in formula (7):

[0095]

[0096] Where M is all the fitting models, I represents the inliers, E is the preset distance threshold, d is the distance calculated by the model, and M * The model that can contain the most inliers is the best model. The centroid of all inliers can be regarded as the predicted longitude and latitude of the point.

[0097] In some embodiments, after step S300, the method further includes:

[0098] Step S500: performing error calculation on the predicted solar position information according to the correct solar position information and optimizing the regression prediction process of the neural network.

[0099] It should be understood that the regression prediction process of optimizing the neural network is to calculate the loss function and transmit it backwards.

[0100] In some embodiments, step S500 includes:

[0101] The mean square error function is used as the loss function. The gradient descent method is used to optimize the neural network with the purpose of minimizing the loss function, and the model parameters of the neural network are updated along the negative gradient direction of the loss function.

[0102] It is important to understand that the error between the results of SE-ResNet network training and the correct sun's geographical location is calculated, and the mean square error (MSE) function is used as the model loss function. The model is optimized using the gradient descent method, and the model parameters are updated along the negative gradient direction of the loss function to minimize the loss function.

[0103] In some embodiments, step S500 includes:

[0104] The correct sun position information is calculated based on the determined geographic information, time information and environmental information of the observation point to obtain the sun position information of the observation point.

[0105] It is important to understand that the solar position algorithm is based on the geographical information (such as longitude and latitude), time information (such as shooting time) and environmental information (such as altitude, average temperature and atmospheric pressure) of the observation point, and can predict the solar zenith angle and solar azimuth angle of the observation point. The details are as follows:

[0106] ① Time conversion and Julian calendar calculation: Convert the current time to Julian day, Julian calendar day, Julian century and Julian calendar millennium.

[0107] ② Calculation of the Earth's position: Calculate the Earth's heliocentric longitude, heliocentric latitude and radius vector based on the Julian calendar millennium.

[0108] ③ Geocentric coordinate calculation: Use the earth's heliocentric longitude, latitude and periodic terms to calculate the geocentric longitude and latitude.

[0109] ④ Nutation calculation: Calculate the longitude nutation and inclination nutation based on the Julian calendar century and periodic terms.

[0110] ⑤Ecliptic inclination and distortion correction: Calculate the true inclination of the ecliptic based on the Julian calendar millennium and inclination nutation, and calculate the distortion correction based on the earth radius vector.

[0111] ⑥ Calculation of solar apparent longitude and Greenwich apparent sidereal time: Calculate the solar apparent longitude based on the geocentric longitude, longitude nutation and distortion correction, and calculate the Greenwich apparent sidereal time based on the Julian day, longitude nutation and true inclination of the ecliptic.

[0112] ⑦ Calculation of solar right ascension and declination and local hour angle: Calculate the geocentric solar right ascension and declination based on the solar apparent longitude, true inclination of the ecliptic and geocentric latitude, and calculate the observer's local hour angle based on Greenwich apparent sidereal time, the observer's geographic longitude and geocentric solar right ascension.

[0113] ⑧ Calculation of station center coordinates: Calculate the station center solar right ascension, station center solar declination and station center local hour angle based on the earth radius vector, observer's geographic latitude, observer's geographic altitude, observer's local hour angle and geocentric solar declination.

[0114] ⑨ Calculation of solar zenith angle and solar azimuth angle: Calculate the elevation angle of the station center without considering atmospheric refraction according to the geographic latitude of the observer, the solar declination of the station center and the local hour angle of the station center, then calculate the atmospheric refraction correction value according to the atmospheric pressure, the average temperature of the year and the elevation angle of the station center, calculate the elevation angle of the station center by the elevation angle of the station center without considering atmospheric refraction and the atmospheric refraction correction value, and finally calculate the zenith angle of the station center according to the elevation angle of the station center. Calculate the astronomical azimuth of the station center according to the local hour angle of the station center, the geographic latitude of the observer and the solar declination of the station center, and calculate the azimuth of the station center according to the astronomical azimuth of the station center.

[0115] The functional results of the present invention are verified in combination with the above-mentioned embodiments.

[0116] Table 1 shows the meaning of the parameters in the experiment.

[0117] Table 1 Experimental parameters

[0118]

[0119]

[0120] The experiment uses a polarization camera to collect polarization images. The polarization camera model used in this experiment is a focal plane polarization camera. The BFS-PGE-51S5P polarization camera produced by FLIR is selected. It is a grayscale camera with a resolution of 2448×2048. In order to make the experimental results more universal, the experiment collected data at three locations with longitude and latitude of (34°9′18″N, 108°54′1″E), (34°3′39″N, 108°53′30″E), and (34°38′28″N, 109°7′8″E). In the experiment, the camera was placed vertically on the horizontal plane, and a set of polarization data collected under clear weather conditions was collected every 30 minutes. Each set of data had 150 images, and each image was separated by 2 seconds. The original images of underwater polarization distribution collected by the experiment, after excluding some inappropriate images, the remaining images will be used as training sets and output sets. The data set of each collection location is image preprocessed.

[0121] like Figure 2 As shown, the I of each pixel is collected 0 ,I 45 ,I 90 ,I 135 , the AOP, DOP and I of each superpixel are calculated by formulas (8) to (10).

[0122]

[0123]

[0124]

[0125] Where, I in formula (8) is the polarized light intensity at that point, I R and I LThey refer to right-handed polarized light and left-handed polarized light respectively. However, in the natural environment, most polarized light contains only a very small amount of circular polarization components, so the V value can be ignored. The DOP value range is [0,1]. The polarization angle (AOP) refers to the angle between the vibration direction of light and a specific direction. The polarization angle value range is [-90°, 90°].

[0126] The experiment combines the three eigenvalues ​​in the polarization mode into a three-channel image. Finally, the preprocessed data is used as input, and the neural network is used for regression prediction, and the sun position (zenith angle and azimuth) information is analyzed and predicted. The number of images collected at each location is as follows:

[0127] Table 2 Datasets collected at four locations

[0128]

[0129] The training sets of locations 1 and 3 are used as the total training set, and the data of location 2 is used as the test set. According to the regression prediction process of the neural network, the training set is trained through the deep learning network model, and the distribution results of the error can be obtained according to the model. The theoretical data and predicted data results of the sun's position are shown in Table 3.

[0130] Table 3 Prediction errors of solar zenith angle and azimuth angle at location 2 (partial)

[0131]

[0132] By classifying the error distribution according to time, we can obtain the error distribution of the predicted zenith angle and azimuth angle as follows: Figure 4 shown.

[0133] from Figure 4 As can be seen from Table 3, the errors of Azi_Error (blue) and Zen_Error (orange) remain relatively small in most time periods, especially between 11:00 and 13:00. The errors in this time period are very small, and the prediction error of the model is low, indicating that the model predicts the azimuth and zenith angles well and stably in this time period.

[0134] Table 4 Prediction errors of solar zenith angle and azimuth angle at location 2 on sunny days

[0135] Maximum error / ° Minimum error / ° Average error / ° Solar zenith angle 0.763 0.472 0.622 Solar azimuth 1.624 0.014 0.295

[0136] Therefore, based on the predicted data of the sun's position on a sunny day and the idea of ​​the compressed approximation algorithm, the longitude and latitude information can be obtained, and the result of geographic location prediction using data from multiple locations can be obtained. Finally, the estimated longitude and latitude of the geographic location are: (34°17'58"N, 108°53'38"E), and the true value is (34°3′39″N, 108°53′30″E), that is, the latitude error is 0.238°, and the longitude error is: 0.0023°.

[0137] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine different embodiments or examples described in this specification.

[0138] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed by the present invention. The specification and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the appended claims.

Claims

1. A passive positioning method based on underwater polarized light, characterized in that: include: Collect voucher images that record polarization pattern information of underwater polarized light; Preprocessing the voucher image to generate a three-channel image integrating underwater polarization angle, polarization degree and polarization light intensity; Using the three-channel image as input of a neural network for regression prediction, and outputting predicted sun position information; The predicted geographical location information is estimated based on the predicted sun position information.

2. The passive positioning method based on underwater polarized light according to claim 1, characterized in that: The step of preprocessing the voucher image to generate a three-channel image integrating underwater polarization angle, polarization degree and polarized light intensity comprises: Extracting polarization mode information of the voucher image, and calculating characteristic values ​​of the underwater polarization angle, polarization degree and polarized light intensity according to the polarization mode information; The characteristic values ​​of the underwater polarization angle, polarization degree and polarized light intensity are respectively taken as a channel and merged into the three-channel image.

3. The passive positioning method based on underwater polarized light according to claim 1, characterized in that: The step of using the three-channel image as an input of a neural network for regression prediction and outputting predicted sun position information comprises: The three-channel image is passed into the convolution layer for convolution operation; wherein the convolution operation of the convolution layer includes subjecting the result of the convolution kernel to normalization processing, maximum pooling processing, residual block fitting processing, average pooling processing, and then outputting the result through a fully connected layer.

4. The passive positioning method based on underwater polarized light according to claim 3, characterized in that: The step of using the three-channel image as an input of a neural network for regression prediction and outputting predicted sun position information comprises: The output result of the convolution operation of the convolution layer is subjected to global feature compression and channel weight learning through a compression excitation module, and a weighted feature map of the three channels is output.

5. The passive positioning method based on underwater polarized light according to claim 4, characterized in that: The step of performing global feature compression and channel weight learning on the output result of the convolution operation of the convolution layer through a compression excitation module and outputting a feature map weighted for the three channels comprises: The global feature compression compresses the spatial information of each channel into a single value through global average pooling.

6. The passive positioning method based on underwater polarized light according to claim 5, characterized in that: The step of performing global feature compression and channel weight learning on the output result of the convolution operation of the convolution layer through a compression excitation module and outputting a feature map weighted for the three channels comprises: The channel weight learning learns a nonlinear relationship through a ReLU activation function, captures the dependency between channels through the nonlinear relationship and outputs the weight of each channel; The learned weight of each channel is assigned to the corresponding channel in the three-channel image to generate the three-channel weighted feature map.

7. The passive positioning method based on underwater polarized light according to claim 1, characterized in that: The step of estimating predicted geographical location information based on the predicted sun position information comprises: The predicted geographical location information is estimated based on the predicted sun position information by using a compression approximation algorithm.

8. The passive positioning method based on underwater polarized light according to any one of claims 1 to 7, characterized in that: After the step of using the three-channel image as the input of the neural network for regression prediction and outputting the predicted sun position information, the method further includes: The predicted solar position information is subjected to error calculation according to the correct solar position information and the regression prediction process of the neural network is optimized.

9. The passive positioning method based on underwater polarized light according to claim 8, characterized in that: The step of performing error calculation on the predicted solar position information according to the correct solar position information and optimizing the regression prediction process of the neural network comprises: The mean square error function is used as the loss function. The gradient descent method is used to optimize the neural network with the purpose of minimizing the loss function, and the model parameters of the neural network are updated along the negative gradient direction of the loss function.

10. The passive positioning method based on underwater polarized light according to claim 8, characterized in that: The step of performing error calculation on the predicted solar position information according to the correct solar position information and optimizing the regression prediction process of the neural network comprises: The correct sun position information is calculated based on the determined geographic information, time information and environmental information of the observation point to obtain the sun position information of the observation point.

Images