Satellite navigation positioning method and system based on deep learning enhancement

Abstract

The application discloses a satellite navigation positioning method and system based on deep learning enhancement, and the method comprises the following steps: inputting a sky gray image into a multi-scale space frequency domain perception model to obtain a sky segmentation image containing shielding information; projecting the position of a satellite in a three-dimensional physical space into the sky segmentation image to obtain a sky satellite perception image; inputting the sky satellite perception image into a deep separable spatial convolution network to obtain spatial discrimination features; fusing GNSS signals and the spatial discrimination features to obtain a fusion feature vector; inputting the fusion feature vector into a multi-head attention-based long short-term memory network model to obtain a navigation signal type discrimination result; dynamically adjusting the weight coefficient of the GNSS signals according to the navigation signal type discrimination result, and performing positioning calculation to obtain accurate positioning information; and the application dynamically identifies and suppresses NLOS signals, thereby improving the positioning performance and accuracy of the positioning system in a complex urban environment.

Description

Technical Field

[0001] This invention belongs to the field of satellite positioning technology, specifically relating to a satellite navigation and positioning method and system based on deep learning enhancement. Background Technology

[0002] The Global Navigation Satellite System (GNSS) provides high-precision positioning, navigation, and timing (PNT) services to users worldwide using satellite signals. However, in typical complex environments such as urban canyons, dense high-rise building areas, dense forests, and under viaducts, GNSS positioning performance still faces significant challenges. Obstacles such as buildings and trees in these environments can block the direct path of satellite signals and cause signal reflection and diffraction, making it difficult for receivers to consistently and stably obtain high-precision positioning information.

[0003] In complex urban environments, received signals can generally be categorized into three types: line-of-sight (LOS) signals, non-line-of-sight (NLOS) signals, and multipath propagation signals. NLOS signals, composed entirely of reflection or diffraction paths, are prone to significant ranging errors, sometimes exceeding 10 meters, and have become one of the main factors affecting positioning accuracy. In contrast, multipath signals typically still contain some LOS signal components, but the errors they introduce are relatively smaller. Therefore, accurately identifying and suppressing the influence of NLOS signals is crucial for improving GNSS positioning performance in complex urban environments.

[0004] With the development of artificial intelligence, machine learning, especially deep learning methods, has provided a new approach to NLOS signal recognition. These methods model LOS / NLOS signal recognition as a classification problem, relying on large-scale data to train highly expressive models that can capture the nonlinear characteristics of signals in complex environments, thus overcoming the limitations of traditional methods to some extent. However, existing deep learning methods often focus on single-dimensional feature extraction, or only consider spatial features, or only rely on temporal features, failing to effectively integrate spatiotemporal information. They struggle to model the dependencies between satellite observations at different times and the spatial correlations between different satellites at the same time, resulting in poor real-time performance and weak generalization ability in dynamic environments. This leads to frequent misclassification or missed recognition of NLOS signals, failing to meet the positioning needs of complex dynamic scenarios. Summary of the Invention

[0005] This invention provides a satellite navigation and positioning method and system based on deep learning enhancement, which aims to achieve dynamic identification and suppression of NLOS signals, reduce the impact of NLOS signals on positioning performance, and thus improve the real-time performance, generalization ability and accuracy of the positioning system in complex urban environments.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The first aspect of this invention provides a satellite navigation and positioning method based on deep learning enhancement, comprising:

[0008] A real-world sky image is obtained by acquiring sky view information above the positioning terminal using a fisheye camera; the real-world sky image is then processed by grayscale conversion and blurring to obtain a grayscale sky image.

[0009] A grayscale sky image is input into the multi-scale spatial frequency domain perception model to obtain a sky segmentation image containing occlusion information; the satellite's position in three-dimensional physical space is projected onto the sky segmentation image to obtain a sky satellite perception image; the sky satellite perception image is input into a depth-separable spatial convolutional network to obtain spatial discrimination features.

[0010] The positioning terminal acquires GNSS signals, and obtains a fused feature vector by fusing GNSS signals and spatial discrimination features; the fused feature vector is then input into a long short-term memory network model based on multi-head attention to obtain the navigation signal type discrimination result;

[0011] The weighting coefficients of the GNSS signal are dynamically adjusted based on the navigation signal type identification result, and positioning calculation is performed to obtain accurate positioning information.

[0012] Furthermore, the grayscale sky image is input into the multi-scale spatial frequency domain perception model to obtain a sky segmentation image containing occlusion information. The specific process is as follows:

[0013] The multi-scale spatial frequency domain perception model includes a frequency domain channel attention (FCAM) ​​module, a frequency domain global spatial attention (FGSA) module, a multi-scale feature extraction (MSFE) module, and a mixed-dimensional convolution module;

[0014] The input of the first convolutional layer is used to receive the grayscale image of the sky; the output of the first convolutional layer is connected to the frequency domain channel attention (FCAM) ​​module, the multi-scale feature extraction (MSFE) module, and the mixed-dimensional convolution module.

[0015] The output of the Frequency Domain Channel Attention (FCAM) ​​module is connected to the second convolutional layer; element-wise multiplication is performed on the output features of the Frequency Domain Channel Attention (FCAM) ​​module and the second convolutional layer to perform frequency domain enhancement, and the results are input to the Frequency Domain Global Spatial Attention (FGSA) module to obtain frequency domain global attention features;

[0016] The first-level and second-level brightness features are output by the multi-scale feature extraction (MSFE) module and the mixed-dimensional convolution module;

[0017] The sky segmentation image is obtained by concatenating the frequency domain global attention features, the sky grayscale image, the first-level brightness features, and the second-level brightness features and then inputting them into the third convolutional layer.

[0018] In this invention, the first, second, and third convolutional layers all use 1×1 convolutional kernels.

[0019] Furthermore, the hybrid-dimensional convolution module includes a vertical convolution kernel, a horizontal convolution kernel, and a global spatial convolution kernel;

[0020] The output of the first convolutional layer is connected to a vertical convolutional kernel, a horizontal convolutional kernel, and a global spatial convolutional kernel; the output features of the vertical convolutional kernel, the horizontal convolutional kernel, and the global spatial convolutional kernel are concatenated to form the second-level brightness feature.

[0021] In this invention, the vertical convolution kernel, the horizontal convolution kernel, and the global spatial convolution kernel are 31×1 convolution kernel, 1×31 convolution kernel, and 31×31 convolution kernel, respectively.

[0022] Furthermore, the satellite's position in three-dimensional physical space is projected onto the sky segmentation image to obtain a sky-satellite perception image, specifically including:

[0023] The satellite's position in the three-dimensional physical space is transformed to the fisheye camera coordinate system using a rotation matrix R and a translation vector D, thus obtaining the satellite's position in the fisheye camera coordinate system. The formula is as follows:

[0024]

[0025] In the formula, T represents the matrix transpose; These are the satellite position coordinates in the fisheye camera coordinate system; These are the satellite's position coordinates in three-dimensional physical space.

[0026] The incident angle and azimuth angle of the satellite relative to the fisheye camera are calculated based on the satellite's position in the fisheye camera coordinate system. The formula is as follows:

[0027]

[0028]

[0029]

[0030] In the formula, The radial distance of the satellite in the camera coordinate system in normalized coordinates; The angle of incidence of the satellite relative to the fisheye camera; This is the azimuth angle of the satellite relative to the fisheye camera;

[0031] The principal point eccentricity between the image point formed by the satellite and the principal point on the image plane is calculated using the incident angle of the satellite relative to the fisheye camera. The principal point of the fisheye is the intersection of the camera's optical axis and the image plane.

[0032] Through the principal point eccentricity The pixel coordinates of the satellite are calculated using the azimuth angle of the satellite relative to the fisheye camera, expressed by the following formula:

[0033]

[0034]

[0035]

[0036]

[0037] In the formula, The physical coordinates of the satellite on the image plane; Principal point coordinates; These are the pixel coordinates of the satellite; The focal length is along the x-axis. The focal length is in the y-axis direction.

[0038] Furthermore, the principal point eccentricity between the image point formed by the satellite on the image plane and the principal point is calculated using the incident angle of the satellite relative to the fisheye camera. Specifically, it includes:

[0039]

[0040] In the formula, The effective focal length of a fisheye camera. Denotes the coefficient of the j-th polynomial; For the Taylor series; This is the angle of incidence of the satellite relative to the fisheye camera.

[0041] Furthermore, the sky satellite perception images are input into a deep separable spatial convolutional network to obtain spatial discriminative features, specifically including:

[0042] The depthwise separable spatial convolutional network includes a separating layer and a gated convolutional block;

[0043] The separation layer decomposes the satellite-sensed image of the sky according to feature scale to obtain high-frequency detail features and low-frequency global features. These high-frequency detail features and low-frequency global features are then input into the gated convolutional block to obtain high-frequency key features and low-frequency key features, expressed as follows:

[0044]

[0045]

[0046] In the formula, These are intermediate features obtained by normalizing high-frequency detail features or low-frequency global features; For input gated convolutional blocks, use either high-frequency detail features or low-frequency global features. For layer normalization operation; , and Represented as linear transformation weights; This is the convolution operation function; Use the Sigmoid activation function; These are the high-frequency or low-frequency key features output by the gated convolutional block. This is element-wise multiplication;

[0047] Spatial discriminant features are obtained by splicing together high-frequency detail features, low-frequency global features, high-frequency key features, and low-frequency key features.

[0048] Furthermore, the fused feature vector is input into a long short-term memory network model based on multi-head attention to obtain the navigation signal type discrimination result; specifically including:

[0049] Long Short-Term Memory (LSTM) network models based on multi-head attention include LSTM neural network modules and multi-head attention modules;

[0050] The fused feature vector is reshaped to obtain a two-dimensional fused feature tensor; the fused feature tensor is input into the multi-head attention module, and the fused feature tensor is mapped to a query vector Q, a key vector K, and a value vector V through three independent fully connected layers; the query vector Q, key vector K, and value vector V are split into m sub-vectors and parallel attention is calculated to obtain m sets of sub-vectors; m is the preset number of attention heads in the multi-head attention module; scaling dot product attention is independently performed on each set of sub-vectors to obtain m sets of spatial correlation sub-features; the m sets of spatial correlation sub-features are concatenated according to the channel dimension to obtain spatial correlation features;

[0051] Spatial correlation features are input into the first discard layer to obtain spatial correlation deduplication features; the spatial correlation deduplication features and the fused feature tensor are sequentially concatenated, normalized, and reshaped to obtain a three-dimensional spatial correlation tensor; the spatial correlation tensor is input into the LSTM neural network module to obtain spatiotemporal correlation features;

[0052] The spatiotemporal correlation features are input into the second discard layer to obtain the spatiotemporal correlation deduplication features. The spatiotemporal correlation deduplication features and the spatiotemporal correlation features are then concatenated and normalized to obtain the navigation signal type discrimination result.

[0053] Furthermore, the spatial correlation tensor is input into the LSTM neural network module to obtain spatiotemporal correlation features, specifically including:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] In the formula, The output features of the forget gate in the LSTM neural network module; The output features of the input gates in the LSTM neural network module; The output features of the output gates in the LSTM neural network module; Use the Sigmoid activation function; , , and For linear transformation weights; , , and This represents the bias term of the weight matrix; The short-term memory features of the LSTM neural network module at time t−1; The current short-term memory features output by the LSTM neural network module are denoted as spatiotemporal correlation features; The spatial input features of the LSTM neural network module at time t; It is the hyperbolic tangent activation function; Based on the current spatial input features Temporal features generated by the short-term memory features ht−1 from the previous moment; and These are the long-term memory features at time t and t-1, respectively.

[0061] Furthermore, the training process for the Long Short-Term Memory (LSTM) network model includes:

[0062] Training samples are constructed using spatial discriminative features output from a depthwise separable spatial convolutional network and GNSS signals; real LOS / NLOS signal labels are assigned to the training samples based on fisheye camera image segmentation and satellite projection results.

[0063] Spatiotemporal correlation features are obtained by training a long short-term memory network model using training samples. Based on spatiotemporal correlation characteristics The training loss is calculated using the real LOS / NLOS (line-of-sight / non-line-of-sight) signal labels, expressed as follows:

[0064]

[0065] In the formula, The training loss for the Long Short-Term Memory (LSTM) network model; For actual LOS / NLOS signal tags;

[0066] The parameters of the Long Short-Term Memory (LSTM) network model are optimized based on the training loss using the gradient descent algorithm. The training process of the LSM network model is repeated iteratively until the maximum number of iterations is reached, at which point the trained LSM network model is output.

[0067] A second aspect of the present invention provides a satellite navigation and positioning system based on deep learning enhancement, comprising:

[0068] The image acquisition unit acquires a real-world sky image by using a fisheye camera to capture sky view information above the positioning terminal; and performs grayscale and blurring processing on the real-world sky image to obtain a grayscale sky image.

[0069] The feature extraction unit inputs a grayscale sky image into the multi-scale spatial frequency domain perception model to obtain a sky segmentation image containing occlusion information; projects the satellite's position in three-dimensional physical space onto the sky segmentation image to obtain a sky satellite perception image; and inputs the sky satellite perception image into a depth-separable spatial convolutional network to obtain spatial discrimination features.

[0070] The signal discrimination unit acquires GNSS signals from the positioning terminal, obtains a fused feature vector by fusing GNSS signals and spatial discrimination features, and inputs the fused feature vector into a long short-term memory network model based on multi-head attention to obtain the navigation signal type discrimination result.

[0071] The navigation and positioning unit dynamically adjusts the weighting coefficients of the GNSS signal based on the navigation signal type discrimination result, and performs positioning calculations to obtain accurate positioning information.

[0072] A third aspect of the present invention provides an electronic terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the satellite navigation and positioning method of the first aspect.

[0073] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0074] This invention inputs a grayscale sky image into the multi-scale spatial frequency domain perception model to obtain a sky segmentation image containing occlusion information; it captures brightness features from fine edges to wide areas through multi-scale convolution, and combines a frequency domain attention mechanism to globally model the brightness distribution, avoiding misjudging occlusions due to local brightness differences in traditional segmentation algorithms; it accurately distinguishes between visible sky and occluded areas, reducing misjudgments of LOS / NLOS signals caused by errors in occlusion information extraction from the source.

[0075] This invention projects the satellite's position in three-dimensional physical space onto a segmented sky image to obtain a sky-satellite perception image; it achieves a precise correlation between the satellite's position and the distribution of sky obstructions. The image is input into a depthwise separable spatial convolutional network, which can adaptively focus on the key spatial features of the satellite and obstructions through a dynamic gating mechanism, while reducing information redundancy in deep networks and lowering computational overhead. The extracted spatial discrimination features accurately characterize the relative positional attributes of the satellite and the obstructing environment, improving the adaptability and discrimination reliability of GNSS positioning in complex environments.

[0076] This invention inputs the fused feature vector into a long short-term memory network model based on multi-head attention to obtain navigation signal type discrimination results. The multi-head attention mechanism can capture the spatial dependency relationship between different satellites at the same time, while the long short-term memory network can model the temporal changes of GNSS signals at different times, effectively reducing the false positive rate and false negative rate of NLOS signals, making the navigation signal type discrimination results more accurate. Attached Figure Description

[0077] Figure 1 This is a flowchart of the satellite navigation and positioning method provided in Embodiment 1 of the present invention;

[0078] Figure 2 This is a flowchart of the segmentation processing method based on fisheye camera images provided in Embodiment 1 of the present invention;

[0079] Figure 3 This is a structural diagram of the multi-scale spatial frequency domain sensing model provided in Embodiment 1 of the present invention;

[0080] Figure 4This is a structural diagram of the depth-separable spatial convolutional network provided in Embodiment 1 of the present invention;

[0081] Figure 5 This is a structural diagram of the long short-term memory network model with multi-head attention provided in Embodiment 1 of the present invention. Detailed Implementation

[0082] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0083] Example 1

[0084] like Figure 1 As shown, this embodiment provides a satellite navigation and positioning method based on deep learning enhancement, including:

[0085] like Figure 2 The process involves acquiring a real-world sky image by using a fisheye camera to capture sky view information above the positioning terminal. The ultra-wide-angle characteristic of the fisheye camera can cover the entire sky field of view above the positioning terminal, and compared with ordinary cameras, it can capture obstructions such as buildings and trees more comprehensively, avoiding the omission of key obstruction information due to field of view limitations. The real-world sky image is then processed by grayscale and blurring to obtain a grayscale sky image. Grayscale processing simplifies the image data dimensions and reduces redundant calculations, while blurring effectively suppresses image noise (such as tiny impurities in the sky and light fluctuations), providing cleaner input data for subsequent segmentation and improving processing efficiency and accuracy.

[0086] The sky grayscale image is input into the multi-scale spatial frequency domain perception model to obtain a sky segmentation image containing occlusion information. The specific process is as follows:

[0087] like Figure 3 As shown, the multi-scale spatial frequency domain perception model includes a frequency domain channel attention (FCAM) ​​module, a frequency domain global spatial attention (FGSA) module, a multi-scale feature extraction (MSFE) module, and a mixed-dimensional convolution module;

[0088] The input of the first convolutional layer is used to receive the grayscale image of the sky; the output of the first convolutional layer is connected to the frequency domain channel attention (FCAM) ​​module, the multi-scale feature extraction (MSFE) module, and the mixed-dimensional convolution module.

[0089] The output of the Frequency Domain Channel Attention (FCAM) ​​module is connected to the second convolutional layer; element-wise multiplication is performed on the output features of the Frequency Domain Channel Attention (FCAM) ​​module and the second convolutional layer to perform frequency domain enhancement, and the results are input to the Frequency Domain Global Spatial Attention (FGSA) module to obtain frequency domain global attention features;

[0090] The first-level brightness features are output by the Multi-Scale Feature Extraction (MSFE) module;

[0091] The second-level brightness features are output by the mixed-dimensional convolutional module, specifically including:

[0092] The hybrid-dimensional convolution module includes a vertical convolution kernel, a horizontal convolution kernel, and a global space convolution kernel;

[0093] The output of the first convolutional layer is connected to a vertical convolutional kernel, a horizontal convolutional kernel, and a global spatial convolutional kernel; the output features of the vertical convolutional kernel, the horizontal convolutional kernel, and the global spatial convolutional kernel are concatenated to form the second-level brightness feature.

[0094] The sky segmentation image is obtained by concatenating the frequency domain global attention features, the sky grayscale image, the first-level brightness features, and the second-level brightness features and then inputting them into the third convolutional layer.

[0095] In this embodiment, the first, second, and third convolutional layers all use 1×1 convolutional kernels; the vertical convolutional kernel, the horizontal convolutional kernel, and the global spatial convolutional kernel are 31×1, 1×31, and 31×31 convolutional kernels, respectively.

[0096] The multi-scale spatial frequency domain perception model captures brightness features from fine edges to wide areas through multi-scale convolution, and combines a frequency domain attention mechanism to globally model the brightness distribution, avoiding the misjudgment of occluded objects by traditional segmentation algorithms due to local brightness differences.

[0097] The satellite's position in three-dimensional physical space is projected onto a segmented sky image to obtain a sky-satellite perception image, specifically including:

[0098] The satellite's position in the three-dimensional physical space is transformed to the fisheye camera coordinate system using a rotation matrix R and a translation vector D, thus obtaining the satellite's position in the fisheye camera coordinate system. The formula is as follows:

[0099]

[0100] In the formula, T represents the matrix transpose; These are the satellite position coordinates in the fisheye camera coordinate system; These are the satellite's position coordinates in three-dimensional physical space.

[0101] The incident angle and azimuth angle of the satellite relative to the fisheye camera are calculated based on the satellite's position in the fisheye camera coordinate system. The formula is as follows:

[0102]

[0103]

[0104]

[0105] In the formula, The radial distance of the satellite in the camera coordinate system in normalized coordinates; The angle of incidence of the satellite relative to the fisheye camera; This is the azimuth angle of the satellite relative to the fisheye camera;

[0106] The principal point eccentricity between the image point formed by the satellite and the principal point on the image plane is calculated using the incident angle of the satellite relative to the fisheye camera. The formula is as follows:

[0107]

[0108] In the formula, The effective focal length of a fisheye camera. Denotes the coefficient of the j-th polynomial; For the Taylor series; The angle of incidence of the satellite relative to the fisheye camera; The principal point eccentricity is the distance between the image point formed by the satellite on the image plane and the principal point; the fisheye principal point is the intersection of the camera optical axis and the image plane.

[0109] Through the principal point eccentricity The pixel coordinates of the satellite are calculated using the azimuth angle of the satellite relative to the fisheye camera, expressed by the following formula:

[0110]

[0111]

[0112]

[0113]

[0114] In the formula, The physical coordinates of the satellite on the image plane; Principal point coordinates; These are the pixel coordinates of the satellite; The focal length is along the x-axis. The focal length is in the y-axis direction.

[0115] Sky satellite-sensing images are input into a depthwise separable spatial convolutional network to obtain spatial discriminative features, specifically including:

[0116] like Figure 4 As shown, the depth-separable spatial convolutional network includes a separation layer and gated convolutional blocks;

[0117] The separation layer decomposes the satellite-sensed image of the sky according to feature scale to obtain high-frequency detail features and low-frequency global features. These high-frequency detail features and low-frequency global features are then input into the gated convolutional block to obtain high-frequency key features and low-frequency key features, expressed as follows:

[0118]

[0119]

[0120] In the formula, These are intermediate features obtained by normalizing high-frequency detail features or low-frequency global features; For input gated convolutional blocks, use either high-frequency detail features or low-frequency global features. For layer normalization operation; , and Represented as linear transformation weights; This is the convolution operation function; Use the Sigmoid activation function; These are the high-frequency or low-frequency key features output by the gated convolutional block. This is element-wise multiplication;

[0121] Spatial discriminant features are obtained by splicing together high-frequency detail features, low-frequency global features, high-frequency key features, and low-frequency key features.

[0122] This embodiment can adaptively focus on the key spatial features of satellites and obstructions through a dynamic gating mechanism, while reducing the redundancy of deep network information and lowering computational overhead. The extracted spatial discrimination features accurately characterize the relative positional attributes of satellites and obstructing environments, improving the adaptability and discrimination reliability of GNSS positioning in complex environments.

[0123] Long Short-Term Memory (LSTM) network models introduce memory units and three gating mechanisms in the hidden layers: forget gate, input gate, and output gate. The gating functions recursively adjust the hidden layer data. These gating functions generate values ​​between 0 and 1 using the sigmoid activation function, controlling the degree of forgetting historical information, the proportion of new information being incorporated, and the output content at the current moment, thereby achieving recursive regulation of the cell state. The training parameters of the LTM network model include the weights and thresholds of the gating functions and memory units, effectively preventing gradient vanishing and gradient exploding during training. The trainable parameters of the LTM network model mainly include the weights and biases corresponding to each gating function and memory unit, which are optimized using the gradient descent algorithm, ultimately enabling the network to effectively capture long-term dependencies in time series data.

[0124] The training process of a long short-term memory network model based on multi-head attention specifically includes:

[0125] Training samples are constructed using spatial discriminative features output from a depthwise separable spatial convolutional network and GNSS signals; real LOS / NLOS signal labels are assigned to the training samples based on fisheye camera image segmentation and satellite projection results.

[0126] Spatiotemporal correlation features are obtained by training a long short-term memory network model using training samples. Based on spatiotemporal correlation characteristics The training loss is calculated using the real LOS / NLOS (line-of-sight / non-line-of-sight) signal labels, expressed as follows:

[0127]

[0128] In the formula, The training loss for the Long Short-Term Memory (LSTM) network model; For actual LOS / NLOS signal tags;

[0129] The parameters of the Long Short-Term Memory (LSTM) network model are optimized based on the training loss using the gradient descent algorithm. The training process of the LSM network model is repeated iteratively until the maximum number of iterations is reached, at which point the trained LSM network model is output.

[0130] The positioning terminal acquires GNSS signals, and a fused feature vector is obtained by fusing the GNSS signals and spatial discrimination features, thereby suppressing interference and improving positioning accuracy. The fused feature vector is then input into a pre-trained long short-term memory network model to obtain the navigation signal type discrimination result, specifically including:

[0131] like Figure 5 As shown, the long short-term memory network model based on multi-head attention includes an LSTM neural network module and a multi-head attention module;

[0132] The fused feature vector is reshaped to obtain a two-dimensional fused feature tensor; the fused feature tensor is input into the multi-head attention module, and the fused feature tensor is mapped to a query vector Q, a key vector K, and a value vector V through three independent fully connected layers; the query vector Q, key vector K, and value vector V are split into m sub-vectors and parallel attention is calculated to obtain m sets of sub-vectors; m is the preset number of attention heads in the multi-head attention module; scaling dot product attention is independently performed on each set of sub-vectors to obtain m sets of spatial correlation sub-features; the m sets of spatial correlation sub-features are concatenated according to the channel dimension to obtain spatial correlation features;

[0133] Spatial correlation features are input into the first discard layer to obtain spatial correlation deduplication features; the spatial correlation deduplication features and the fused feature tensor are sequentially concatenated, normalized, and reshaped to obtain a three-dimensional spatial correlation tensor; the spatial correlation tensor is input into the LSTM neural network module to obtain spatiotemporal correlation features; the expression formula is:

[0134]

[0135]

[0136]

[0137]

[0138]

[0139]

[0140] In the formula, The output features of the forget gate in the LSTM neural network module; The output features of the input gates in the LSTM neural network module; The output features of the output gates in the LSTM neural network module; Use the Sigmoid activation function; , , and For linear transformation weights; , , and This represents the bias term of the weight matrix; The short-term memory features of the LSTM neural network module at time t−1; The current short-term memory features output by the LSTM neural network module are denoted as spatiotemporal correlation features; The spatial input features of the LSTM neural network module at time t; It is the hyperbolic tangent activation function; Based on the current spatial input features Temporal features generated by the short-term memory features ht−1 from the previous moment; and These are the long-term memory features at time t and t-1, respectively.

[0141] The spatiotemporal correlation features are input into the second discard layer to obtain the spatiotemporal correlation deduplication features. The spatiotemporal correlation deduplication features and the spatiotemporal correlation features are then concatenated and normalized to obtain the navigation signal type discrimination result.

[0142] In this embodiment, the multi-head attention mechanism can capture the spatial dependencies between different satellites at the same time, while the long short-term memory network can model the temporal changes of GNSS signals at different times, effectively reducing the false positive and false negative rates of NLOS signals and making the navigation signal type identification results more accurate.

[0143] The weighting coefficients of GNSS signals are dynamically adjusted based on the navigation signal type discrimination results, and positioning calculations are performed to obtain accurate positioning information. In this embodiment, reliable LOS signals are given high weights, while NLOS signals with significant errors are given low weights (or are eliminated). The interference of NLOS signals is directly suppressed in the positioning calculation stage, which solves the core pain point of "fixed weights leading to NLOS error accumulation" in traditional positioning and greatly improves the accuracy of positioning calculations.

[0144] Example 2

[0145] This embodiment provides a satellite navigation and positioning system based on deep learning enhancement. The satellite navigation and positioning system is used to execute the satellite navigation and positioning method described in Embodiment 1. The satellite navigation and positioning system includes:

[0146] The image acquisition unit acquires a real-world sky image by using a fisheye camera to capture sky view information above the positioning terminal; and performs grayscale and blurring processing on the real-world sky image to obtain a grayscale sky image.

[0147] The feature extraction unit inputs a grayscale sky image into the multi-scale spatial frequency domain perception model to obtain a sky segmentation image containing occlusion information; projects the satellite's position in three-dimensional physical space onto the sky segmentation image to obtain a sky satellite perception image; and inputs the sky satellite perception image into a depth-separable spatial convolutional network to obtain spatial discrimination features.

[0148] The signal discrimination unit acquires GNSS signals from the positioning terminal, obtains a fused feature vector by fusing GNSS signals and spatial discrimination features, and inputs the fused feature vector into a long short-term memory network model based on multi-head attention to obtain the navigation signal type discrimination result.

[0149] The navigation and positioning unit dynamically adjusts the weighting coefficients of the GNSS signal based on the navigation signal type discrimination result, and performs positioning calculations to obtain accurate positioning information.

[0150] Example 3

[0151] This embodiment provides an electronic terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the satellite navigation and positioning method described in Embodiment 1.

[0152] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0154] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0155] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0156] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A satellite navigation and positioning method based on deep learning enhancement, characterized in that, include: A real-world image of the sky is obtained by capturing sky view information above the positioning terminal using a fisheye camera; A grayscale sky image is obtained by performing grayscale and blurring processing on a real-world sky image; A sky grayscale image is input into a multi-scale spatial frequency domain sensing model to obtain a sky segmentation image containing occlusion information; the satellite's position in three-dimensional physical space is projected onto the sky segmentation image to obtain a sky satellite sensing image; Sky satellite-sensing images are input into a depthwise separable spatial convolutional network to obtain spatial discriminative features, specifically including: The depth-separable spatial convolutional network includes a separating layer and a gated convolutional block; The separation layer decomposes the satellite-sensed image of the sky according to feature scale to obtain high-frequency detail features and low-frequency global features. These high-frequency detail features and low-frequency global features are then input into the gated convolutional block to obtain high-frequency key features and low-frequency key features. The formula is as follows: ； ； In the formula, These are intermediate features normalized from high-frequency detail features or low-frequency global features; For input gated convolutional blocks, use either high-frequency detail features or low-frequency global features. For layer normalization operation; , and Represented as linear transformation weights; This is the convolution operation function; Use the Sigmoid activation function; These are the high-frequency or low-frequency key features output by the gated convolutional block. This is element-wise multiplication; Spatial discriminant features are obtained by concatenating high-frequency detail features, low-frequency global features, high-frequency key features, and low-frequency key features; The positioning terminal acquires GNSS signals, and obtains a fused feature vector by fusing GNSS signals and spatial discrimination features; the fused feature vector is then input into a long short-term memory network model based on multi-head attention to obtain the navigation signal type discrimination result; The weighting coefficients of the GNSS signal are dynamically adjusted based on the navigation signal type identification result, and positioning calculation is performed to obtain accurate positioning information.

2. The satellite navigation and positioning method according to claim 1, characterized in that, The sky grayscale image is input into a multi-scale spatial frequency domain sensing model to obtain a sky segmentation image containing occlusion information. The specific process is as follows: The multi-scale spatial frequency domain perception model includes a frequency domain channel attention module, a frequency domain global spatial attention module, a multi-scale feature extraction module, and a hybrid dimension convolution module. The input of the first convolutional layer is used to receive the grayscale image of the sky; the output of the first convolutional layer is connected to the frequency domain channel attention module, the multi-scale feature extraction module, and the mixed-dimensional convolution module. The output of the frequency domain channel attention module is connected to the second convolutional layer; element-wise multiplication is performed on the output features of the frequency domain channel attention module and the second convolutional layer to enhance the frequency domain, and the results are input to the frequency domain global spatial attention module to obtain frequency domain global attention features; The first-level brightness feature and the second-level brightness feature are output by the multi-scale feature extraction module and the mixed-dimensional convolution module; The sky segmentation image is obtained by concatenating the frequency domain global attention features, the sky grayscale image, the first-level brightness features, and the second-level brightness features and then inputting them into the third convolutional layer.

3. The satellite navigation and positioning method according to claim 2, characterized in that, The hybrid-dimensional convolution module includes a vertical convolution kernel, a horizontal convolution kernel, and a global space convolution kernel; The output of the first convolutional layer is connected to a vertical convolutional kernel, a horizontal convolutional kernel, and a global spatial convolutional kernel; the output features of the vertical convolutional kernel, the horizontal convolutional kernel, and the global spatial convolutional kernel are concatenated to form the second-level brightness feature.

4. The satellite navigation and positioning method according to claim 1, characterized in that, The satellite's position in three-dimensional physical space is projected onto a segmented sky image to obtain a sky-satellite perception image, specifically including: The satellite's position in the three-dimensional physical space is transformed to the fisheye camera coordinate system using a rotation matrix R and a translation vector D, thus obtaining the satellite's position in the fisheye camera coordinate system. The formula is as follows: ； In the formula, T represents the matrix transpose; These are the satellite position coordinates in the fisheye camera coordinate system; These are the satellite's position coordinates in three-dimensional physical space. The incident angle and azimuth angle of the satellite relative to the fisheye camera are calculated based on the satellite's position in the fisheye camera coordinate system. The formula is as follows: ； ； ； In the formula, The radial distance of the satellite in the camera coordinate system in normalized coordinates; The angle of incidence of the satellite relative to the fisheye camera; This is the azimuth angle of the satellite relative to the fisheye camera; The principal point eccentricity between the image point formed by the satellite and the principal point on the image plane is calculated using the incident angle of the satellite relative to the fisheye camera. ; Through the principal point eccentricity The pixel coordinates of the satellite are calculated using the azimuth angle of the satellite relative to the fisheye camera, expressed by the following formula: ； ； ； ； In the formula, The physical coordinates of the satellite on the image plane; Principal point coordinates; These are the pixel coordinates of the satellite; The focal length is along the x-axis. The focal length is in the y-axis direction.

5. The satellite navigation and positioning method according to claim 4, characterized in that, The principal point eccentricity between the image point formed by the satellite and the principal point on the image plane is calculated using the incident angle of the satellite relative to the fisheye camera. Specifically, it includes: ； In the formula, The effective focal length of a fisheye camera. Denotes the coefficient of the j-th polynomial; Expand the Taylor series; This is the angle of incidence of the satellite relative to the fisheye camera.

6. The satellite navigation and positioning method according to claim 1, characterized in that, The fused feature vector is input into a long short-term memory network model based on multi-head attention to obtain the navigation signal type discrimination result; specifically including: Long Short-Term Memory (LSTM) network models based on multi-head attention include LSTM neural network modules and multi-head attention modules; The fused feature vector is reshaped to obtain a two-dimensional fused feature tensor; the fused feature tensor is input into the multi-head attention module, and the fused feature tensor is mapped to a query vector Q, a key vector K, and a value vector V through three independent fully connected layers; the query vector Q, key vector K, and value vector V are split into m sub-vectors and parallel attention is calculated to obtain m sets of sub-vectors; m is the preset number of attention heads in the multi-head attention module; scaling dot product attention is independently performed on each set of sub-vectors to obtain m sets of spatial correlation sub-features; the m sets of spatial correlation sub-features are concatenated according to the channel dimension to obtain spatial correlation features; Spatial correlation features are input into the first discard layer to obtain spatial correlation deduplication features; the spatial correlation deduplication features and the fused feature tensor are sequentially concatenated, normalized, and reshaped to obtain a three-dimensional spatial correlation tensor; the spatial correlation tensor is input into the LSTM neural network module to obtain spatiotemporal correlation features; The spatiotemporal correlation features are input into the second discard layer to obtain the spatiotemporal correlation deduplication features. The spatiotemporal correlation deduplication features and the spatiotemporal correlation features are then concatenated and normalized to obtain the navigation signal type discrimination result.

7. The satellite navigation and positioning method according to claim 6, characterized in that, The spatial correlation tensor is input into the LSTM neural network module to obtain spatiotemporal correlation features; the expression formula is: ； ； ； ； ； ； In the formula, The output features of the forget gate in the LSTM neural network module; The output features of the input gates in the LSTM neural network module; The output features of the output gates in the LSTM neural network module; Use the Sigmoid activation function; , , and For linear transformation weights; , , and This represents the bias term of the weight matrix; The short-term memory features of the LSTM neural network module at time t−1; The current short-term memory features output by the LSTM neural network module are denoted as spatiotemporal correlation features; The spatial input features of the LSTM neural network module at time t; It is the hyperbolic tangent activation function; Based on the current spatial input features Temporal features generated by the short-term memory features ht−1 from the previous moment; and These are the long-term memory features at time t and t-1, respectively.

8. A satellite navigation and positioning system based on deep learning enhancement, characterized in that, include: The image acquisition unit obtains a real-world sky image by capturing sky view information above the positioning terminal using a fisheye camera; A grayscale sky image is obtained by performing grayscale and blurring processing on a real-world sky image; The feature extraction unit inputs the grayscale sky image into a multi-scale spatial frequency domain perception model to obtain a sky segmentation image containing occlusion information; projects the satellite's position in three-dimensional physical space onto the sky segmentation image to obtain a sky satellite perception image; and inputs the sky satellite perception image into a depth-separable spatial convolutional network to obtain spatial discrimination features. The signal discrimination unit acquires GNSS signals from the positioning terminal and obtains a fused feature vector by fusing the GNSS signals and spatial discrimination features; The fused feature vectors are input into a long short-term memory network model based on multi-head attention to obtain the navigation signal type discrimination result; The navigation and positioning unit dynamically adjusts the weighting coefficients of the GNSS signal based on the navigation signal type discrimination result, and performs positioning calculations to obtain accurate positioning information; The feature extraction unit inputs the sky satellite perception image into a depth-separable spatial convolutional network to obtain spatial discriminative features, specifically including: The depth-separable spatial convolutional network includes a separating layer and a gated convolutional block; The separation layer decomposes the satellite-sensed image of the sky according to feature scale to obtain high-frequency detail features and low-frequency global features. These high-frequency detail features and low-frequency global features are then input into the gated convolutional block to obtain high-frequency key features and low-frequency key features. The formula is as follows: ； ； In the formula, These are intermediate features normalized from high-frequency detail features or low-frequency global features; For input gated convolutional blocks, use either high-frequency detail features or low-frequency global features. For layer normalization operation; , and Represented as linear transformation weights; This is the convolution operation function; Use the Sigmoid activation function; These are the high-frequency or low-frequency key features output by the gated convolutional block. This is element-wise multiplication; Spatial discriminant features are obtained by splicing together high-frequency detail features, low-frequency global features, high-frequency key features, and low-frequency key features.

9. An electronic terminal, characterized in that, It includes a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the satellite navigation and positioning method according to any one of claims 1 to 7.

Images