Indoor visible light positioning method and system in occluded scene

Abstract

The present application belongs to the field of wireless communication network, and particularly relates to a kind of indoor visible light positioning method and system in occlusion scene.The method constructs sample level normalization, double view robust training, space-time feature extraction, LED geometric graph prior modeling and multi-task joint output indoor visible light positioning network, so that the model can still maintain high positioning accuracy, robustness and inference efficiency under the condition of channel state change, transmission power fluctuation, partial LED signal occlusion or observation missing.Experimental results show that, compared with the baseline model, the present application can obtain more concentrated positioning error distribution and lower error tail under mixed signal-to-noise ratio test conditions and higher occlusion ratio conditions.Therefore, the present application has better positioning accuracy, occlusion robustness and engineering application stability in complex indoor visible light positioning scenarios.

Description

Technical Field

[0001] This invention pertains to wireless communication networks, and particularly relates to a method and system for indoor visible light positioning in obstructed environments. Background Technology

[0002] Visible Light Positioning (VLP) utilizes the combined lighting and communication features of LED lighting facilities. It collects light signals through photodiodes (PDs) at the receiving end and infers spatial location. It has advantages such as abundant spectrum resources, strong resistance to electromagnetic interference, low deployment cost, and high positioning accuracy, and is suitable for scenarios such as industrial IoT, mobile robots, warehousing and logistics, smart services, and indoor navigation.

[0003] However, real-world indoor VLP systems are not always in a state of complete, stable, and unobstructed optical link. Human presence, obstructions from pillars or equipment, changes in receiver orientation, temporary light source failures, partial link interruptions, and changes in environmental noise can all lead to missing LED signals or incomplete observations. The original TNSE_R1 paper summarizes these problems as optical link interruption, partial link failure, and incomplete optical observations—that is, incomplete optical observations caused by obstructions and dynamic environments in optical wireless communication networks.

[0004] In complex indoor dynamic occlusion scenarios, existing visible light positioning technology is difficult to adapt to incomplete light observation conditions, and has obvious technical shortcomings, specifically: Classical model-driven methods such as RSS, TOA / TDOA, and AOA have strong physical interpretability, but they heavily rely on ideal line-of-sight propagation, accurate angle / time measurements, and stable channel assumptions; once multipath, occlusion, or measurement noise occurs, the positioning performance will rapidly decline.

[0005] Learning-based VLP methods such as CNN, RNN, and Transformer can directly learn the nonlinear mapping from signal to coordinate, but most assume that the input signal is complete during the training and inference stages and lack specialized modeling for local link failures and temporally continuous occlusion.

[0006] Simple random occlusion or noise enhancement is insufficient to characterize the spatial correlation and temporal continuity of real occlusion, and cannot adequately guarantee that the model maintains a stable error distribution as the occlusion ratio increases.

[0007] For example, Chinese patent document CN119402814A discloses an indoor visible light fusion positioning method based on deep learning, including the following steps: Step 1: Collect visible light AoA information and UWB distance information data at different locations, perform data preprocessing, and then transform them into a two-dimensional matrix format for easy network input; Step 2: Construct a network model to extract features from the preprocessed data, perform feature fusion, process the fused features, and output the continuous coordinate values ​​of the acquired data; Step 3: Train the network model using samples with and without AoA data to ensure that the network model can learn effectively even with only UWB data. However, it mainly addresses the fusion positioning problem of visible light AoA information and UWB ranging information, and the positioning performance depends to some extent on the availability and deployment conditions of UWB ranging data; at the same time, its training strategy only enhances the model's adaptability to a single UWB input by using samples with and without AoA data, and does not establish an effective robust feature learning mechanism for visible light link obstruction, partial LED signal loss, incomplete observation data, and signal fluctuations in complex indoor environments. Furthermore, this method does not fully incorporate spatial geometric relationships of the light source, distance constraints, and multi-task collaborative supervision information, making it difficult to fully exploit the spatial structural features inherent in visible light signals. Therefore, when relying solely on visible light signals or with incomplete visible light observations, its positioning accuracy, robustness, and generalization ability still have room for improvement.

[0008] Therefore, there is an urgent need for a visible light positioning method that can adapt to complex indoor dynamic occlusion environments, balance positioning accuracy and robustness, and efficiently process incomplete optical observation data, so as to meet the positioning requirements of high precision and high stability indoors. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes an indoor visible light localization method and system for occluded scenarios. The method explicitly models the fixed spatial layout of the LED light source as a graph structure, extracts spatial geometric priors through a graph attention network, and modulates signal features using a gating mechanism. Simultaneously, sample-level RMS energy normalization is performed on each complex baseband sample to reduce amplitude scale differences under different channel states. Furthermore, a dual-view progressive training strategy, jointly optimizing the complete view and the occluded view, is employed to ensure stable localization performance even under conditions of communication interruption and partial signal loss.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A method for indoor visible light localization in an occluded scene includes the following steps: S1. Collect indoor visible light communication received signals and extract the original samples of complex baseband corresponding to each LED light source through pilot extraction; S2. Perform sample-level RMS normalization on the original complex baseband samples to obtain normalized complex baseband signals. Construct normalized complete views and normalized occluded views from the normalized complex baseband signals respectively. Construct a dual-view dataset with paired views. The dual-view dataset includes a training set and a validation set. S3. Construct a GeoGated-PiTNet model, which includes a feature embedding and sequence modeling module, a spatial feature extraction module, a geometric prior and gating module, and a multi-task output head module. Input the normalized complete view and the normalized occluded view into the GeoGated-PiTNet model respectively to obtain the optional angle estimate, position coordinates, and LED distance estimate for each view. Use a dual-view progressive training strategy to train the GeoGated-PiTNet model through the training set, and use the validation set to select the optimal parameters of the GeoGated-PiTNet model to obtain the optimal model. S4. Deploy the optimal model to locate the original samples of the complex baseband corresponding to each LED light source obtained through pilot extraction.

[0011] Preferably, the original sample of complex base bands mentioned in step S1 is represented as follows: (1) in, For the complex matrix carrying the original samples of the complex basis band, Q is the real part matrix, j is the imaginary part matrix, P is the pilot number, and M is the subcarrier number. Preferably, step S2, which involves performing sample-level RMS normalization, includes: a. Traverse all sampling points of the original sample of the complex baseband and calculate the single-point amplitude energy of each sampling point, as shown in the following formula; (2) in, This represents the single-point amplitude energy of the sampling point corresponding to the p-th pilot and the m-th subcarrier. Represents the real part matrix The element in the p-th row and m-th column; Let represent the element in the p-th row and m-th column of the imaginary part matrix Q; b. Calculate the sample-level RMS factor based on the single-point amplitude energy, using the following formula: (3) c. Based on the sample-level RMS factor, the I and Q components are uniformly scaled to obtain the normalized complex baseband signal, where the normalized real and imaginary parts are represented as follows: (4) in, For numerically stable terms, The normalized real part, This is the normalized imaginary part.

[0012] Preferably, step S3 specifically includes: S31. The normalized complete view / occluded view is input into the feature embedding and sequence modeling module to obtain the corresponding sequence features; the sequence features are processed by the spatial feature extraction module to output the angle feature representation and the signal features extracted by the backbone network. ; Angle feature representation outputs optional angle estimates through the multi-task output header module; S32, LED coordinates are used as fixed-space prior input to the graph attention network, and the updated LED node features are obtained through neighbor-weighted aggregation. These node features are then input into the geometric prior and gating module, along with the signal features. Element-wise modulation is used to generate enhanced features. These enhanced features are then input into the multi-task output head module, which outputs the receiver position coordinates and LED distance estimate. S33. Based on the receiver's true coordinates, receiver's angle label, LED-receiver distance label calculated from the receiver's true coordinates, and the optional angle estimates, position coordinates, and LED distance estimates corresponding to the complete view and occluded view output by the GeoGated-PiTNet model, calculate the complete view loss and the occluded view regression loss respectively to train the GeoGated-PiTNet model.

[0013] Further preferably, the feature embedding and sequence modeling module in step S31 includes Patch Embedding, Bi-LSTM, and multi-head attention; the spatial feature extraction module includes a feature sharing layer, an angle feature extraction branch, and a time-frequency-spatial feature transformation branch; and the multi-task output head module includes an angle output head, a coordinate output head, and a distance output head.

[0014] More preferably, step S31 specifically includes: The real and imaginary parts of the normalized complete / occluded view are used as dual-channel complex baseband signals. These signals are partitioned and mapped into several embedding vectors using a patch embedding operation to form a sequence representation. This sequence representation is then input into a Bi-LSTM and a multi-head attention mechanism to extract the temporal dependencies of the pilot signal in the time sampling dimension and the LED emission order dimension, resulting in sequence features containing information related to the pilot sampling order and the LED emission order. Subsequently, the sequence features are input through a feature sharing layer to the angle feature extraction layer of the angle feature extraction branch and the time-frequency-space feature transformation layer of the time-frequency-space feature transformation branch, respectively. Based on the angle feature extraction layer, angle feature representations related to the receiver's attitude change are extracted from the sequence features. These angle feature representations are then input into the angle output head in the multi-task output head, outputting a selectable angle estimate for the receiver. The time-frequency-space feature transformation layer rearranges the sequence features according to the LED emission order and time sampling order corresponding to the pilot signal, reconstructing a two-dimensional time-frequency-space feature map. A CNN extracts spatially relevant features, and the feature map output by the CNN or its pooled feature vector is used as the signal features extracted by the backbone network. .

[0015] Preferably, step S32 specifically includes: S321. The LED coordinates are used as the fixed-space prior input graph for the attention network. For node i in the LED coordinates and its neighbor node j, GAT calculates the attention score based on their encoding features, as shown in the following formula: (5) in, This represents the unnormalized attention score between node i and its neighbor node j; LeakyReLU is the activation function. This represents the learnable attention weight vector in a graph attention network. This represents the feature representation of node i after a linear transformation. This represents the feature representation of node j after a linear transformation. This represents the vector concatenation operation, which connects the features of node i and node j along their dimensions to form a joint feature; S322. Based on the attention score, calculate the attention weight of node i to its neighbor node j, using the following formula: (6) in, Let exp() represent the importance weights of node iii to each of its neighboring nodes. Let i represent the set of neighboring nodes. This represents the unnormalized attention score between node i and its neighbor node k. S323. Based on the attention weights of node i to its neighbor node j, calculate the features of the i-th LED node after the graph attention network update, as follows: (7) in, The features of the i-th LED node after the graph attention network update are: Represents a nonlinear activation function; S324. All LED node features are subjected to global average pooling to obtain spatial prior features, and spatial gating vectors are generated by the sigmoid function: (8) in, This represents the global spatial prior features corresponding to the LED array, where N is the total number of LED nodes.

[0016] (9) Where g represents the spatial gating vector generated by inputting global spatial prior features into the Sigmoid activation function; S325, Gating Vector and Backbone Network Extracted Signal Features Element-wise multiplication yields the enhanced features: (10) in, Indicates the enhanced signal characteristics, ⊙ represents the signal features extracted by the backbone network; ⊙ represents element-wise multiplication. S326. Input the enhanced features after geometric gating into the multi-task output head to obtain the receiver position coordinates and LED distance estimate.

[0017] Preferably, the dual-view progressive training strategy described in step S3 includes dual-view progressive pre-training and frozen backbone network fine-tuning. During the dual-view progressive pre-training phase, the optimization objective is represented as: (11) (12) Where θ represents all trainable parameters of the network, and X is the number of training samples. and These represent the regression losses for the full view and the occluded view, respectively. The progressive weighting coefficients increase gradually with the number of training rounds.

[0018] After each training round, the validation loss is calculated on the full validation set, and the current optimal model parameters are saved. During the backbone network fine-tuning phase, the optimization objective can be expressed as: (13) in, This indicates that the regression layer parameters for the task are being fine-tuned, while the backbone network parameters remain fixed. Fixed weights for the second-stage occlusion view loss.

[0019] This invention also provides a system for implementing the above-described indoor visible light positioning method in an obstructed scene, comprising: Data acquisition module: used to receive LED pilot signals, extract complex baseband I / Q signals, and record or generate coordinate, distance and angle labels.

[0020] Sample normalization module: used to calculate the RMS energy scaling factor for each sample and to uniformly scale the I / Q components.

[0021] Graph Prior Modeling Module: Used to encode the 3D coordinates and topological relationships of LEDs into a graph structure, and generate spatial prior features through GAT.

[0022] Geometric Gated Feature Fusion Module: Used to generate gating vectors based on spatial priors to enhance or suppress the signal features of the backbone network.

[0023] Dual-view training module: used to construct complete view and continuously occluded view, and improve occlusion robustness through two-stage progressive optimization.

[0024] Positioning output module: used to output the receiver's three-dimensional coordinates, distance estimate, and optional angle estimate.

[0025] In another aspect of the invention, an electronic device is also provided, comprising: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the indoor visible light positioning method for occluded scenes as described above.

[0026] In another aspect of the invention, a machine-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform the indoor visible light positioning method for occluded scenes as described above.

[0027] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs an indoor visible light positioning network with sample-level normalization, dual-view robust training, spatiotemporal feature extraction, LED geometry prior modeling, and multi-task joint output, so that the model can maintain high positioning accuracy, robustness and inference efficiency even under conditions of channel state changes, transmission power fluctuations, partial LED signal obstruction or observation loss.

[0028] (2) This invention employs sample-level RMS normalization to normalize the energy scale of complex baseband signals in both the complete and occluded views, ensuring that the real and imaginary parts of the input signal maintain stable amplitude scales within the sample. This reduces the differences in signal amplitude caused by different channel states, variations in transmit power, and occlusion, preventing the model from over-relying on absolute amplitude characteristics and improving the model's stability and generalization ability under different signal strength conditions.

[0029] (3) The present invention adopts a dual-view training technique that combines complete view and occluded view training, enabling the model to learn effective features under complete observation conditions and robust features under occluded observation conditions simultaneously. This improves the model's adaptability to situations such as partial LED signal loss, link interruption, and incomplete local observation, allowing the model to maintain stable positioning performance even under occluded environments.

[0030] (4) The present invention adopts a progressive occlusion loss weight adjustment technique to reduce the impact of occlusion view loss on model optimization in the early stage of training, and gradually increase the weight of occlusion view loss during training. This can avoid the model being excessively interfered with by occlusion noise or incomplete observations in the early stage of training, so that the model first learns the basic localization rules under complete signals, and then gradually enhances the robust expression ability under occlusion scenarios, thereby improving training stability.

[0031] (5) This invention employs a temporal feature extraction technique that combines Patch Embedding, Bi-LSTM, and multi-head attention mechanisms to map the normalized complex baseband signal into sequence features and extract the temporal dependencies in the pilot sampling order and LED emission order. This enhances the model's ability to express temporal variation features and long-range dependencies in visible light signals and improves the accuracy of subsequent localization feature extraction.

[0032] (6) This invention employs time-frequency-spatial feature conversion and CNN spatial feature extraction techniques to rearrange the sequence features according to the LED emission order and time sampling order corresponding to the pilot signal, reconstructing a two-dimensional time-frequency-spatial feature map, and further extracting spatially related features. This enables the one-dimensional sequence features to be converted into a two-dimensional feature expression that is more suitable for spatial structure modeling, improving the model's ability to perceive the spatial correlation of the LED array and the changing patterns of the receiver position.

[0033] (7) This invention uses a fixed geometric layout of LEDs to construct a graph prior and enhances the backbone signal features through a graph attention network and a geometric gating mechanism. This allows the physical spatial structure of the LED array to be explicitly introduced into the positioning network, so that the model is still subject to spatial geometric constraints under conditions of partial signal loss or occlusion, reducing position prediction jumps that do not conform to the physical structure and improving the rationality and continuity of the positioning results.

[0034] (8) This invention employs a multi-task output technique consisting of a coordinate output head, a distance output head, and an angle output head, simultaneously constraining the receiver's position coordinates, the distance from the receiver to each LED, and the receiver's selectable angle estimation. This enhances the model's ability to learn about spatial geometric relationships and receiver attitude changes through the synergistic effect of coordinate supervision, distance supervision, and angle supervision, thereby improving the accuracy and stability of the positioning results.

[0035] (9) This invention employs a dual-view learning approach during the training phase and a single-view forward propagation approach during the inference phase. This allows the dual-view mechanism to only require additional computation during the training phase, eliminating the need to input both the complete view and the occluded view simultaneously during actual online positioning. This improves the model's occlusion robustness while avoiding an increase in the number of model parameters and the complexity of online positioning during the inference phase, thus meeting the requirements of low complexity and high efficiency for indoor real-time positioning applications.

[0036] (10) Experimental results show that, under mixed signal-to-noise ratio test conditions and high occlusion ratio conditions, the present invention can achieve a more concentrated positioning error distribution and a lower error tail compared to the baseline model. This indicates that the present invention has better positioning accuracy, occlusion robustness and engineering application stability in complex indoor visible light positioning scenarios. Attached Figure Description

[0037] Figure 1 This is a diagram showing the overall structure of the GeoGated-PiTNet model described in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the dual-view progressive robust training strategy of Embodiment 1 of the present invention; Figure 3 This is a comparison chart of the accuracy of three models—CRNN-MLP, STAK-Ploc, and GeoGated-PiTNet—without RMS normalization processing on the input data of Embodiment 1 of the present invention. Among them, (a) and (b) are comparison charts of the cumulative distribution function (CDF) curves and box plots of the localization error of the three models on test set A, respectively; (c) and (d) are comparison charts of the cumulative distribution function (CDF) curves and box plots of the localization error of the three models on test set B, respectively. Figure 4This is a comparison chart of the accuracy range distribution of the CRNN-MLP model in Embodiment 1 of the present invention before and after performing RMS normalization on test set A and input data; Figure 5 This is a comparison chart of the accuracy range distribution of the STAK-Ploc model in Embodiment 1 of the present invention before and after performing RMS normalization on test set A and input data; Figure 6 This is a distribution of the accuracy range of the GeoGated-PiTNet model in Embodiment 1 of the present invention before and after RMS normalization on test set A; Figure 7 This is a comparison chart of the accuracy range distribution of the CRNN-MLP model in Embodiment 1 of the present invention before and after performing RMS normalization on the test set B and input data; Figure 8 This is a comparison chart of the accuracy range distribution of the STAK-Ploc model in Embodiment 1 of the present invention before and after performing RMS normalization on the test set B and input data; Figure 9 This is a distribution of the accuracy range of the GeoGated-PiTNet model in Embodiment 1 of the present invention before and after RMS normalization processing on test set B. Figure 10 The following is a comparison chart of CDF positioning accuracy after RMS normalization processing of the input data of the three models in Embodiment 1 of the present invention; wherein, (a) is a comparison chart of positioning accuracy on dataset A; and (b) is a comparison chart of positioning accuracy on dataset B. Figure 11 This is a comparison chart of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set A, 0% occlusion ratio; Figure 12 This is a comparison chart of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set A, occlusion ratio of 5%; Figure 13 This is a comparison chart of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set A, occlusion ratio of 10%; Figure 14 This is a comparison of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set A, occlusion ratio of 20%; Figure 15 This is a comparison of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set A, occlusion ratio of 30%; Figure 16This is a comparison chart of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set B, occlusion ratio of 0%; Figure 17 This is a comparison chart of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set B, occlusion ratio of 5%; Figure 18 This is a comparison chart of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set B, occlusion ratio of 10%; Figure 19 This is a comparison of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set B, occlusion ratio of 20%; Figure 20 This is a comparison of the localization accuracy of the GeoGated-PiTNet model in Embodiment 1 of the present invention under different training conditions with test set B, occlusion ratio of 30%; Figure 21 This is a spatial distribution map of the localization error in an occluded environment for the GeoGated-PiTNet model in Embodiment 1 of the present invention under the optimal training strategy; wherein, (a)-(c) are spatial distribution maps of localization error in test set A; (d)-(f) are spatial distribution maps of localization error in test set B; Figure 22 The following is a spatial distribution map of the positioning error in an occlusion environment with only a normalized baseline model in Embodiment 1 of the present invention; wherein (a)-(c) are spatial distribution maps of the positioning error in test set A; and (d)-(f) are spatial distribution maps of the positioning error in test set B. Detailed Implementation

[0038] The present invention will be further described below with reference to embodiments. The following embodiments are used to explain the present invention and do not constitute a limitation on the scope of protection.

[0039] In indoor MIMO-VLP systems, multiple LEDs serve both lighting and communication functions. The receiving device (PD) moves in space and receives pilot signals from different LEDs. Ideally, the receiver can obtain complete multi-source observations; however, in real-world scenarios, human occlusion, structural obstruction, changes in receiver posture, or temporary light source failures can cause partial optical link disruptions. In such cases, the input signal exhibits localized gaps, timing interruptions, and amplitude distribution variations. Traditional models may mistakenly identify energy attenuation caused by occlusion as a change in position.

[0040] Example 1: A method for indoor visible light localization in an occluded scene includes the following steps: S1. Collect indoor visible light communication received signals and extract the original samples of complex baseband corresponding to each LED light source through pilot extraction; The original sample of the complex base band is represented as follows: (1) in, For the complex matrix carrying the original samples of the complex basis band, Let j be the real part matrix, j be the imaginary part matrix, and Q be the imaginary part matrix. It is a set of real numbers, where P is the pilot number and M is the subcarrier number; S2. Perform sample-level RMS normalization on the original complex baseband samples to obtain normalized complex baseband signals. Construct normalized complete views and normalized occluded views from the normalized complex baseband signals respectively. Construct a dual-view dataset with paired views. The dual-view dataset includes a training set and a validation set. The sample-level RMS normalization is performed as follows: a. Traverse all sampling points of the original sample of the complex baseband and calculate the single-point amplitude energy of each sampling point; Specifically, the amplitude energy is calculated independently for each sample, using the following formula: (2) in, This represents the single-point amplitude energy of the sampling point corresponding to the p-th pilot and the m-th subcarrier. Represents the real part matrix The element in the p-th row and m-th column; Let represent the element in the p-th row and m-th column of the imaginary part matrix Q; b. Calculate the sample-level RMS factor based on the single-point amplitude energy, according to the following formula; Specifically, the mean and square root of the amplitude energy at all sampling positions of the current sample are calculated, and the sample-level RMS factor is obtained according to equation (3): (3) c. Based on the sample-level RMS factor, the I and Q components are uniformly scaled to obtain the normalized complex baseband signal.

[0041] Specifically, the I and Q components are scaled synchronously using the same RMS factor, and the normalized real and imaginary parts are obtained according to equation (4).

[0042] (4) in, For numerically stable terms, The normalized real part, This is the normalized imaginary part.

[0043] Based on the obtained normalized real and imaginary parts, a normalized complex baseband signal is output. A dual-view input is constructed based on the normalized complex baseband signal. One path retains all pilot observation information to form a normalized complete view, while the other path applies an occlusion mask according to a preset occlusion ratio or occlusion mode to generate continuous occlusion segments in the time dimension, forming a normalized occlusion view. The complete view and the occlusion view are used for subsequent feature extraction and localization model training.

[0044] Specifically, leaving the normalized complex baseband signal unchanged is called "clean"; performing a short-term continuous masking on the normalized complex baseband signal, i.e., taking a portion and turning it to 0, is called "mask". The complete view x_clean represents an unobstructed or normal reception state; the occluded view x_mask uses TemporalMask to randomly select continuous intervals in the time dimension and set them to zero or replace them with zero-mean noise, which is used to simulate timing loss caused by human occlusion, posture changes and short-term light source failure.

[0045] This normalization method has three characteristics: first, it only relies on the currently observable signal, making it suitable for practical deployment; second, it shares the same scaling factor for all locations of the same sample, preserving the spatial amplitude relationship; and third, it uses synchronous I / Q scaling to preserve the phase structure.

[0046] S3. Construct the GeoGated-PiTNet model, such as Figure 1 As shown, the GeoGated-PiTNet model includes a feature embedding and sequence modeling module, a spatial feature extraction module, a geometric prior and gating module, and a multi-task output head module. Normalized complete views and normalized occluded views are input into the GeoGated-PiTNet model to obtain the optional angle estimates, position coordinates, and LED distance estimates for each view. A dual-view progressive training strategy is adopted, training the GeoGated-PiTNet model using the training set, and selecting the optimal parameters of the GeoGated-PiTNet model using the validation set to obtain the optimal model. S31. The normalized complete view / occluded view is input into the feature embedding and sequence modeling module to obtain the corresponding sequence features; the sequence features are processed by the spatial feature extraction module to output the angle feature representation and the signal features extracted by the backbone network. ; Angle feature representation outputs optional angle estimates through the multi-task output header module; Specifically, the feature embedding and sequence modeling module includes Patch Embedding, Bi-LSTM, and multi-head attention mechanism; the spatial feature extraction module includes a feature sharing layer, an angle feature extraction branch, and a time-frequency-spatial feature transformation branch; and the multi-task output head module includes an angle output head, a coordinate output head, and a distance output head.

[0047] For the normalized complete view and the occluded view, a backbone network with shared parameters is used to extract features respectively.

[0048] Specifically, the real and imaginary parts of the normalized complete / occluded view are used as dual-channel complex baseband signals. These signals are then partitioned and mapped into several embedding vectors using a patch embedding operation to form a sequence representation. This sequence representation is input into a Bi-LSTM and a multi-head attention mechanism to extract the temporal dependencies of the pilot signal in the time sampling dimension and the LED emission order dimension, resulting in sequence features containing information related to the pilot sampling order and the LED emission order. Subsequently, the sequence features are input through a feature sharing layer to the angle feature extraction layer of the angle feature extraction branch and the time-frequency-space feature transformation layer of the time-frequency-space feature transformation branch, respectively. The angle feature extraction layer extracts angle feature representations related to the receiver's attitude change from the sequence features and inputs these angle feature representations into the angle output head in the multi-task output head, outputting a selectable angle estimate for the receiver. The time-frequency-space feature transformation layer rearranges the sequence features according to the LED emission order and time sampling order corresponding to the pilot signal, reconstructing a two-dimensional time-frequency-space feature map. A CNN extracts spatially relevant features, and the feature map output by the CNN or its pooled feature vector is used as the signal features extracted by the backbone network. .

[0049] The time-frequency-space conversion layer rearranges the sequence features according to the LED emission order and time sampling order corresponding to the pilot signal, ensuring that data with adjacent spatial or temporal relationships remain distributed in a close proximity in the two-dimensional feature map, thus preserving the spatial neighborhood information in the pilot signal. The convolutional neural network uses local convolutional kernels to extract features from the two-dimensional feature map, learning the correlation between local regions to obtain spatial correlation features that reflect different LED spatial layouts and signal propagation characteristics.

[0050] The fixed three-dimensional coordinates of the LED array are used to construct graph nodes, and a graph structure is established based on the spatial adjacency or full connectivity between LEDs.

[0051] S32, LED coordinates are used as fixed-space prior input to the graph attention network, and the updated LED node features are obtained through neighbor-weighted aggregation. These node features are then input into the geometric prior and gating module, along with the signal features. Element-wise multiplication generates enhanced features, which are then input into the multi-task output head module to output the receiver position coordinates and LED distance estimate. The LED coordinates serve as the fixed-space prior input graph attention network. For nodes i and j, GAT calculates attention coefficients based on their encoded features and obtains the node context representation through neighbor-weighted aggregation.

[0052] S321. The LED coordinates are used as the fixed-space prior input graph for the attention network. For node i in the LED coordinates and its neighbor node j, GAT calculates the attention score based on their encoding features, as shown in the following formula: (5) in, represents the unnormalized attention score between node i and its neighbor node j, used to measure the spatial correlation between LED node j and LED node i. The larger the value, the greater the contribution of node j to the feature update of node i; j represents the neighbor node index of node i; LeakyReLU is the activation function. represents the learnable attention weight vector in the graph attention network, used to calculate the importance of the concatenated features. Its parameters are also automatically optimized through model training, and T is the transpose. This represents the feature representation of node i after a linear transformation. This represents the feature representation of node j after a linear transformation. The vector concatenation operation is represented, which connects the features of node i and node j along the dimension to form a joint feature, which is used to describe the spatial relationship between the two LED nodes; W represents the learnable linear transformation weight matrix. The input feature vector representing neighbor node j can be composed of LED spatial coordinates, geometric prior information, or distance-related features. This is achieved by analyzing the features of neighbor nodes. By performing attention-weighted aggregation, graph attention networks can adaptively model the spatial relationships between different LED nodes.

[0053] Formula (1) represents the spatial correlation strength between nodes i and j, where the nodes correspond to the position coordinates of fixed indoor LEDs. After linear mapping, the geometric correlation between the two LEDs is calculated using an attention mechanism. The greater the correlation strength, the more important the spatial relationship between the two LEDs is to the current position reasoning.

[0054] S322. Based on the attention score, calculate the attention weight of node i to its neighbor node j, using the following formula: (6) in, Let be the attention weights of node i for each neighbor node j, used to characterize the importance of neighbor node j in the feature aggregation process of node i; This represents an exponential mapping. The formula as a whole uses softmax normalization to convert the attention scores of each neighbor node into normalized weights. This represents the unnormalized attention score between node i and its neighbor node k. Let represent the set of neighboring nodes of node i, and let represent the set of LED nodes that are spatially associated with the i-th LED node. Formula (2) represents the importance weight of node i to each neighbor node. This weight is obtained by Softmax normalization. The sum of the weights of each neighbor is 1, which is used to reflect the contribution of different LED spatial priors to the representation of the current node.

[0055] S323. Based on the attention weights of node i to its neighbor node j, calculate the features of the i-th LED node after the graph attention network update, as follows: (7) in, The features of the i-th LED node after the graph attention network update are: Represents a nonlinear activation function; Formula (3) represents the node representation after fusing neighbor information. It not only includes the spatial coordinate information of the current LED itself, but also incorporates the geometric topological relationship between adjacent LEDs, thus forming a spatial prior feature with global spatial constraints.

[0056] S324. All LED node features are subjected to global average pooling to obtain spatial prior features, and spatial gating vectors are generated by the sigmoid function: (8) in, This represents the global spatial prior features corresponding to the LED array, where N is the total number of LED nodes.

[0057] Formula (8) represents the global average pooling of the updated features of all LED nodes to obtain the global spatial prior features corresponding to the entire LED array. This feature comprehensively describes the spatial topological relationships and geometric distribution information between all LED nodes. It no longer reflects the local features of a single LED, but forms a global spatial representation of the entire indoor LED layout, providing a unified spatial prior constraint for subsequent signal feature enhancement.

[0058] (9) Where g represents the spatial gating vector generated by inputting global spatial prior features into the Sigmoid activation function; The gating vector's elements range from 0 to 1, representing the importance of different feature channels. When the spatial geometry of a channel is more consistent with the current scene, its gating value approaches 1; conversely, when the channel is more likely to be affected by noise, occlusion, or signal loss, its gating value approaches 0. Therefore, the gating vector can adaptively adjust the weights of different channel features based on the LED's spatial geometry prior.

[0059] S325, Gating Vector and Backbone Network Extracted Signal Features Element-wise multiplication yields enhanced features.

[0060] (10) in, Indicates the enhanced signal characteristics, ⊙ represents the signal features extracted by the backbone network, and ⊙ represents element-wise multiplication.

[0061] Specifically, the gating vector and the signal features extracted by the backbone network Element-wise multiplication yields enhanced features, thereby strengthening channels consistent with spatial geometry and suppressing unreliable channels caused by noise or signal loss.

[0062] Formula (10) represents the enhanced signal features obtained by weighting the backbone signal features channel by channel using spatial gating vectors. Specifically, for channels with larger gate values, their corresponding signal features are preserved or enhanced; for channels with smaller gate values, their corresponding signal features are suppressed, thereby reducing the impact of unreliable features caused by occlusion, communication interruption, or noise on the positioning results. The enhanced features after spatial prior constraints not only retain the effective positioning information in the original signal but also integrate the spatial geometric relationship of the LED array, thus improving the positioning accuracy and robustness of the model under incomplete signal conditions.

[0063] S326. Input the enhanced features after geometric gating into the multi-task output head to obtain the receiver position coordinates and LED distance estimate.

[0064] Specifically, the feature input coordinate output head after geometric gating enhancement is used to obtain the receiver position coordinates; the feature input distance output head after geometric gating enhancement is used to obtain the LED distance estimate.

[0065] S33. Based on the receiver's true coordinates, receiver's angle label, LED-receiver distance label calculated from the receiver's true coordinates, and the optional angle estimates, position coordinates, and LED distance estimates corresponding to the complete view and occluded view output by the GeoGated-PiTNet model, calculate the complete view loss and the occluded view regression loss respectively to train the GeoGated-PiTNet model.

[0066] Specifically, true coordinates are the actual measured spatial coordinates, which are obtained manually during dataset construction. The same applies to angle labels, which can be measured using inertial sensing devices such as IMUs. Angle means attitude, the attitude of the receiving device in space, expressed in angles. Distance is the distance between the transmitting LED and the receiving device.

[0067] The loss function consists of complete view loss and occluded view loss. In the early stages of training, the weights of occluded views are relatively small in the first stage, prioritizing the learning of stable spatial mappings from complete signals. As training progresses, the weights of occluded views are gradually increased, allowing the model to gradually adapt to incomplete observations. In the second stage, the backbone network is frozen, and only the high-level regression modules are fine-tuned to maintain the robust feature manifold obtained in the first stage, while improving the localization accuracy under complete signals.

[0068] The loss function formula is as follows: (11) (12) Where L is the loss function, For the regression loss of the complete view, Let λ(t) represent the regression loss for the occluded view, where t represents the current training epoch, and λ(t) represents the asymptotic weighting coefficient of the occluded view loss. As the training epoch t increases, λ(t) gradually increases from a small value to 1, allowing the model to prioritize learning stable features under the complete view in the early stages of training, and gradually enhance its robust learning ability for occluded views in the later stages of training. When t is greater than or equal to 80, λ(t) remains at 1.

[0069] Phase 1: Progressive pre-training with dual views, such as Figure 2 As shown: In the first stage, the model is jointly trained using complete and occluded views. A progressive loss weighting method guides the model to gradually learn the spatial mapping relationship under complete signals and robust feature representations under incomplete signals. Its optimization objective can be expressed as: (13) Where θ represents all trainable parameters of the network, and X is the number of training samples. and They represent the first The regression loss for the complete view and the occluded view corresponding to each sample. These are progressive weighting coefficients that increase gradually with the number of training epochs. In the early stages of training, the model primarily relies on the complete view to learn stable spatial mapping relationships; as training progresses, the occluded view loss gradually plays a greater role, thereby progressively improving the model's adaptability to scenarios with missing signals.

[0070] After each training round, the validation loss is calculated on the complete validation set (Formula 11), and the optimal model parameters corresponding to the global minimum validation loss are continuously tracked and saved. The regression loss is calculated using the following method: (14) (15) in, and y represents the receiver position coordinates predicted by the model in the full view and the occluded view, respectively, and y represents the actual receiver position coordinates. and d represents the distance estimates from the receiver to each LED predicted by the model in the full view and the occluded view, respectively, and d represents the actual distance label of the LED; and These represent the optional angle estimates predicted by the model under the full view and the occluded view, respectively. a represents the angle label at the receiver; α represents the angle loss weighting coefficient; and ||·||2² represents the L2 squared value, used to calculate the squared error between the predicted result and the true label.

[0071] Phase Two: Freezing Backbone Network Fine-tuning After the first phase of training, the model parameters with the best performance on the validation set are loaded, and the first half of the parameters of the shared feature extraction structure are frozen, using the channel adjustment sublayer in the feature sharing layer as the freeze boundary. Specifically, the network parameters from the input end to or including the bias parameter of the first channel adjustment sublayer in the feature sharing layer are frozen, while the remaining network parameters remain trainable. Subsequently, the trainable modules after the freeze boundary are fine-tuned using the complete training set input, including the second half of the feature sharing layer, the angle feature extraction branch, the time-frequency-spatial feature transformation branch, the geometric gating enhancement module, and the multi-task output head module. This further improves the model's localization accuracy under complete signal conditions while retaining the stable low-level feature representations learned in the first phase.

[0072] The optimization objective at this stage can be expressed as: (16) in, This indicates that the regression layer parameters for the task are being fine-tuned, while the backbone network parameters remain fixed. The fixed weights are used for the occlusion view loss in the second stage. Compared to the first stage, this stage no longer uses progressive weight adjustment. Instead, it refines the high-level regression module while maintaining robust feature representation, thereby further improving the accuracy of localization coordinates and distance prediction, while reducing multi-task gradient interference and improving model convergence stability.

[0073] In the second stage, only the coordinate and distance regression losses are retained, and the optimal model parameters are output after fine-tuning.

[0074] (17) (18) After each training round, the validation loss is calculated on the complete validation set, and the optimal model parameters corresponding to the global minimum validation loss are continuously tracked and saved. S4. Deploy the optimal model to locate the original samples of the complex baseband corresponding to each LED light source obtained through pilot extraction.

[0075] Specifically, in step S4, the original complex baseband samples corresponding to each LED light source obtained through pilot extraction are subjected to sample-level RMS normalization to obtain normalized complex baseband signals. The normalized complex baseband signals are then input into the deployed optimal model to output positioning results. The positioning results include optional angle estimation, position coordinates, and LED distance estimation.

[0076] Performance comparisons were conducted using three models: CRNN-MLP, STAK-Ploc, and GeoGated-PiTNet. CRNN-MLP is a traditional convolutional recurrent model, STAK-Ploc is an ablation model with geometric gating removed, and GeoGated-PiTNet is a complete model incorporating geometric prior gating.

[0077] Table 1 General Experimental Parameters

[0078] Test set A and test set B are two test sets with different sample values.

[0079] Table 2 Inference Time (s)

[0080] The normalization described in this embodiment refers to performing RMS normalization processing on the input data (original samples of complex baseband).

[0081] Table 2 compares the inference times of the three models. After introducing the temporal and multi-head attention modules, the inference time of STAK-Ploc increases compared to the traditional CRNN-MLP. Compared to the ablation model STAK-Ploc, the full model GeoGated-PiTNet further increases the inference time due to the addition of geometric prior modulation and gating graph attention geometric gating modules, which bring a small amount of computational overhead. Overall, the inference time remains at 3.37-3.47s.

[0082] The results show that the geometric gating module introduces a certain computational cost, but does not cause an order-of-magnitude increase in inference time. The overall inference latency still meets the requirements of indoor visible light positioning engineering applications. The model achieves improved positioning accuracy and robustness in occluded scenarios at an acceptable inference time cost.

[0083] Figure 3 The figures show a comparison of the accuracy of CRNN-MLP, STAK-Ploc, and GeoGated-PiTNet models on test sets A and B without RMS normalization. Experimental results show that GeoGated-PiTNet's CDF curves are furthest to the left on both test sets A and B, and its center error and interquartile range in the box plots are smaller, indicating that it is more stable under mixed signal-to-noise ratio conditions. STAK-Ploc's performance degrades after removing geometric gating, indicating that LED spatial prior gating makes a substantial contribution to reducing positioning errors. Although TransEnc can capture global dependencies, it lacks explicit geometric constraints, and there are still many samples in the medium-to-high error range.

[0084] Figures 4-9The diagram compares the accuracy distribution of three models—CRNN-MLP, STAK-Ploc, and GeoGated-PiTNet—before and after RMS normalization. It shows that the traditional CRNN-MLP model is highly dependent on the scale of the input signal; without RMS normalization, a large number of samples have localization errors exceeding 100cm, resulting in severe performance degradation. RMS normalization significantly improves performance. The STAK-Ploc ablation model has good feature processing capabilities, avoiding many extreme errors even without normalization; RMS normalization further improves the error distribution. The GeoGated-PiTNet model of this invention, without RMS normalization preprocessing, already has nearly half of the samples with localization errors in the 1-5cm range, with a small proportion of samples exhibiting extreme large errors, demonstrating stronger robustness to the magnitude of the input signal. After RMS normalization, the samples are further concentrated in the low-error range within 20cm, resulting in the best overall localization performance.

[0085] Figure 10 The images show a comparison of the CDF localization accuracy of the three models on test sets A and B after normalization. The GeoGated-PiTNet model from this invention exhibits the best performance.

[0086] like Figures 11-20 As shown, at occlusion ratios of 0%, 5%, 10%, 20%, and 30%, only the normalized model shows that the error migrates to the mid-to-high range at high occlusion ratios; normalization + progressive training is effective under mild occlusion, but has limited compensation for high occlusion; normalization + dual-view training can suppress high-error samples; normalization + dual-view + progressive training shows the most stable performance on both test sets.

[0087] Table 3 Comparison of localization errors (cm) of GeoGated-PiTNet model under different training strategies on test set A

[0088] Table 4 Comparison of localization errors (cm) of GeoGated-PiTNet model under different training strategies on test set B

[0089] Figure 21 This is a spatial distribution map of localization error in an occluded environment under the optimal training strategy (normalization + dual views + progressive). Depend on Figure 21It can be seen that under unobstructed conditions, the positioning error of most indoor spatial points is low, with only a small number of discrete high-error points. As the occlusion ratio increases from 0% to 20% and 30%, the positioning error in local areas of the space gradually increases, and the number of high-error samples increases, but there is no global positioning failure, and most indoor areas can still maintain a low positioning error.

[0090] Figure 22 This is a spatial distribution map of the localization error in an occluded environment using only the normalized baseline model. Under unoccluded conditions, the baseline model can achieve basic localization, but a certain number of discrete high-error points already exist. As the occlusion ratio increases to 20% and 30%, the number of high-error samples in the baseline model space increases significantly, and high-error points spread from the scene edge to the center of the room, with a significant degradation in the localization accuracy of a large number of sampling points. Comparing the results of the optimal training strategy of this invention in Figure 8, it can be seen that under the same occlusion conditions, the baseline model is more affected by signal loss, and its localization robustness is worse than the complete model with geometric prior gating, verifying the role of the geometric prior gating module in improving localization performance in occluded environments.

[0091] Example 2 This embodiment provides a system for implementing an indoor visible light positioning method in an occluded scene, including: Data acquisition module: used to receive LED pilot signals, extract complex baseband I / Q signals, and record or generate coordinate, distance and angle labels.

[0092] Sample normalization module: used to calculate the RMS energy scaling factor for each sample and to uniformly scale the I / Q components.

[0093] Graph Prior Modeling Module: Used to encode the 3D coordinates and topological relationships of LEDs into a graph structure, and generate spatial prior features through GAT.

[0094] Geometric Gated Feature Fusion Module: Used to generate gating vectors based on spatial priors to enhance or suppress the signal features of the backbone network.

[0095] Dual-view training module: used to construct complete view and continuously occluded view, and improve occlusion robustness through two-stage progressive optimization.

[0096] Positioning output module: used to output the receiver's three-dimensional coordinates, distance estimate, and optional angle estimate.

[0097] Example 3 This embodiment also provides an electronic device, including: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the indoor visible light positioning method for occluded scenes as described above.

[0098] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable computing device, consumer electronic device, etc.

[0099] Example 4 This embodiment also provides a machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the indoor visible light positioning method for occluded scenes as described above.

[0100] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0101] In this case, the program code read from the readable medium itself can perform the functions of any of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification.

[0102] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0103] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for indoor visible light positioning in an occluded scene, characterized in that, Includes the following steps: S1. Collect indoor visible light communication received signals and extract the original samples of complex baseband corresponding to each LED light source through pilot extraction; S2. Perform sample-level RMS normalization on the original complex baseband samples to obtain normalized complex baseband signals. Construct normalized complete views and normalized occluded views from the normalized complex baseband signals respectively. Construct a dual-view dataset with paired views. The dual-view dataset includes a training set and a validation set. S3. Construct the GeoGated-PiTNet model, which includes a feature embedding and sequence modeling module, a spatial feature extraction module, a geometric prior and gating module, and a multi-task output head module; The normalized complete view and the normalized occluded view are input into the GeoGated-PiTNet model respectively to obtain the optional angle estimate, position coordinates and LED distance estimate for each view; a dual-view progressive training strategy is adopted to train the GeoGated-PiTNet model through the training set, and the optimal parameters of the GeoGated-PiTNet model are obtained through the validation set to obtain the optimal model; S4. Deploy the optimal model to locate the original samples of the complex baseband corresponding to each LED light source obtained through pilot extraction.

2. The indoor visible light positioning method for occluded scenes according to claim 1, characterized in that, Step S2, which involves performing sample-level RMS normalization, includes: a. Traverse all sampling points of the original sample of the complex baseband and calculate the single-point amplitude energy of each sampling point, as shown in the following formula; (2) in, This represents the single-point amplitude energy of the sampling point corresponding to the p-th pilot and the m-th subcarrier. Represents the real part matrix The element in the p-th row and m-th column; Let represent the element in the p-th row and m-th column of the imaginary part matrix Q; b. Calculate the sample-level RMS factor based on the single-point amplitude energy, using the following formula: （3） Where P is the number of pilots and M is the number of subcarriers; c. Based on the sample-level RMS factor, the I and Q components are uniformly scaled to obtain the normalized complex baseband signal, where the normalized real and imaginary parts are represented as follows: (4) in, For numerically stable terms, The normalized real part, This is the normalized imaginary part.

3. The indoor visible light positioning method for occluded scenes according to claim 1, characterized in that, Step S3 is as follows: S31. The normalized complete view / occluded view is input into the feature embedding and sequence modeling module to obtain the corresponding sequence features; the sequence features are processed by the spatial feature extraction module to output the angle feature representation and the signal features extracted by the backbone network. ; Angle feature representation outputs selectable angle estimates via a multi-task output header module; S32, LED coordinates are used as fixed-space prior input to the graph attention network, and the updated LED node features are obtained through neighbor-weighted aggregation. These node features are then input into the geometric prior and gating module, along with the signal features. Element-wise modulation is used to generate enhanced features. These enhanced features are then input into the multi-task output head module, which outputs the receiver position coordinates and LED distance estimate. S33. Based on the receiver's true coordinates, receiver's angle label, LED-receiver distance label calculated from the receiver's true coordinates, and the optional angle estimates, position coordinates, and LED distance estimates corresponding to the complete view and occluded view output by the GeoGated-PiTNet model, calculate the complete view loss and the occluded view regression loss respectively to train the GeoGated-PiTNet model.

4. The indoor visible light positioning method for occluded scenes according to claim 3, characterized in that, The feature embedding and sequence modeling module in step S31 includes Patch Embedding, Bi-LSTM, and multi-head attention; the spatial feature extraction module includes a feature sharing layer, an angle feature extraction branch, and a time-frequency-spatial feature transformation branch. The multi-task output head module includes an angle output head, a coordinate output head, and a distance output head.

5. The indoor visible light positioning method for occluded scenes according to claim 4, characterized in that, Step S31 is as follows: The real and imaginary parts of the normalized complete view / occluded view are used as dual-channel complex baseband signals, which are divided and mapped into several embedding vectors by the PatchEmbedding operation to form a sequence representation. The sequence representation is then input into Bi-LSTM and a multi-head attention mechanism to extract the temporal dependence of the pilot signal in the time sampling dimension and the LED emission order dimension, thereby obtaining sequence features containing the correlation information between the pilot sampling order and the LED emission order. Subsequently, the sequence features are input through a feature sharing layer to the angle feature extraction layer of the angle feature extraction branch and the time-frequency-space feature conversion layer of the time-frequency-space feature conversion branch, respectively. Based on the angle feature extraction layer, angle feature representations related to the attitude change of the receiver are extracted from the sequence features, and the angle feature representations are input into the angle output head in the multi-task output head to output the selectable angle estimate of the receiver. The time-frequency-space feature conversion layer rearranges the sequence features according to the LED transmission order and time sampling order corresponding to the pilot signal, reconstructs it into a two-dimensional time-frequency-space feature map, and the CNN extracts spatially related features. The feature map output by the CNN or its pooled feature vector is used as the signal features extracted by the backbone network. .

6. The indoor visible light positioning method for occluded scenes according to claim 3, characterized in that, Step S32 is as follows: S321. The LED coordinates are used as the fixed-space prior input graph for the attention network. For node i in the LED coordinates and its neighbor node j, GAT calculates the attention score based on their encoded features, as shown in the following formula: （5） in, This represents the unnormalized attention score between node i and its neighbor node j; LeakyReLU is the activation function. This represents the learnable attention weight vector in a graph attention network. This represents the feature representation of node i after a linear transformation. This represents the feature representation of node j after a linear transformation. This represents the vector concatenation operation, which connects the features of node i and node j along their dimensions to form a joint feature; S322. Based on the attention score, calculate the attention weight of node i to its neighbor node j, using the following formula: (6) in, Let exp() represent the importance weights of node i to each of its neighboring nodes. Let i represent the set of neighboring nodes. This represents the unnormalized attention score between node i and its neighbor node k. S323. Based on the attention weights of node i to its neighbor node j, calculate the features of the i-th LED node after the graph attention network update, as follows: (7) in, The features of the i-th LED node after the graph attention network update are: Represents a nonlinear activation function; S324. All LED node features are subjected to global average pooling to obtain spatial prior features, and spatial gating vectors are generated by the sigmoid function: (8) in, This represents the global spatial prior features corresponding to the LED array, where N is the total number of LED nodes; (9) Where g represents the spatial gating vector generated by inputting global spatial prior features into the Sigmoid activation function; S325, Gating Vector and Backbone Network Extracted Signal Features Element-wise multiplication yields the enhanced features: (10) in, Indicates the enhanced signal characteristics, ⊙ represents the signal features extracted by the backbone network; ⊙ represents element-wise multiplication. S326. Input the geometrically gated enhanced features into the multi-task output head to obtain the receiver position coordinates and LED distance estimate.

7. The indoor visible light positioning method for occluded scenes according to claim 1, characterized in that, The dual-view progressive training strategy described in step S3 includes dual-view progressive pre-training and frozen backbone network fine-tuning. During the dual-view progressive pre-training phase, the optimization objective is represented as: （11） (12) Where θ represents all trainable parameters of the network, and X is the number of training samples. and These represent the regression losses for the full view and the occluded view, respectively. The progressive weighting coefficients increase gradually with the number of training rounds; After each training round, the validation loss is calculated on the full validation set, and the current optimal model parameters are saved. During the backbone network fine-tuning phase, the optimization objective can be expressed as: (13) in, This indicates that the regression layer parameters involved in the fine-tuning are maintained, while the backbone network parameters remain fixed. Fixed weights for the second-stage occlusion view loss.

8. A system for implementing the indoor visible light positioning method in an occluded scene as described in any one of claims 1-7, characterized in that, The system includes: Data acquisition module: used to receive LED pilot signals, extract complex baseband I / Q signals, and record or generate coordinate, distance, and angle labels; Sample normalization module: used to calculate the RMS energy scaling factor for each sample and to uniformly scale the I / Q components; Graph Prior Modeling Module: Used to encode the 3D coordinates and topological relationships of LEDs into a graph structure, and generate spatial prior features through GAT; Geometric Gated Feature Fusion Module: Used to generate gating vectors based on spatial priors to enhance or suppress the signal features of the backbone network; Dual-view training module: used to construct complete view and continuously occluded view, and improve occlusion robustness through two-stage progressive optimization; Positioning output module: used to output the receiver's three-dimensional coordinates, distance estimate, and optional angle estimate.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory that stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the indoor visible light positioning method for an occluded scene as described in any one of claims 1 to 7.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores executable instructions that, when executed, cause the machine to perform the indoor visible light positioning method for occluded scenes as described in any one of claims 1 to 7.

Images