A Dynamic Indoor Visible Light Positioning Method Based on Spatiotemporal Feature Information

Abstract

The present invention discloses a dynamic indoor visible light positioning method based on spatio-temporal feature information, which relates to the technical field of visible light communication. First, the present invention constructs a target motion model based on spatio-temporal feature information to analyze the spatio-temporal correlation in the received signal strength (RSS) sequence. The forward circular convolution module extracts the spatial texture feature information within the historical time period and combines the target motion model to predict the change trend of the spatial texture feature information within the future time period. The backward circular convolution module uses the future observation information to compensate the spatial texture feature information within the historical time period, so as to fully mine the time pattern, spatial features, and potential spatio-temporal correlation features from the spatial texture feature information of the RSS sequence in the historical time period, and achieve high-precision positioning of the moving target.

Description

Technical Field

[0001] The present invention relates to the technical field of visible light communication, and particularly relates to a dynamic indoor visible light positioning method, device, equipment and medium based on spatio-temporal feature information. Background Art

[0002] With the rapid development of wireless communication technology, the application of precise indoor positioning technology in multiple fields has gradually become a reality, such as emerging industries like intelligent warehouses, robot navigation, autonomous driving, and intelligent parking; visible light positioning (VLP), as a new type of indoor positioning technology, uses light-emitting diode (LED) signals for target positioning and attitude tracking, and has advantages such as strong stability, low-carbon environmental protection, and flexible coverage. Compared with traditional positioning technologies such as Wi-Fi, Bluetooth, and infrared, it provides higher accuracy and resolution.

[0003] However, current visible light dynamic positioning and attitude tracking technologies can achieve relatively high positioning accuracy in static environments. However, in dynamic environments, due to the influence of environmental factors (such as signal intensity fluctuations, occlusion effects, and light changes), the positioning accuracy and robustness of the system will significantly decrease; traditional methods based on signal propagation models (SPMs) rely on preset and stable signal propagation models for positioning calculations. However, in complex dynamic environments, the signal propagation characteristics will change, resulting in an increase in positioning errors. Especially in complex environmental conditions such as light, reflection, and multi-path propagation, the accuracy of the SPM model is difficult to guarantee; traditional fingerprint-based positioning methods estimate the position by establishing a fingerprint database and matching the received signal with the fingerprint database, without relying on an accurate SPM model. However, the FG method has high requirements for environmental staticity. For example, the parameters of the emission light source and receiver need to be stable; in reality, due to factors such as human activities and changes in indoor temperature and humidity, the signal propagation environment will change dynamically, resulting in a decrease in the effectiveness of the fingerprint database, which limits its application in dynamic scenarios; traditional photography-based positioning methods sense the LED array projection through a camera, but this method has been affected by photography errors and smear effects in dynamic scenarios.

[0004] In response to such problems, in recent years, some positioning methods based on machine learning have begun to receive attention; traditional deep learning methods, such as convolutional neural networks (CNNs), although have achieved certain results in some static or near-static environments, when facing the rapid changes in dynamic environments, due to the movement of the target, the signal propagation path and environment change, which will cause large fluctuations in signal intensity, resulting in difficulty in identifying the time features in the observed data at consecutive moments, and problems such as poor positioning accuracy and unstable real-time performance; therefore, in dynamic environments, the current methods are difficult to fully exploit the potential spatio-temporal correlation features in time series data, resulting in the positioning system being difficult to achieve high-precision positioning of moving targets. Summary of the Invention

[0005] An embodiment of the present invention provides a dynamic indoor visible light positioning method based on spatio-temporal feature information, which can solve the problem in the prior art that the existing methods are difficult to fully exploit the potential spatio-temporal correlation features in time series data, resulting in the positioning system being difficult to achieve high-precision positioning of moving targets.

[0006] An embodiment of the present invention provides a dynamic indoor visible light positioning method based on spatio-temporal feature information, including the following steps:

[0007] When a moving target moves indoors, extract the received signal strength RSS sequence of the visible light signal received by the moving target;

[0008] When the moving target makes a regular or random movement indoors, predict the position coordinates and pose parameters of the moving target at time t to construct a target motion model;

[0009] Construct a bidirectional recurrent convolutional neural network Bi-RCNN including a forward recurrent convolutional module and a backward recurrent convolutional module; use the forward recurrent convolutional module to extract the spatial texture features in the received signal strength RSS sequence of the historical period, and combine with the target motion model to predict the spatial texture features in the received signal strength RSS sequence of the future period; use the backward recurrent convolutional module to extract the spatial texture features in the predicted received signal strength RSS sequence of the future period to correct the spatial texture features in the received signal strength RSS sequence of the historical period;

[0010] According to the corrected spatial texture features in the received signal strength RSS sequence of the historical period, and combine with the target motion model to predict the pose parameters of the moving target at the current moment, and determine the position of the moving target at the current moment.

[0011] Preferably, the constructing of the target motion model includes:

[0012] Assume that the moving target makes a regular or random movement indoors, and the regular or random movement is respectively characterized as uniform motion or random motion;

[0013] Let the uniform motion speed be Assume Where Represents the absolute value of the speed; let Represents the random motion vector at time t, which follows a Gaussian Gauss random process, and is expressed as Where Represents The covariance matrix of, assume that the pose of the moving target Randomly changes; then the predicted position coordinates of the moving target at time t + 1 Expressed as:

[0014]

[0015] Where: w u and w s ∈ [0, 1] represent the weights of the uniform component and the random component respectively, and w u + w s = 1;

[0016] In the positioning of a moving target, the samples of the received signal strength RSS sequence depend on the pose parameters of the maneuvering target at time t, and the pose parameters are expressed as Using the pose parameters Predict the time-varying pose information of the target, and perform the positioning and tracking of the moving target;

[0017] In the actual process of moving target positioning, continuously obtain the received signal strength RSS of the moving target to obtain the sample sequence of the received signal strength RSS The corresponding positioning label is

[0018] Preferably, the method for correcting the spatial texture features in the received signal strength RSS sequence in the historical period includes:

[0019] Both the forward circular convolution module and the backward circular convolution module include forget gate branches; the forward circular convolution module Generate a weight vector by fusing the hidden state information at past times and the current observation information The backward circular convolution module Fuse the hidden state information in the future and the current observation information To form a weight

[0020] In the forward circular convolution module, the forget gate weights and inherits the historical memory And fuses with the observation information at the current time To obtain a new feature And stores it in the memory cell, expressed as:

[0021]

[0022] Where: ⊙ represents the Hadamard product, the output weight of the forget gate And the output of the memory gate

[0023] The memory cell state of the backward circular convolution module is expressed as:

[0024]

[0025] The forward cyclic convolution module uses a memory gate and a forget gate to extract the received signal strength (RSS) observation samples at the current moment and the historical hidden information and the stable spatial texture features in it. The backward cyclic convolution module uses a memory gate and a forget gate to extract the received signal strength (RSS) observation samples at the current moment and the future hidden information the stable spatial texture features in it.

[0026] Preferably, the pose parameters of the moving target at the current moment include:

[0027] Using the perception layer of the bidirectional cyclic convolutional neural network (Bi-RCNN) to fuse the hidden states obtained by the forward cyclic convolution module and the backward cyclic convolution module to determine the pose parameters of the moving target

[0028] Let:

[0029]

[0030] be the input of the perception layer that fuses the hidden states of the forward cyclic convolution module and the backward cyclic convolution module. Let:

[0031]

[0032] and represent the weights and biases of the perception layer respectively. Then the pose parameters of the moving target are estimated as:

[0033]

[0034] where ReLU(·) represents the ReLU activation function.

[0035] Preferably, after the received signal strength (RSS) sequence is obtained, the sliding window method is used to intercept and normalize the received signal strength (RSS) sequence.

[0036] An embodiment of the present invention further provides a dynamic indoor visible light positioning device based on spatio-temporal feature information, including:

[0037] An acquisition module, when the moving target moves indoors, extracts the received signal strength (RSS) sequence of the visible light signal received by the moving target;

[0038] Motion module, where the moving target makes regular or random movements indoors, predicts the position coordinates and pose parameters of the moving target at time t to construct a target motion model;

[0039] Recognition module, used to construct a bidirectional recurrent convolutional neural network Bi-RCNN including a forward recurrent convolutional module and a backward recurrent convolutional module; uses the forward recurrent convolutional module to extract the spatial texture features in the received signal strength RSS sequence in the historical period, and combines with the target motion model to predict the spatial texture features in the received signal strength RSS sequence in the future period; uses the backward recurrent convolutional module to extract the spatial texture features in the predicted received signal strength RSS sequence in the future period to correct the spatial texture features in the received signal strength RSS sequence in the historical period;

[0040] Positioning module, used to determine the position of the moving target at the current moment according to the corrected spatial texture features in the received signal strength RSS sequence in the historical period and combine with the target motion model to predict the pose parameters of the moving target at the current moment.

[0041] An embodiment of the present invention also provides an electronic device, including a memory and a processor;

[0042] The memory is used to store a computer program;

[0043] The processor, when executing the computer program stored in the memory, implements the steps of a dynamic indoor visible light positioning method based on spatio-temporal feature information as described above.

[0044] An embodiment of the present invention also provides a computer-readable storage medium, used to store a computer program, and when the computer program is executed by a processor, it implements the steps of a dynamic indoor visible light positioning method based on spatio-temporal feature information as described above.

[0045] An embodiment of the present invention provides a dynamic indoor visible light positioning method based on spatio-temporal feature information. Compared with the prior art, its beneficial effects are as follows:

[0046] In the present invention, a target motion model is constructed by a moving target performing regular or random motion indoors. The forward cyclic convolution module is used to extract the spatial texture features within the received signal strength (RSS) sequence in the historical period, and combined with the target motion model, to predict the spatial texture features within the RSS sequence in the future period. The backward cyclic convolution module is used to extract the spatial texture features within the predicted RSS sequence in the future period to correct the spatial texture features within the RSS sequence in the historical period. According to the corrected spatial texture features within the RSS sequence in the historical period, the position of the moving target is determined. This process first constructs a target motion model based on spatio-temporal feature information to analyze the spatio-temporal correlation in the RSS sequence, extracts the spatial texture feature information in the historical time period through the forward cyclic convolution module, and combines the target motion model to predict the change trend of the spatial texture feature information in the future time period. The backward cyclic convolution module then uses the future observation information to compensate the spatial texture feature information in the historical time period, so as to fully mine the time pattern, spatial features, and potential spatio-temporal correlation features from the spatial texture feature information of the RSS sequence in the historical time period, and achieve high-precision positioning of the moving target.

[0047] Moreover, in the present invention, by setting memory cells and forget gate mechanisms in the forward cyclic convolution module and the backward cyclic convolution module, the response ability to signal strength fluctuations in a dynamic environment is enhanced; at the same time, the three-dimensional convolution network is also combined to further analyze the time pattern and spatial features in the RSS sequence, effectively improving the positioning and attitude tracking performance of dynamic targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 FIG. is a schematic diagram of the overall solution of a dynamic indoor visible light positioning method based on spatio-temporal feature information provided by an embodiment of the present invention;

[0049] Figure 2 FIG. is a schematic diagram of a dynamic target visible light positioning system of a dynamic indoor visible light positioning method based on spatio-temporal feature information provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given with reference to the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0051] See Figure 1, embodiments of the present invention provide a dynamic indoor visible light positioning method based on spatio-temporal feature information; currently, existing visible light positioning (VLP) technologies can be mainly divided into methods based on signal propagation models (SPM), fingerprint-based positioning methods, neural network methods based on machine learning, and photography-based positioning methods.

[0052] Methods based on signal propagation models (SPM); this method uses a photodiode (PD) to receive visible light signals and inversely calculates the target position and attitude parameters based on the signal propagation model; although the positioning accuracy is theoretically high, it is highly dependent on environmental conditions in practical applications; when the indoor environment changes (such as small-scale fading, scattering, and reflection, etc.), the SPM model is difficult to accurately reflect the actual signal propagation characteristics, resulting in a significant decrease in positioning accuracy and limited adaptability in dynamic scenarios.

[0053] Fingerprint-based positioning methods; such methods estimate the position by establishing a fingerprint database and matching the received signal with the fingerprint library, without relying on an accurate SPM model; however, the FG method has high requirements for environmental staticness, for example, the parameters of the emission light source and the receiver need to be stable; in reality, due to factors such as human activities, changes in indoor temperature and humidity, etc., the signal propagation environment will change dynamically, resulting in a decrease in the effectiveness of the fingerprint library, limiting its application in dynamic scenarios.

[0054] Neural network-based positioning methods; in recent years, convolutional neural networks (CNNs) and deep neural networks (DNNs) based on deep learning have been applied to the VLP field; these methods improve the positioning accuracy through feature learning, but most methods ignore the temporal correlation of continuous time-series data, resulting in insufficient positioning performance in high-speed moving targets and complex dynamic environments, especially significant bottlenecks in real-time response capabilities and accuracy.

[0055] Photography-based positioning methods; the visible light positioning (Image-based VLP) method based on a camera senses the LED array projection through a camera, but this method is easily affected by photography errors (such as focal length errors, camera jitter, LED camera vignetting, etc.) and the trailing effect in dynamic scenarios, resulting in great limitations in the application of fast-moving targets and dynamic environments.

[0056] The present invention aims to solve the problem that the positioning accuracy of existing visible light positioning (VLP) systems significantly decreases in a dynamic environment, especially the problem of signal intensity fluctuations caused by the movement of the target, which leads to changes in the signal propagation path and the environment. Traditional positioning methods based on the signal propagation model (SPM) and fingerprint method (FG) have insufficient adaptability in a dynamic environment and are difficult to maintain high-precision positioning. In addition, the existing machine learning models have limited information mining and generalization capabilities in dynamic scenarios and are difficult to effectively utilize the spatio-temporal correlation of target movement. The present invention proposes a bidirectional recurrent convolutional network model based on spatio-temporal correlation, aiming to achieve high-precision estimation of the target position and attitude in a dynamic environment by mining the time patterns and spatial features in the received signal strength (RSS) sequence. The technology of the present invention mainly includes:

[0057] Step 1: Preprocessing of received signal strength.

[0058] The RSS sequence is intercepted and normalized by using the sliding window technique to reduce the influence of noise and improve the data quality, providing high-quality input data for subsequent model training and pose estimation.

[0059] Step 2: Spatio-temporal feature extraction.

[0060] A model based on spatio-temporal feature information is constructed to analyze the spatio-temporal correlation in the RSS sequence. The forward recurrent memory module extracts the spatial texture information in the historical time period and predicts the change trend in the future time period. The backward recurrent memory module compensates the historical texture information by using the future observation information. The convolutional network is combined to extract the spatial texture, enhancing the model's response ability to signal intensity fluctuations.

[0061] Step 3: Prediction of the pose of a dynamic target.

[0062] The extracted spatio-temporal feature information is used to accurately predict the pose trajectory of the target through the model, ensuring stable and reliable positioning and attitude tracking in a dynamic environment.

[0063] By effectively utilizing the spatio-temporal features in the RSS sequence, the present invention breaks through the dependence on static models of traditional methods, improves the adaptability of the visible light positioning system in a complex dynamic environment, and has significant practical value and broad application prospects.

[0064] Specifically:

[0065] I. Visible light dynamic positioning model.

[0066] 1. Target motion model.

[0067] As Figure 2As shown in the figure, assume that the maneuvering target moves randomly indoors and consider the general case; the movement of the maneuvering target consists of two components: (predictable) regular movement and (difficult-to-predict) random movement, which are characterized by uniform motion and random movement respectively; assume that the speed of uniform motion is Assume that where represents the absolute value of the speed; let represent the random movement vector at time t, which follows a Gaussian random process, that is where represents the covariance matrix of Assume that the target attitude changes randomly; according to the above assumptions, the position coordinates of the maneuvering target at time t+1

[0068]

[0069] where: w u and w s ∈[0,1] represent the weights of the uniform component and the random component respectively, and w u +w s =1; here, it is considered that in the case of no high-order statistical information, the Gaussian process is the maximum entropy model, with the strongest model generalization ability and the smallest risk of modeling mismatch.

[0070] 2. Continuous RSS sequence observation model.

[0071] In the static observation model of the RSS sequence, the observation sample depends on the relative position and angular relationship between the LED and the PD; in dynamic positioning, the RSS sample ultimately depends on the pose parameters of the maneuvering target at time t, denoted as As long as the RSS sample depends on the maneuvering target positioning parameter it can be used to estimate the time-varying pose information of the target, so as to perform positioning and tracking; in the actual positioning process, it is necessary to continuously sample the RSS observation data of the moving target to obtain the RSS sample sequence The corresponding positioning label is Since the propagation parameters in the dynamic environment change dynamically with time, the observation sample and its label also change randomly.

[0072] II. Specific design.

[0073] As Figure 1As shown in the figure, it is divided into two stages: offline training and online positioning. In the offline training stage, the network parameters are optimized to obtain the mapping relationship between the observation samples and the pose parameters of the maneuvering target. Subsequently, in the online positioning and tracking stage, the real-time visible light observation sample sequence is input into the trained network to output the pose parameter estimation of the moving target. Since the received signal strength (RSS) observation samples are affected by the environmental parameters and the random changes in the target motion posture, their fluctuations show random characteristics. Therefore, the static parameters change from time-invariant to time-varying parameters. The proposed network will extract the stable temporal and spatial texture structures in the observation data, thus achieving robust positioning and tracking results. The specific steps are as follows:

[0074] 1. Data preprocessing.

[0075] In the input data preprocessing stage, first, the received signal strength sequence and the position information sequence are intercepted through a long sliding window to form the k-th input sequence of the model. Next, these sequences are normalized to eliminate the scale difference between the data and ensure the effectiveness and stability of model training. Specifically, the received signal strength sequence and the position information sequence are normalized by calculating their mean and standard deviation respectively to generate the normalization center and scale parameters.

[0076] In this process, the normalization center is used to adjust the center position of the data, and the normalization scale is used to control the distribution range of the data. Finally, these normalization center and scale parameters are integrated into the joint parameters of the input data, providing the necessary input for the subsequent forward recurrent convolutional memory and backward recurrent convolutional memory modules, ensuring that the model can effectively extract spatio-temporal features when processing dynamic target positioning.

[0077] 2. Forgetting gate branch.

[0078] The design of the forgetting gate branch is for the forward and backward recurrent memory modules. The forward module generates a weight vector by fusing the past hidden state information and the current observation information while the backward module fuses the future hidden state information and the current observation information to form a weight These weight coefficients are calculated through a fully connected layer and processed by a sigmoid activation function to effectively retain or discard historical information when updating the memory cell information, thereby enhancing the model's ability to extract temporal and spatial features.

[0079] 3. Memory gate branch.

[0080] The memory gate enhances the texture feature extraction of current samples through a 3D convolutional neural network; the memory gate of the forward recurrent module consists of multiple 3D convolutional layers and fully connected layers. After the input data extracts spatial texture features through the 3D convolutional network, it is then input into the fully connected layer to extract spatio-temporal features; the input of the forward memory gate at the current moment is stacked by the past hidden state and observation information; the 3D convolutional output is rearranged into a vector form and combined with weights and biases to calculate the memory gate output; the structure of the backward recurrent module is similar to that of the forward module, and the output is calculated through the corresponding input and convolutional kernel parameters; this structure effectively improves the model's ability to fuse spatial and temporal features.

[0081] 4. Output gate branch.

[0082] The output gate branch is similar to the forget gate. After the hidden state information and RSS observation information are concatenated, they pass through a single fully connected layer to obtain the output of the output gate. The output gate state is divided into the output gate of the forward recurrent module and the output gate of the backward recurrent module, which are respectively defined as:

[0083]

[0084] 5. Update memory cells and hidden state cells.

[0085] The forward recurrent convolutional module uses memory cells to store the spatio-temporal texture features of the observed samples extracted by this module, and cyclically inputs them into the processing of the next moment; then, by fusing the feature of the observed data at different moments, the positioning and tracking performance is enhanced.

[0086] First, in the forward recurrent module, the forget gate weights and inherits the historical memory and then fuses it with the current observation information to obtain a new feature

[0087]

[0088] where: ⊙ represents the Hadamard product, the output weight of the forget gate and the output of the memory gate

[0089] Similarly, the memory cell state of the backward recurrent module is expressed as:

[0090]

[0091] The forward recurrent convolutional module uses the memory gate and the forget gate to extract the current moment RSS observed samples and the historical hidden information and stable spatial texture features, and the backward recurrent module extracts actual RSS observation samples using memory gates and forget gates and future hidden information stable spatial texture features.

[0092] 6. Output pose trajectory estimation.

[0093] Using the hidden states obtained by the perception layer fusing the forward and backward recurrent convolutional modules to determine the pose parameters of the maneuvering target Let:[[]]

[0094]

[0095] be the input of the perception layer for fusing the forward and backward hidden states. Let:[[]]

[0096] and

[0097] and represent the weights and biases of the perception layer respectively. Then the pose estimation of the moving target is:[[]]

[0098]

[0099] where: ReLU(·) represents the ReLU activation function. For the sake of concise expression, let[[]]

[0100] The present invention proposes a dynamic visible light positioning scheme based on spatio-temporal feature information to address the positioning accuracy problem caused by the signal propagation path and environmental changes due to the movement of the target in the existing visible light positioning system; the scheme of the present invention intercepts and normalizes the received signal strength (RSS) sequence through the sliding window technique, laying a foundation for subsequent model training and pose estimation; in the model, by analyzing the spatio-temporal correlation in the RSS sequence, the forward recurrent memory module extracts the spatial texture information at historical moments to predict and correct the spatial texture information at future moments; while the backward recurrent memory module extracts the spatial texture information at future moments to compensate for the spatial texture information at historical moments; this design enhances the discriminative features of the moving target; the bidirectional recurrent neural network significantly improves the positioning and tracking performance by fusing the spatio-temporal texture features at the front and back moments and enhances the response ability to dynamic signal fluctuations.

[0101] Aiming at the problem that in the existing visible light positioning (VLP) system, the signal propagation path and environment change due to the movement of the target, so that the static parameter model is difficult to maintain high-precision positioning in a dynamic environment, the present invention proposes a bidirectional recurrent convolutional network model based on spatiotemporal correlation, and the core steps of the model are as follows: first, the received signal strength (RSS) sequence is intercepted and normalized by using a sliding window technology to lay a foundation for subsequent model training and pose estimation; second, the spatiotemporal correlation in the RSS sequence is analyzed, the temporal pattern and spatial features in the data are identified, and the spatiotemporal texture in the RSS observation sequence is extracted by using a bidirectional convolution module, and the spatiotemporal texture feature information in the RSS sequence is extracted by introducing a memory cell structure, and the cached spatiotemporal information is reasonably inherited by using a forget gate mechanism to ensure the long-term memory of these features, thereby enhancing the model's response ability to dynamic signal fluctuations; finally, based on the extracted spatiotemporal correlation, the pose trajectory of the target is predicted to achieve stable and accurate estimation of the dynamic target.

[0102] The present invention assumes that the target terminal moves at a uniform speed or randomly in the room (obeying Gaussian random process), the spatial position obeys the random distribution function, and the target spatial direction information obeys the truncated Laplace distribution, and is based on the typical visible light positioning boundary conditions (5m×5m×3m, the emitting light source is 4×4 LEDs). 6 For spatial data points, when the moving speed is ≤15km / h, the CDF-90% accuracy is ≤10cm; the average yaw angle estimation error α error is ≤3°, and the time required for a positioning calculation is no more than 100ms.

[0103] The present invention proposes a method combining a bidirectional loop structure with a three-dimensional convolutional network to respectively extract the time features and spatial texture features in the received signal strength (RSS) data; by analyzing the spatiotemporal correlation in the RSS sequence, the forward loop memory module extracts the spatial texture information of the historical moments to predict and correct the spatial texture information of the future moments; and the backward loop memory module extracts the spatial texture information of the future moments to compensate for the spatial texture information of the historical moments; in a complex environment, the present invention enhances the response capability to dynamic signal fluctuations and realizes stable and accurate posture trajectory prediction of dynamic targets; compared with the traditional dynamic VLP positioning method based on the signal propagation model (SPM), under the same environmental conditions and data points, the CDF-90% accuracy is improved by 50%.

[0104] Creative Aspects:

[0105] The present invention innovatively introduces a bidirectional recurrent convolutional neural network based on spatio-temporal correlation, combines a forward recurrent memory module and a backward recurrent memory module to extract the spatial texture information at historical and future moments respectively, and enhances the response ability to signal intensity fluctuations in a dynamic environment through a memory cell and a forgetting gate mechanism; at the same time, combines a three-dimensional convolutional network to further analyze the temporal patterns and spatial features in the received signal strength (RSS) sequence, effectively improving the positioning and attitude tracking performance of dynamic targets; the method of the present invention is significantly superior to traditional positioning methods based on static parameter models or signal propagation models (SPM), solves the problems of insufficient positioning accuracy and robustness in a dynamic environment, and realizes the accurate prediction of the pose trajectory in a dynamic environment.

[0106] Novelty aspect:

[0107] The present invention proposes a brand-new dynamic visible light positioning method, incorporates the spatio-temporal correlation of the RSS sequence into the model design, and breaks through the bottleneck of the insufficient adaptability of traditional static fingerprint methods and ranging models to dynamic environments; through the combination of a bidirectional recurrent structure and a three-dimensional convolutional network, the method of the present invention can deeply excavate the spatio-temporal texture information in the RSS sequence, and store and update the features in a dynamic environment through a memory unit, effectively overcoming the problem of insufficient generalization ability of existing machine learning models in dynamic scenarios; experimental results show that in a typical dynamic scenario, the CDF-90% positioning accuracy of the method of the present invention is improved by 50% compared with traditional methods, and the average yaw angle estimation error is controlled within 3°, having strong technological leadership and scientific innovation.

[0108] Practicality aspect:

[0109] The present invention has a wide range of application prospects, especially suitable for scenarios requiring high-precision dynamic positioning and attitude tracking, such as intelligent warehousing, industrial automation, unmanned driving navigation and other fields; the present invention adopts a dynamic positioning technology based on spatio-temporal feature information, can quickly and stably realize the positioning and attitude estimation of a target in a complex dynamic environment; by improving the adaptability of the system to dynamic signal fluctuations, it meets the actual requirements of high dynamics and high precision; in addition, the present invention has the advantage of low computational cost, the positioning calculation time does not exceed 100 ms, and can stably achieve an average positioning error ≤ 10 cm under the condition of a moving speed ≤ 15 km / h, providing a practical solution for the positioning requirements in dynamic scenarios.

[0110] The above-described embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A dynamic indoor visible light positioning method based on spatio-temporal feature information, characterized in that, Including the following steps: When the moving target is moving indoors, extract the received signal strength RSS sequence of the visible light signal received by the moving target; When the moving target makes regular or random movements indoors, predict the position coordinates and pose parameters of the moving target at time t to construct a target motion model; Construct a bidirectional recurrent convolutional neural network Bi-RCNN including a forward recurrent convolutional module and a backward recurrent convolutional module; Use the forward recurrent convolutional module to extract the spatial texture features in the received signal strength RSS sequence of the historical period, and combine with the target motion model to predict the spatial texture features in the received signal strength RSS sequence of the future period; Use the backward recurrent convolutional module to extract the spatial texture features in the predicted received signal strength RSS sequence of the future period to correct the spatial texture features in the received signal strength RSS sequence of the historical period; According to the corrected spatial texture features in the received signal strength RSS sequence of the historical period, and combine with the target motion model to predict the pose parameters of the moving target at the current moment, and determine the position of the moving target at the current moment.

2. The dynamic indoor visible light positioning method based on spatio-temporal feature information according to claim 1, characterized in that, The construction of the target motion model includes: Assume that the moving target makes regular or random movements indoors, and the regular or random movements are respectively characterized as uniform motion or random motion; Let the uniform motion speed be Suppose where represents the absolute value of the speed; let represent the random motion vector at time t, which follows a Gaussian random process and is expressed as where represents the covariance matrix of Suppose the attitude of the moving target changes randomly; then the predicted position coordinates of the moving target at time t + 1 where: w u and w s ∈ [0, 1] represent the weights of the uniform component and the random component respectively, and w u + w s = 1; In the positioning of a moving target, the samples of the received signal strength (RSS) sequence depend on the pose parameters of the maneuvering target at time t, and the pose parameters are expressed as Using the pose parameters Predict the time-varying pose information of the target, and perform the positioning and tracking of the moving target; During the actual moving target positioning process, continuously obtain the received signal strength (RSS) of the moving target to obtain a sample sequence of the received signal strength RSS The corresponding positioning label is 3. A dynamic indoor visible light positioning method based on spatio-temporal feature information according to claim 2, characterized in that, The correction of the spatial texture features in the received signal strength RSS sequence of the historical period includes: Both the forward cyclic convolution module and the backward cyclic convolution module include forget gate branches; the forward cyclic convolution module generates a weight vector by fusing the hidden state information at past time steps and the current observation information The backward cyclic convolution module fuses the hidden state information in the future and the current observation information to form a weight In the forward cyclic convolution module, the forget gate weights and inherits the historical memory and fuses it with the observation information at the current time to obtain a new feature which is stored in the memory cell and can be expressed as: where: ⊙ represents the Hadamard product, the output weight of the forget gate and the output of the memory gate The memory cell state of the backward recurrent convolutional module is expressed as: The forward cyclic convolution module uses memory gates and forget gates to extract the current received signal strength (RSS) observation samples and the historical hidden information and the stable spatial texture features in it. The backward cyclic convolution module uses memory gates and forget gates to extract the current received signal strength (RSS) observation samples and the future hidden information the stable spatial texture features in it.

4. A dynamic indoor visible light positioning method based on spatio-temporal feature information according to claim 3, characterized in that, The prediction of the pose parameters of the moving target at the current moment includes: The hidden state obtained by fusing the forward recurrent convolutional module and the backward recurrent convolutional module in the perception layer of the bidirectional recurrent convolutional neural network Bi-RCNN to determine the pose parameters of the moving target Let: For the input of the perception layer that fuses the hidden states of the forward recurrent convolutional module and the backward recurrent convolutional module, let: and represent the weights and biases of the perception layer respectively, then the pose parameters of the moving target are estimated as: Where: ReLU(·) represents the ReLU activation function.

5. A dynamic indoor visible light positioning method based on spatio-temporal feature information according to claim 1, characterized in that, After the received signal strength RSS sequence is obtained, use the sliding window method to intercept and normalize the received signal strength RSS sequence.

6. A dynamic indoor visible light positioning device based on spatio-temporal feature information, characterized in that, Including: A collection module that extracts the received signal strength RSS sequence of the visible light signal received by the moving target when the moving target is moving indoors; A motion module that predicts the position coordinates and pose parameters of the moving target at time t when the moving target makes regular or random movements indoors to construct a target motion model; An identification module for constructing a bidirectional recurrent convolutional neural network Bi-RCNN including a forward recurrent convolutional module and a backward recurrent convolutional module; Use the forward recurrent convolutional module to extract the spatial texture features in the received signal strength RSS sequence of the historical period, and combine with the target motion model to predict the spatial texture features in the received signal strength RSS sequence of the future period; Use the backward recurrent convolutional module to extract the spatial texture features in the predicted received signal strength RSS sequence of the future period to correct the spatial texture features in the received signal strength RSS sequence of the historical period; A positioning module for predicting the pose parameters of the moving target at the current moment according to the corrected spatial texture features in the received signal strength RSS sequence of the historical period, and combining with the target motion model to determine the position of the moving target at the current moment.

7. An electronic device, characterized in that, Including: A memory and a processor; The memory is used to store computer programs; When the processor is used to execute the computer program stored in the memory, the steps of a dynamic indoor visible light positioning method based on spatio-temporal feature information as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium, characterized in that, For storing a computer program, when the computer program is executed by a processor, the steps of a dynamic indoor visible light positioning method based on spatio-temporal feature information as described in any one of claims 1 to 5 are implemented.

Images