Unsupervised dual-target localization method based on geometric consistency of ocean environment

Abstract

The present invention discloses an unsupervised dual-target positioning method based on geometric consistency of an ocean environment, which relates to the technical field of dual-target positioning, including: obtaining infrared images and visible light images of different ocean scenes; obtaining a feature extraction network and establishing a feature extraction model; obtaining a geometric feature map; using an overall loss function to obtain an overall loss parameter; obtaining the optimal parameters of a geometric transformation matrix T; and finally calibrating the binocular system parameters of the infrared image and the visible light image. The present invention is used to solve the problem that existing dual-target positioning methods for visible light and infrared light cannot solve the problem that, when facing environmental interference and noise, they are easily affected by the differences in statistical features of multimodal images, resulting in inaccurate feature matching and difficulty in achieving robust calibration.

Description

Technical Field

[0001] The present invention relates to the technical field of dual-target positioning, and in particular to an unsupervised dual-target positioning method based on geometric consistency of an ocean environment. Background Art

[0002] Dual-target calibration refers to determining the geometric relationship between two cameras, including the intrinsic parameter matrix, extrinsic parameter matrix, and fundamental matrix. The main purpose of dual-target calibration is to accurately calculate the relative position and posture between the two cameras in applications such as 3D reconstruction and stereo matching, so as to accurately reconstruct objects in 3D space. Visible light and infrared dual-target calibration refers to the simultaneous calibration of visible light cameras and infrared cameras to determine the relative position relationship between them. This calibration method is very important when fusing visible light and infrared images, and can improve the accuracy and efficiency of image processing.

[0003] Existing methods for dual-target calibration of visible light and infrared light usually correct the external parameters of the camera by extracting matching feature points, and align the infrared image and the visible light image based on mutual information, thereby achieving joint calibration based on the positional relationship between the visible light image and the infrared image. Although this improved method can correct the external parameters before alignment to improve the accuracy of joint calibration, it can only calibrate the external parameters of the camera by maximizing the global and local mutual information between images of different modalities. When faced with environmental interference and noise, it is easily affected by the differences in statistical features of multimodal images, resulting in inaccurate feature matching and difficulty in achieving robust calibration. For example, in the publication No. The patent application CN110969669A discloses a joint calibration method for visible light and infrared cameras based on mutual information registration. This solution extracts matching feature points and utilizes the positional relationship between visible light images and infrared images to more effectively perform joint self-calibration of infrared cameras and visible light cameras. Other improvements in visible light and infrared dual-target calibration, usually in the selection of calibration objects, still cannot solve the problem that when faced with environmental interference and noise, they are easily affected by the differences in statistical features of multimodal images, resulting in inaccurate feature matching and difficulty in achieving robust calibration. In view of this, it is necessary to improve the existing methods for visible light and infrared dual-target calibration. Summary of the Invention

[0004] The present invention aims to solve, at least to a certain extent, one of the technical problems in the prior art. By proposing an unsupervised dual-target positioning method based on geometric consistency of the ocean environment, it is used to solve the problem that the existing dual-target positioning methods for visible light and infrared light are easily affected by the differences in statistical features of multimodal images when facing environmental interference and noise, resulting in inaccurate feature matching and difficulty in achieving robust calibration.

[0005] To achieve the above objectives, this application provides an unsupervised dual-target localization method based on geometric consistency of the ocean environment, comprising the following steps:

[0006] Acquire infrared and visible light images of different ocean scenes and store them in a multi-environment ocean library; perform data enhancement on the images in the multi-environment ocean library based on ocean environment characteristics;

[0007] Obtain a feature extraction network based on big data, and establish a feature extraction model based on the feature extraction network; use the feature extraction model to extract features from images in a multi-environment ocean image library, and obtain a geometric feature map based on the extraction results;

[0008] Obtain the geometric consistency loss, photometric consistency loss, and feature alignment loss in the geometric feature map, and use the overall loss function to obtain the overall loss parameter; obtain the geometric transformation matrix T based on unsupervised training, and obtain the optimal parameters of the geometric transformation matrix T based on the overall loss parameter;

[0009] The binocular system parameters of the infrared image and the visible light image are calibrated, wherein the binocular system parameters include external parameters and internal parameters.

[0010] Furthermore, infrared images and visible light images of different ocean scenes are obtained and stored in a multi-environment ocean image library, including:

[0011] Acquire a variety of ocean scenes based on possible weather conditions, possible lighting conditions, and possible sea conditions, and acquire infrared and visible light images of different ocean scenes based on real-scene acquisition and scene simulation;

[0012] The infrared images and visible light images of all ocean scenes are stored in a multi-environment ocean library, and the infrared images and visible light images of the same ocean scene are recorded as analysis image pairs.

[0013] Furthermore, data enhancement of images in the multi-environment ocean image library based on ocean environment characteristics includes: for any analysis image pair α in the multi-environment ocean image library, when the analysis image pair α is obtained by scene simulation, data enhancement of the analysis image pair α is performed using a scene enhancement method.

[0014] Furthermore, the scene enhancement method includes: recording the ocean scene corresponding to the scene simulation as the simulated real scene; obtaining the light intensity and sea conditions in the simulated real scene based on big data, and recording them as the simulated light intensity and the simulated sea conditions respectively; when the simulated light intensity is not equal to the light intensity in the scene simulation, adjusting the light intensity in the scene simulation to the simulated light intensity;

[0015] The ocean noise corresponding to the simulated sea condition is obtained and recorded as the simulated noise; the simulated noise is added to the scene simulation, and the analysis image pair in the scene simulation obtained at this time replaces the analysis image pair α in the multi-environment ocean library.

[0016] Furthermore, a feature extraction network is obtained based on the big data, and a feature extraction model is established based on the feature extraction network; the feature extraction model is used to extract features from images in the multi-environment ocean image library, and a geometric feature map is obtained based on the extraction results, including:

[0017] A feature extraction network is obtained based on big data, and the obtained feature extraction network is stored in a feature extraction model, wherein the input end of the feature extraction model is used to input image data, and the output end is used to output a geometric feature map obtained after processing by the feature extraction network.

[0018] Furthermore, the feature extraction network includes: when image data is input into the feature extraction model, feature extraction is performed on the image data based on the convolutional layer of the feature extraction network, and a feature image corresponding to the image data is obtained; for any position (x, y) in the image data, a position acquisition algorithm is used based on the Sinusoidal function to obtain a position code of the image data, and the position acquisition algorithm includes:

[0019] Among them, i is the dimension index, d model is the feature dimension of the feature extraction model, F(x) is the feature image output by the convolution layer, and F encoded It is a geometric feature map.

[0020] Furthermore, the geometric consistency loss, photometric consistency loss, and feature alignment loss in the geometric feature map are obtained, and the overall loss function is used to obtain the overall loss parameters, including:

[0021] Use the geometric loss algorithm to obtain the geometric consistency loss in the geometric feature map. The geometric loss algorithm is:

[0022] Among them, L geo is the geometric consistency loss, N is the number of feature points in the geometric feature map, T is the geometric transformation function, and F IR (i) is the pixel value of the infrared image of the i-th feature point in the geometric feature map, F VIS (i) is the pixel value of the visible light image of the i-th feature point in the geometric feature map;

[0023] The photometric loss algorithm is used to obtain the photometric consistency loss in the geometric feature map. The photometric loss algorithm is:

[0024] Among them, L cosineis the loss of photometric consistency;

[0025] The feature loss algorithm is used to obtain the feature alignment loss in the geometric feature map. The feature loss algorithm is:

[0026] Among them, L photometric is the feature alignment loss, I IR (i) is the pixel value of the i-th feature point in the infrared image, I VIS is the pixel value of the i-th feature point in the visible light image.

[0027] Furthermore, obtaining the geometric consistency loss, photometric consistency loss, and feature alignment loss in the geometric feature map and using the overall loss function to obtain the overall loss parameters also includes:

[0028] Use the overall loss function to obtain the joint optimization loss. The overall loss function is: L total =γ1×L geo +γ2×L cosine +γ3×L photometric , where L total To jointly optimize the loss, γ1 to γ3 are the first loss weight, the second loss weight, and the third loss weight, respectively.

[0029] Furthermore, obtaining the geometric transformation matrix T based on unsupervised training and obtaining the optimal parameters of the geometric transformation matrix T based on the overall loss parameter include:

[0030] The parameters in the feature extraction network are optimized using a backpropagation algorithm and gradient descent optimization methods until the infrared image in the feature extraction model is aligned with the features of its corresponding visible light image after transformation. The parameter optimization process is unsupervised learning.

[0031] When the infrared image in the feature extraction model can be aligned with the features of its corresponding visible light image after transformation, the parameters in the feature extraction network at this time are recorded as the geometric transformation matrix T; the minimum loss of the infrared features and visible light features after transformation is obtained based on the joint optimization loss, and the parameters of the geometric transformation matrix T at this time are recorded as the optimal parameters.

[0032] Furthermore, calibrating the binocular system parameters of the infrared image and the visible light image includes:

[0033] The calibration method of internal parameters is as follows: based on the calibration algorithm in computer vision, the internal parameters of the camera when shooting and analyzing the image pair are obtained;

[0034] The calibration method of the external parameters is: use the coordinate transformation algorithm to transform the coordinates in the visible light image and align the coordinates of the infrared image with the coordinates of the visible light image; the coordinate transformation algorithm is: Xhw =XT×X kj , where X hw is the coordinate in the infrared image, XT is the geometric transformation matrix T, X kj are the coordinates in the visible light image.

[0035] Beneficial effects of the present invention: The present application first obtains infrared images and visible light images of different ocean scenes; performs data enhancement on images in a multi-environment ocean library based on ocean environment characteristics; obtains a feature extraction network based on big data, and establishes a feature extraction model based on the feature extraction network; uses the feature extraction model to extract features from images in the multi-environment ocean library, and obtains a geometric feature map based on the extraction results. The advantage of this is that, by performing data enhancement on the image, it can ensure that the image to be analyzed can truly reflect the actual ocean environment, which helps subsequent analysis results to be in line with reality, thereby improving data accuracy; by obtaining a geometric feature map based on the feature extraction model, it is possible to obtain a characteristic relationship between the visible light image and the infrared image, which helps to obtain the characteristic difference between the visible light image and the infrared image under the influence of environmental interference and noise, thereby performing subsequent difference analysis;

[0036] This application also obtains the geometric consistency loss, photometric consistency loss and feature alignment loss in the geometric feature map, and uses the overall loss function to obtain the overall loss parameters; obtains the geometric transformation matrix T based on unsupervised training, and obtains the optimal parameters of the geometric transformation matrix T based on the overall loss parameters; finally, calibrates the binocular system parameters of the infrared image and the visible light image. The advantage of this is that by obtaining the overall loss parameters and obtaining the optimal parameters of the geometric transformation matrix T based on the overall loss parameters, the mutual conversion parameters of the visible light image and the infrared image when affected by environmental interference and noise can be obtained, thereby ensuring that the feature matching can be more accurate and robust calibration can be achieved during actual analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the steps of the method of the present invention;

[0038] Figure 2 Schematic diagram of the structure of the feature extraction model of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] Example 1, please refer to Figure 1 As shown, the present application provides an unsupervised dual-target positioning method based on geometric consistency of the ocean environment, comprising the following steps:

[0041] Step S1, acquiring infrared images and visible light images of different ocean scenes and storing them in a multi-environment ocean library; performing data enhancement on the images in the multi-environment ocean library based on ocean environment characteristics; Step S1 includes:

[0042] Step S101, acquiring a variety of different ocean scenes based on possible weather conditions, possible illumination conditions, and possible sea conditions, and acquiring infrared images and visible light images of the different ocean scenes based on real-scene capture and scene simulation;

[0043] In a specific implementation, the weather conditions in various ocean scenes may include: sunny days, foggy days, rainy days, snowy days and other weather conditions that can occur in the ocean; the lighting conditions in various ocean scenes may include: daytime, nighttime, sunset and other lighting conditions with varying lighting conditions in the ocean; and the sea conditions in various ocean scenes may include: calm, wavy and stormy sea surface conditions. By acquiring infrared images and visible light images in different ocean scenes, it can be ensured that the subsequent analysis parameters are more in line with the actual ocean scenes.

[0044] Step S102: storing infrared images and visible light images of all ocean scenes into a multi-environment ocean image library, and recording infrared images and visible light images of the same ocean scene as analysis image pairs;

[0045] Step S103: For any analysis image pair α in the multi-environment ocean image library, when the analysis image pair α is obtained by scene simulation, data enhancement is performed on the analysis image pair α using a scene enhancement method. The scene enhancement method comprises: step S1031: recording the ocean scene corresponding to the scene simulation as a simulated real scene; obtaining the light intensity and sea conditions in the simulated real scene based on big data, and recording them as simulated light intensity and simulated sea conditions, respectively; when the simulated light intensity is not equal to the light intensity in the scene simulation, adjusting the light intensity in the scene simulation to the simulated light intensity;

[0046] In the specific implementation process, while adjusting the illumination intensity, the brightness, contrast, and saturation of the analysis image pair corresponding to the scene simulation can also be adjusted to achieve consistency between the image obtained from the analysis scene simulation and the image obtained from the actual ocean, thereby improving the accuracy of data analysis;

[0047] Step S1032, obtain the ocean noise corresponding to the simulated sea condition and record it as the simulated noise; add the simulated noise to the scene simulation, and replace the analysis image pair α in the multi-environment ocean library with the analysis image pair in the scene simulation obtained at this time; in the specific implementation process, the simulated noise may include: simulating waves, sea surface reflections, hull shaking, and other noises that can be generated in the ocean due to sea conditions.

[0048] Step S2, obtaining a feature extraction network based on the big data, and establishing a feature extraction model based on the feature extraction network; using the feature extraction model to perform feature extraction on images in the multi-environment ocean image library, and obtaining a geometric feature map based on the extraction results;

[0049] Step S2 includes:

[0050] Step S201, please refer to Figure 2 As shown, a feature extraction network is obtained based on big data, and the obtained feature extraction network is stored in a feature extraction model, wherein the input end of the feature extraction model is used to input image data, and the output end is used to output a geometric feature map obtained after processing by the feature extraction network;

[0051] The feature extraction network includes: when image data is input into the feature extraction model, the convolutional layer of the feature extraction network extracts features from the image data and obtains a feature image corresponding to the image data; in a specific implementation, when obtaining the feature extraction network, a feature extraction network that can adapt to the marine environment can be selected, such as ResNet, EfficientNet, and Transformer;

[0052] For any position (x, y) in the image data, the position code of the image data is obtained using a position acquisition algorithm based on the Sinusoidal function. The position acquisition algorithm includes:

[0053] Among them, i is the dimension index, d model is the feature dimension of the feature extraction model, F(x) is the feature image output by the convolution layer, and F encoded is the geometric feature map;

[0054] In the specific implementation process, position encoding is used to enhance the model's perception of spatial position; for image data, position encoding can effectively guide the feature extraction network to understand the positional relationship of the target, calculate the position encoding for each input image, and use the Sinusoidal function; add position encoding to the feature map after convolution processing to enhance the perception of the target position.

[0055] Step S3: obtaining the geometric consistency loss, photometric consistency loss, and feature alignment loss in the geometric feature map, and obtaining the overall loss parameter using the overall loss function; obtaining the geometric transformation matrix T based on unsupervised training, and obtaining the optimal parameter of the geometric transformation matrix T based on the overall loss parameter;

[0056] Step S3 includes: step S301, using a geometric loss algorithm to obtain a geometric consistency loss in a geometric feature map, the geometric loss algorithm is:

[0057] Among them, L geo is the geometric consistency loss, N is the number of feature points in the geometric feature map, T is the geometric transformation function, and F IR (i) is the pixel value of the infrared image of the i-th feature point in the geometric feature map, F VIS (i) is the pixel value of the visible light image of the i-th feature point in the geometric feature map;

[0058] Step S302: Use a photometric loss algorithm to obtain the photometric consistency loss in the geometric feature map. The photometric loss algorithm is:

[0059] Among them, L cosine is the loss of photometric consistency;

[0060] Step S303: Use a feature loss algorithm to obtain the feature alignment loss in the geometric feature graph. The feature loss algorithm is:

[0061] Among them, L photometric is the feature alignment loss, I IR (i) is the pixel value of the i-th feature point in the infrared image, I VIS is the pixel value of the i-th feature point in the visible light image;

[0062] Step S304: Use the overall loss function to obtain the joint optimization loss. The overall loss function is: L total =γ1×L geo +γ2×L cosine +γ3×L photometric , where L total To jointly optimize the loss, γ1 to γ3 are the first loss weight, the second loss weight, and the third loss weight, respectively;

[0063] Step S305, optimizing the parameters in the feature extraction network based on a back-propagation algorithm and a gradient descent optimization method until the infrared image in the feature extraction model is aligned with the features of its corresponding visible light image after transformation, wherein the parameter optimization process is unsupervised learning;

[0064] Step S306: When the infrared image in the feature extraction model can be aligned with the features of its corresponding visible light image after transformation, the parameters in the feature extraction network at this time are recorded as the geometric transformation matrix T; based on the joint optimization loss, the minimum loss of the infrared features and the visible light features after transformation is obtained, and the parameters of the geometric transformation matrix T at this time are recorded as the optimal parameters.

[0065] Step S4, calibrating binocular system parameters of the infrared image and the visible light image, wherein the binocular system parameters include external parameters and internal parameters; Step S4 includes:

[0066] Step S401, a method for calibrating internal parameters is to obtain the internal parameters of the camera when shooting and analyzing the image pair based on a calibration algorithm in computer vision; in a specific implementation, the calibration algorithm can be the Zhang Zhengyou calibration method in OpenCV; the internal parameters are internal parameters, and the external parameters are external parameters;

[0067] Step S402: The external parameter calibration method is: using a coordinate transformation algorithm to transform the coordinates in the visible light image, and aligning the coordinates of the infrared image with the coordinates of the visible light image; the coordinate transformation algorithm is: X hw =XT×X kj , where X hw is the coordinate in the infrared image, XT is the geometric transformation matrix T, X kj are the coordinates in the visible light image.

[0068] In the embodiments provided in this application, it should be understood that the disclosed system or method can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or units can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of systems, modules and units can be electrical, mechanical or other forms.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. Unsupervised dual-target localization method based on geometric consistency of ocean environment, characterized by: The steps include: Acquire infrared and visible light images of different ocean scenes and store them in a multi-environment ocean library; perform data enhancement on the images in the multi-environment ocean library based on ocean environment characteristics; Obtain a feature extraction network based on big data, and establish a feature extraction model based on the feature extraction network; Use the feature extraction model to extract features from images in the multi-environment ocean image library, and obtain geometric feature maps based on the extraction results; Obtain the geometric consistency loss, photometric consistency loss, and feature alignment loss in the geometric feature map, and use the overall loss function to obtain the overall loss parameter; obtain the geometric transformation matrix T based on unsupervised training, and obtain the optimal parameters of the geometric transformation matrix T based on the overall loss parameter; Calibrate the binocular system parameters of the infrared image and the visible light image, wherein the binocular system parameters include external parameters and internal parameters; Among them, obtaining infrared images and visible light images of different ocean scenes and storing them in the multi-environment ocean library includes: Acquire a variety of ocean scenes based on possible weather conditions, possible lighting conditions, and possible sea conditions, and acquire infrared and visible light images of different ocean scenes based on real-scene acquisition and scene simulation; The infrared images and visible light images of all ocean scenes are stored in the multi-environment ocean image library, and the infrared images and visible light images of the same ocean scene are recorded as analysis image pairs; The data enhancement of images in the multi-environment ocean image library based on ocean environment characteristics includes: for any analysis image pair α in the multi-environment ocean image library, when the analysis image pair α is obtained by scene simulation, using a scene enhancement method to perform data enhancement on the analysis image pair α; The scene enhancement method is as follows: the ocean scene corresponding to the scene simulation is recorded as the simulated real scene; the light intensity and sea conditions in the simulated real scene are obtained based on big data, and recorded as the simulated light intensity and simulated sea conditions respectively; when the simulated light intensity is not equal to the light intensity in the scene simulation, the light intensity in the scene simulation is adjusted to the simulated light intensity; The ocean noise corresponding to the simulated sea condition is obtained and recorded as the simulated noise; the simulated noise is added to the scene simulation, and the analysis image pair in the scene simulation obtained at this time replaces the analysis image pair α in the multi-environment ocean library.

2. The unsupervised dual-target positioning method based on geometric consistency of ocean environment according to claim 1 is characterized in that: Obtain a feature extraction network based on big data, and establish a feature extraction model based on the feature extraction network; Use the feature extraction model to extract features from images in the multi-environment ocean image library, and obtain geometric feature maps based on the extraction results, including: A feature extraction network is obtained based on big data, and the obtained feature extraction network is stored in a feature extraction model, wherein the input end of the feature extraction model is used to input image data, and the output end is used to output a geometric feature map obtained after processing by the feature extraction network.

3. The unsupervised dual-target positioning method based on geometric consistency of ocean environment according to claim 2 is characterized in that: The feature extraction network includes: when image data is input into the feature extraction model, the convolutional layer of the feature extraction network extracts features from the image data and obtains the feature image corresponding to the image data; for any position (x, y) in the image data, the position code of the image data is obtained using the position acquisition algorithm based on the Sinusoidal function. The position acquisition algorithm includes: , where i is the dimension index, d model is the feature dimension of the feature extraction model, F(x) is the feature image output by the convolution layer, and F encoded It is a geometric feature map.

4. The unsupervised dual-target positioning method based on geometric consistency of ocean environment according to claim 1, characterized in that: Obtain the geometric consistency loss, photometric consistency loss, and feature alignment loss in the geometric feature map, and use the overall loss function to obtain the overall loss parameters including: Use the geometric loss algorithm to obtain the geometric consistency loss in the geometric feature map. The geometric loss algorithm is: , where L geo is the geometric consistency loss, N is the number of feature points in the geometric feature map, and F IR (i) is the pixel value of the infrared image of the i-th feature point in the geometric feature map, F VIS (i) is the pixel value of the visible light image of the i-th feature point in the geometric feature map; The photometric loss algorithm is used to obtain the photometric consistency loss in the geometric feature map. The photometric loss algorithm is: , where L photometric is the photometric consistency loss, I IR (i) is the pixel value of the i-th feature point in the infrared image, I VIS is the pixel value of the i-th feature point in the visible light image; The feature loss algorithm is used to obtain the feature alignment loss in the geometric feature map. The feature loss algorithm is: , where L cosine is the feature alignment loss.

5. The unsupervised dual-target positioning method based on geometric consistency of ocean environment according to claim 4 is characterized in that: Obtaining the geometric consistency loss, photometric consistency loss, and feature alignment loss in the geometric feature map, and using the overall loss function to obtain the overall loss parameters also includes: Use the overall loss function to obtain the joint optimization loss. The overall loss function is: L total =γ1×L geo +γ2×L cosine +γ3×L photometric , where L total To jointly optimize the loss, γ1 to γ3 are the first loss weight, the second loss weight, and the third loss weight, respectively.

6. The unsupervised dual-target positioning method based on geometric consistency of ocean environment according to claim 1, characterized in that: The geometric transformation matrix T is obtained based on unsupervised training, and the optimal parameters of the geometric transformation matrix T are obtained based on the overall loss parameter. The parameters in the feature extraction network are optimized using a backpropagation algorithm and gradient descent optimization methods until the infrared image in the feature extraction model is aligned with the features of its corresponding visible light image after transformation. The parameter optimization process is unsupervised learning. When the infrared image in the feature extraction model can be aligned with the features of its corresponding visible light image after transformation, the parameters in the feature extraction network at this time are recorded as the geometric transformation matrix T; the minimum loss of the infrared features and visible light features after transformation is obtained based on the joint optimization loss, and the parameters of the geometric transformation matrix T at this time are recorded as the optimal parameters.

7. The unsupervised dual-target positioning method based on geometric consistency of ocean environment according to claim 1, characterized in that: Calibration of binocular system parameters for infrared and visible light images includes: The calibration method of internal parameters is as follows: based on the calibration algorithm in computer vision, the internal parameters of the camera when shooting and analyzing the image pair are obtained; The calibration method of the external parameters is: use the coordinate transformation algorithm to transform the coordinates in the visible light image and align the coordinates of the infrared image with the coordinates of the visible light image; the coordinate transformation algorithm is: X hw =T×X kj , where X hw is the coordinate in the infrared image, X kj are the coordinates in the visible light image.

Images