Passive domain high-confidence large-space positioning method and system

Abstract

The invention discloses a passive domain high-confidence large space positioning method and system, and the method comprises the steps: receiving a position signal of a to-be-positioned object in a VR environment through a receiver, and carrying out the normalization processing; constructing a position positioning model based on the knowledge distillation network; and inputting the processed position signal into a position positioning model to complete large-space positioning of the passive domain. According to the method, source domain data does not need to be accessed, data privacy is protected, and transmission cost is reduced. Meanwhile, through a k-NN correction strategy, the confidence coefficient of the pseudo tag is improved, and the sensitivity of the model to noise data is reduced. In addition, a teacher network is updated through EMA, and the stability of the model on a target domain is kept; and the model can quickly adapt to different target environments and has good expandability. According to the method, the robustness of the model in a noise environment is improved through screening and correction of the high-confidence pseudo labels.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer vision and computer graphics, and in particular to a large-space positioning method and system with high confidence in a passive domain. Background Art

[0002] In large-scale immersive tour experiences, precise location positioning is not only a key step in technological implementation, but also a crucial factor in enhancing the user experience. Through precise positioning, the system can identify the visitor's current location and provide customized virtual content and services accordingly. For example, in an immersive AR / VR museum tour experience, when a visitor approaches a specific cultural relic, the system can automatically load relevant background information, historical stories, or 3D reconstruction models based on their location information, allowing visitors to gain a more in-depth and vivid cultural experience. In addition, the system can recommend tour highlights based on the visitor's location and optimize the tour route to ensure that every visitor can explore the essence of the museum in the most efficient way.

[0003] Although deep learning-based large-scale spatial localization models have made significant progress, they still face numerous challenges in practical applications. First, existing unsupervised domain adaptation methods typically rely on large amounts of labeled data from the initial (source) domain. However, in many scenarios, this reliance on source data is a major obstacle to their widespread application. This is particularly true in privacy-sensitive applications, such as healthcare or personal identity protection. Accessing source domain data may violate privacy policies or laws and regulations, preventing data sharing or use. Furthermore, storage and computing resource limitations, as well as data transmission limitations, make direct use of source domain data impractical. Another challenge is the issue of target domain data. In many cases, the target domain data is small and noisy, making direct transfer of source domain models to the target domain ineffective. Due to the distributional differences between the target and source domains, simply applying a model from one domain to the other often results in degraded performance. Furthermore, when generating pseudo-labels in the target domain, if the confidence level of the pseudo-labels is low, this can lead to model overfitting, thereby reducing localization accuracy. This issue is particularly prominent in small-sample learning or high-noise environments, where a lack of sufficient high-quality training data directly impacts the model's learning performance and generalization ability. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a large-space positioning method and system that does not rely on source domain data. This method abandons the traditional practice of relying on source domain labeled data and relies only on unlabeled data in the target domain for self-supervised learning, which greatly reduces the dependence on source data and privacy risks. In order to improve the confidence of pseudo-labels, this method introduces a teacher network to generate pseudo-labels, and combines the k-nearest neighbor (k-NN) correction strategy to further improve the quality of pseudo-labels, reducing the model's dependence on low-confidence pseudo-labels. At the same time, the exponential moving average (EMA) mechanism is used to maintain the stability of the teacher network, allowing the student network to be dynamically adjusted on the target domain data, thereby enhancing the adaptability and robustness of the model.

[0005] To achieve the above object, the present invention provides a large-space positioning method with high confidence in the passive domain, comprising the following steps:

[0006] Use the receiver to receive the position signal of the object to be located in the VR environment and perform normalization processing;

[0007] Build a location positioning model based on the knowledge distillation network;

[0008] The processed position signal is input into the position positioning model to complete large-space positioning in the passive domain.

[0009] Preferably, the position positioning model is constructed based on a knowledge distillation network; wherein the knowledge distillation network realizes self-supervised learning through a teacher-student network structure; the teacher network generates pseudo labels to guide the training of the student network; the student network uses these pseudo labels for training to adapt to the target domain data.

[0010] Preferably, the position positioning model uses a Bayesian optimization framework to optimize the parameters θ of the position positioning model.

[0011]

[0012] Among them, θ represents the optimal parameter; It is a standard notation in optimization problems; Denotes the joint distribution P X，Y，Z expected value; Indicates the difference between the actual position (X, Y) and the predicted position in the X and Y directions The sum of the absolute errors of Where Var(·) represents the variance, represents the variance of the positioning error in the X direction, represents the variance of the positioning error in the Y direction; ψ(t) represents the time decay weight; T represents the total number of time steps; represents the sum of the absolute errors between the true position and the predicted position at time step t; α1, α2 and α3 are weight coefficients.

[0013] Preferably, the method for training the position location model includes: using multi-objective optimization to measure the accuracy, stability and dynamic adaptability of the model by combining absolute error, variance and time decay weight; at the same time, introducing time decay weight ψ(t) to adapt to dynamic environment:

[0014]

[0015] in, Represents the original input data of the target domain; Represents the target domain data after noise enhancement; ∈ i represents the i-th noise type; represents the indicator function, indicating whether the i-th noise type is selected, 1 represents selected, and 0 represents unselected; N represents the total number of noise types; p i represents the probability that the i-th noise type is selected.

[0016] Preferably, the position localization model adds feature consistency loss, noise robustness loss and prediction confidence loss on the basis of knowledge distillation loss:

[0017]

[0018] in, represents the knowledge distillation loss; represents feature consistency loss; ||·|| 2 represents the Euclidean norm; represents the noise robust loss; Represents the prediction confidence loss; v1, v2, and v3 represent weight coefficients, which are used to control the weights of different items in the total loss.

[0019] Preferably, the knowledge distillation loss The expression is as follows:

[0020]

[0021] in, is the student network loss function; θ s are the parameters of the student network; Represents the expected value of all possible target domain samples; where X′ t and Y′ t is the pseudo label generated by the teacher network, indicating the true location of the target domain sample, and is the predicted position of the student network; B represents the batch size; in and is the pseudo label of the b-th sample generated by the teacher network, and is the predicted position of the bth sample of the target domain data by the student network.

[0022] The present invention also provides a large-space positioning system with high confidence in a passive domain, and the system is used to implement the above method, comprising: a receiving module, a building module and a positioning module;

[0023] The receiving module is used to use a receiver to receive the position signal of the object to be located in the VR environment and perform normalization processing;

[0024] The building module is used to build a location positioning model based on the knowledge distillation network;

[0025] The positioning module is used to input the processed position signal into the position positioning model to complete large-space positioning in the passive domain.

[0026] Preferably, the workflow of the construction module includes: constructing the position positioning model based on the knowledge distillation network; wherein the knowledge distillation network realizes self-supervised learning through a teacher-student network structure; the teacher network generates pseudo labels to guide the training of the student network; the student network uses these pseudo labels for training to adapt to the target domain data.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] This method does not require access to source domain data, protecting data privacy and reducing transmission costs. Furthermore, through a k-NN correction strategy, the confidence of pseudo-labels is improved, reducing the model's sensitivity to noisy data. Furthermore, by updating the teacher network through EMA, the model maintains stability in the target domain. The model can also quickly adapt to different target environments, demonstrating good scalability. By screening and correcting high-confidence pseudo-labels, the method improves the model's robustness in noisy environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Example 1

[0034] This embodiment provides a large-space positioning method with high confidence in a passive domain, the steps including:

[0035] S1. Use a receiver to receive the position signal of the object to be located in the VR environment and perform normalization processing.

[0036] This embodiment supports "multimodal input data", and the received signal can be one or more of Wi-Fi RSSI, Bluetooth signal strength, ultra-wideband data, magnetometer data (auxiliary function), accelerometer data (auxiliary function), and other data, such as radio frequency identification (RFID) and channel state information (CSI).

[0037] After receiving the signal, it is normalized, and the steps include:

[0038] This embodiment uses Wi-Fi RSSI as an example and adopts minimum-maximum normalization:

[0039]

[0040] Among them, min(RSSI) and max(RSSI) are determined based on the training set statistics and mapped to the [0,1] interval.

[0041] Different normalization processing methods are selected for other types of received signals according to their types. For example, symmetric normalization is used for accelerometer data; while for magnetometer data, each of its three-axis data is normalized separately.

[0042] S2. Build a location positioning model based on the knowledge distillation network.

[0043] Self-supervised learning is achieved through a teacher-student network structure, in which the teacher network generates pseudo-labels to guide the training of the student network; the student network uses these pseudo-labels for training and adapts to the target domain data. To improve the robustness of the model, this embodiment introduces a k-nearest neighbor correction strategy to correct uncertain pseudo-labels. It should be noted that the "domain data" in the position positioning model only refers to the target domain data (i.e., unlabeled test data) and does not include source domain data (i.e., labeled training data); this embodiment can obtain high-precision positioning results based solely on the target domain data.

[0044] In order to optimize the parameters θ of the location positioning model, this embodiment uses the Bayesian optimization framework to train the model by maximizing the log-likelihood of the true location. Specifically, the goal of the model is to minimize the predicted location The error between the actual position (X, Y):

[0045]

[0046] Among them, θ represents the optimal parameter, that is, the parameter found by the optimization process to minimize the objective function; It is a standard notation in optimization problems, used to find the parameter θ that minimizes the objective function; is the joint distribution P X，Y，Z The expected value of all possible signals and the expected value of the corresponding true value (X, Y); Indicates the difference between the actual position (X, Y) and the predicted position in the X and Y directions The sum of the absolute errors is the total error under the Laplace error model, which is used to measure the overall positioning accuracy of the model; Where Var(·) represents the variance, represents the variance of the positioning error in the X direction, represents the variance of the positioning error in the Y direction; ψ(t) represents the time decay weight, which decreases with time t. Its function is to give higher weight to the most recent data, so that the model pays more attention to the latest positioning data; T is the total number of time steps, which represents the time window size used for optimization; Represents the sum of the absolute errors between the true position and the predicted position at time step t; α1, α2, and α3 are weight coefficients that control the absolute error term, variance term, and time decay term, respectively, thereby balancing the importance of different terms.

[0047] This example uses multi-objective optimization, combining absolute error, variance, and time-decay weighting to comprehensively measure the model's accuracy, stability, and dynamic adaptability. Furthermore, the introduction of a time-decay weight, ψ(t), allows the model to focus more on recent data, enabling better adaptation to dynamic environments. The inclusion of an error variance term reduces the model's sensitivity to noisy data, improving overall robustness.

[0048] Methods for training the above model include:

[0049] (1) Initialization: Both the teacher network and the student network are initialized as localization models pre-trained on the source domain data.

[0050] (2) Pseudo-label generation: The teacher network generates pseudo-labels for the target domain data Z t Generate pseudo labels (X′ t , Y′ t ).

[0051] (3) Data enhancement: The student network receives the target domain data after noise enhancement:

[0052]

[0053] in, Represents the original input data of the target domain; Represents the target domain data after noise enhancement; ∈ i represents the i-th noise type, ∈ i It can be Gaussian noise, salt and pepper noise, or impulse noise; represents the indicator function, indicating whether the i-th noise type is selected, 1 represents selected, and 0 represents unselected; N represents the total number of noise types; p i represents the probability that the i-th noise type is selected.

[0054] The purpose of data augmentation is to simulate multiple noise types, enabling the model to adapt to more complex noise environments. Furthermore, combining multiple noise types increases the diversity of the training data and more realistically simulates noise conditions in real-world scenarios. Furthermore, data augmentation prevents the model from overfitting to a single noise type and improves its performance in unseen noise conditions.

[0055] (4) Knowledge distillation: The student network is trained by minimizing the difference between the pseudo labels generated by the teacher network and the loss function is

[0056] (5) Teacher network update: The parameters of the teacher network are updated through EMA (Exponential Moving Average):

[0057] θ′t ←ρ·θ s +(1-ρ)·θ t

[0058] Among them, θ′ t represents the updated parameters of the teacher network; θ t represents the parameters of the teacher network before updating; ρ represents the decay rate of EMA, which controls the speed of updating the teacher network parameters; θ s Represents the parameters of the student network.

[0059] In order to further improve the robustness and adaptability of the student network in dynamic and noisy environments, feature consistency loss, noise robustness loss and prediction confidence loss are added on the basis of knowledge distillation loss:

[0060]

[0061] in, represents the knowledge distillation loss, which measures the consistency between the student network prediction and the teacher network pseudo-label; Represents feature consistency loss, which measures the feature consistency between the student network and the teacher network in the target domain. in Pair Distribution The expectation of all possible target domain input signals The expected value of is the feature extracted by the student network, represents the features extracted by the teacher network; ||·|| 2 represents the Euclidean norm, which is used to measure the consistency of features; represents the noise robust loss, which measures the sensitivity of the student network to noisy data, Where ∈ represents the noise added to the target domain data, Var represents the variance, which is used to measure the sensitivity of the feature extractor to noise; represents the prediction confidence loss, which is used to measure the importance the student network attaches to high-confidence predictions. Among them, Conf(·) represents the prediction confidence, which is used to measure the confidence of the discriminator in the target domain features; MAE(·) represents the mean absolute error, which is used to measure the prediction accuracy of the model; v1, v2 and v3 represent weight coefficients, which are used to control the weights of different items in the total loss.

[0062] The above knowledge distillation loss The expression is as follows:

[0063]

[0064] in, is the student network loss function, which measures the error between the student network predicted position and the pseudo label generated by the teacher network; θ s are the parameters of the student network, through To optimize θ s ; is the expected value, which represents the expected value of all possible target domain samples; where X′ t and Y′ t is the pseudo label generated by the teacher network, indicating the true location of the target domain sample, and is the predicted position of the student network; B represents the batch size, which is used for batch gradient descent to improve computational efficiency; in and is the pseudo label of the b-th sample generated by the teacher network, and is the predicted position of the bth sample of the target domain data by the student network.

[0065] To improve the confidence of pseudo labels, this embodiment introduces a k-NN correction strategy, which includes the following steps:

[0066] (1) Uncertain pseudo-label recognition

[0067] For each target domain sample, noise is added repeatedly multiple times to generate pseudo labels, and the standard deviation is calculated; if the standard deviation exceeds a preset threshold, the pseudo label is considered uncertain.

[0068] (2) K-NN correction: For uncertain pseudo-labels, the weighted average correction is performed using the confidence pseudo-labels of the k nearest neighbors.

[0069]

[0070] Among them, (X′ t , Y′ t ) is the corrected pseudo-label, which indicates the target position used by the student network for training; K is the number of nearest neighbors, used to calculate the weighted average; ω i It is the comprehensive weight of the i-th nearest neighbor that combines distance, confidence and local density, and its expression is:

[0071]

[0072] Among them, d i is the distance between the uncertain sample and the i-th nearest neighbor, which is used to measure the similarity between the uncertain sample and the nearest neighbor sample. is the confidence of the i-th nearest neighbor, which is used to measure the reliability of the pseudo label and can be calculated by the standard deviation of multiple predictions; is the high-confidence pseudo label of the verified i-th nearest neighbor generated by the teacher network; Density(c i ) is the local density of the i-th nearest neighbor, which is used to measure the density of samples in the feature space and can be calculated by clustering algorithm; It is a normalization factor, which ensures that the sum of the weighted average is 1, prevents the weight from being too large or too small, and ensures that the corrected pseudo-label is reasonable.

[0073] It should be noted that the specific network structures of the teacher model and the student model in the "knowledge distillation network" are not limited. The output of this embodiment is the corrected pseudo-label output by the teacher model and the "predicted position" output by the student model.

[0074] S3. Input the processed position signal into the position positioning model to complete large-space positioning in the passive domain.

[0075] The processed position signal is input into the position positioning model, and the signal recorded by the receiver is analyzed (in this embodiment, there are N receivers and one transmitter, and the signal emitted by the transmitter interacts with the objects in the environment and is captured by the receiver). Through a mapping (i.e., the position positioning model of this embodiment), the signal recorded by the receiver is converted into The real position of the object is mapped to the real position of the object, thereby inferring the real position of the signal source (such as the RFID tag). The method flow of this embodiment is as follows Figure 1 shown.

[0076] Example 2

[0077] This embodiment also provides a large-space positioning system with high confidence in the passive domain, including: a receiving module, a building module and a positioning module; the receiving module is used to use a receiver to receive the position signal of the object to be positioned in the VR environment and perform normalization processing; the building module is used to construct a position positioning model based on the knowledge distillation network; the positioning module is used to input the processed position signal into the position positioning model to complete large-space positioning in the passive domain.

[0078] The following will describe in detail how the present invention solves technical problems in real life in conjunction with the present embodiment.

[0079] First, the receiving module is used to receive the position signal of the object to be located in the VR environment and perform normalization processing.

[0080] This embodiment supports "multimodal input data", and the received signal can be one or more of Wi-Fi RSSI, Bluetooth signal strength, ultra-wideband data, magnetometer data (auxiliary function), accelerometer data (auxiliary function), and other data, such as radio frequency identification (RFID) and channel state information (CSI).

[0081] After receiving the signal, it is normalized. The process includes:

[0082] This embodiment uses Wi-Fi RSSI as an example and adopts minimum-maximum normalization:

[0083]

[0084] Among them, min(RSSI) and max(RSSI) are determined based on the training set statistics and mapped to the [0,1] interval.

[0085] Different normalization processing methods are selected for other types of received signals according to their types. For example, symmetric normalization is used for accelerometer data; while for magnetometer data, each of its three-axis data is normalized separately.

[0086] The construction module then builds a location positioning model based on the knowledge distillation network.

[0087] Self-supervised learning is achieved through a teacher-student network structure, in which the teacher network generates pseudo-labels to guide the training of the student network; the student network uses these pseudo-labels for training and adapts to the target domain data. To improve the robustness of the model, this embodiment introduces a k-nearest neighbor correction strategy to correct uncertain pseudo-labels. It should be noted that the "domain data" in the position positioning model only refers to the target domain data (i.e., unlabeled test data) and does not include source domain data (i.e., labeled training data); this embodiment can obtain high-precision positioning results based solely on the target domain data.

[0088] In order to optimize the parameters θ of the location positioning model, this embodiment uses the Bayesian optimization framework to train the model by maximizing the log-likelihood of the true location. Specifically, the goal of the model is to minimize the predicted location The error between the actual position (X, Y):

[0089]

[0090] Among them, θ represents the optimal parameter, that is, the parameter found by the optimization process to minimize the objective function; It is a standard notation in optimization problems, used to find the parameter θ that minimizes the objective function; is the joint distribution P X,Y,Z The expected value of all possible signals and the expected value of the corresponding true value (X, Y); Indicates the difference between the actual position (X, Y) and the predicted position in the X and Y directions The sum of the absolute errors is the total error under the Laplace error model, which is used to measure the overall positioning accuracy of the model; Where Var(·) represents the variance, represents the variance of the positioning error in the X direction, represents the variance of the positioning error in the Y direction; ψ(t) represents the time decay weight, which decreases with time t. Its function is to give higher weight to the most recent data, so that the model pays more attention to the latest positioning data; T is the total number of time steps, which represents the time window size used for optimization; Represents the sum of the absolute errors between the true position and the predicted position at time step t; α1, α2, and α3 are weight coefficients that control the absolute error term, variance term, and time decay term, respectively, thereby balancing the importance of different terms.

[0091] This example uses multi-objective optimization, combining absolute error, variance, and time-decay weighting to comprehensively measure the model's accuracy, stability, and dynamic adaptability. Furthermore, the introduction of a time-decay weight, ψ(t), allows the model to focus more on recent data, enabling better adaptation to dynamic environments. The inclusion of an error variance term reduces the model's sensitivity to noisy data, improving overall robustness.

[0092] Methods for training the above model include:

[0093] (1) Initialization: Both the teacher network and the student network are initialized as localization models pre-trained on the source domain data.

[0094] (2) Pseudo-label generation: The teacher network generates pseudo-labels for the target domain data Z t Generate pseudo labels (X′ t , Y′ t ).

[0095] (3) Data enhancement: The student network receives the target domain data after noise enhancement.

[0096]

[0097] in, Represents the original input data of the target domain; Represents the target domain data after noise enhancement; ∈ i represents the i-th noise type, ∈ i It can be Gaussian noise, salt and pepper noise, or impulse noise; represents the indicator function, indicating whether the i-th noise type is selected, 1 represents selected, and 0 represents unselected; N represents the total number of noise types; p i represents the probability that the i-th noise type is selected.

[0098] The purpose of data augmentation is to simulate multiple noise types, enabling the model to adapt to more complex noise environments. Furthermore, combining multiple noise types increases the diversity of the training data and more realistically simulates noise conditions in real-world scenarios. Furthermore, data augmentation prevents the model from overfitting to a single noise type and improves its performance in unseen noise conditions.

[0099] (4) Knowledge distillation: The student network is trained by minimizing the difference between the pseudo labels generated by the teacher network and the loss function is

[0100] (5) Teacher network update: The parameters of the teacher network are updated through EMA (Exponential Moving Average):

[0101] θ′ t ρ·θ s +(1-ρ)·θ t

[0102] Among them, θ′ t represents the updated parameters of the teacher network; θ t represents the parameters of the teacher network before updating; ρ represents the decay rate of EMA, which controls the speed of updating the teacher network parameters; θ s Represents the parameters of the student network.

[0103] In order to further improve the robustness and adaptability of the student network in dynamic and noisy environments, feature consistency loss, noise robustness loss and prediction confidence loss are added on the basis of knowledge distillation loss:

[0104]

[0105] in, represents the knowledge distillation loss, which measures the consistency between the student network prediction and the teacher network pseudo-label; Represents feature consistency loss, which measures the feature consistency between the student network and the teacher network in the target domain. in Pair Distribution The expectation of all possible target domain input signals The expected value of is the feature extracted by the student network, represents the features extracted by the teacher network; ||·|| 2 represents the Euclidean norm, which is used to measure the consistency of features; represents the noise robust loss, which measures the sensitivity of the student network to noisy data, Where ∈ represents the noise added to the target domain data, Var represents the variance, which is used to measure the sensitivity of the feature extractor to noise; represents the prediction confidence loss, which is used to measure the importance the student network attaches to high-confidence predictions. Among them, Conf(·) represents the prediction confidence, which is used to measure the confidence of the discriminator in the target domain features; MAE(·) represents the mean absolute error, which is used to measure the prediction accuracy of the model; v1, v2 and v3 represent weight coefficients, which are used to control the weights of different items in the total loss.

[0106] The above knowledge distillation loss The expression is as follows:

[0107]

[0108] in, is the student network loss function, which measures the error between the student network predicted position and the pseudo label generated by the teacher network; θ s are the parameters of the student network, through To optimize θ s ; is the expected value, which represents the expected value of all possible target domain samples; where X′ t and Y′ t is the pseudo label generated by the teacher network, indicating the true location of the target domain sample, and is the predicted position of the student network; B represents the batch size, which is used for batch gradient descent to improve computational efficiency; in and is the pseudo label of the b-th sample generated by the teacher network, and is the predicted position of the bth sample of the target domain data by the student network.

[0109] To improve the confidence of pseudo labels, this embodiment introduces a k-NN correction strategy, which includes the following steps:

[0110] (1) Uncertain pseudo-label recognition

[0111] For each target domain sample, noise is added repeatedly multiple times to generate pseudo labels, and the standard deviation is calculated; if the standard deviation exceeds a preset threshold, the pseudo label is considered uncertain.

[0112] (2) K-NN correction: For uncertain pseudo-labels, the weighted average correction is performed using the confidence pseudo-labels of the k nearest neighbors.

[0113]

[0114] Among them, (X′ t , Y′ t ) is the corrected pseudo-label, which indicates the target position used by the student network for training; K is the number of nearest neighbors, used to calculate the weighted average; ω i It is the comprehensive weight of the i-th nearest neighbor that combines distance, confidence and local density, and its expression is:

[0115]

[0116] Among them, d i is the distance between the uncertain sample and the i-th nearest neighbor, which is used to measure the similarity between the uncertain sample and the nearest neighbor sample. is the confidence of the i-th nearest neighbor, which is used to measure the reliability of the pseudo label and can be calculated by the standard deviation of multiple predictions; is the high-confidence pseudo label of the verified i-th nearest neighbor generated by the teacher network; Density(c i ) is the local density of the i-th nearest neighbor, which is used to measure the density of samples in the feature space and can be calculated by clustering algorithm; It is a normalization factor, which ensures that the sum of the weighted average is 1, prevents the weight from being too large or too small, and ensures that the corrected pseudo-label is reasonable.

[0117] It should be noted that the specific network structures of the teacher model and the student model in the "knowledge distillation network" are not limited. The output of this embodiment is the corrected pseudo-label output by the teacher model and the "predicted position" output by the student model.

[0118] Finally, the positioning module inputs the processed position signal into the positioning model to complete the large-space positioning in the passive domain. By analyzing the signals recorded by the receiver (in this embodiment, there are N receivers and one transmitter, and the signal emitted by the transmitter interacts with the objects in the environment and is captured by the receiver), the signals recorded by the receiver are mapped (i.e., the positioning model of this embodiment) Mapped to the real position of the object, the real position of the signal source (such as an RFID tag) can be inferred.

[0119] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A high-confidence large-space positioning method in passive domain, characterized by the following steps: include: Use the receiver to receive the position signal of the object to be located in the VR environment and perform normalization processing; Build a location positioning model based on the knowledge distillation network; The processed position signal is input into the position positioning model to complete large-space positioning in the passive domain.

2. The high-confidence large-space positioning method in the passive domain according to claim 1 is characterized in that: The position positioning model is constructed based on a knowledge distillation network. The knowledge distillation network implements self-supervised learning through a teacher-student network structure. The teacher network generates pseudo labels to guide the training of the student network. The student network uses these pseudo labels for training and adapts to the target domain data.

3. The high-confidence large-space positioning method in the passive domain according to claim 1, characterized in that: The position positioning model uses the Bayesian optimization framework to optimize the parameters θ of the position positioning model. Among them, θ represents the optimal parameter; It is a standard notation in optimization problems; Denotes the joint distribution P X，Y，Z expected value; Indicates the difference between the actual position (X, Y) and the predicted position in the X and Y directions The sum of the absolute errors of Where Var(·) represents the variance, represents the variance of the positioning error in the X direction, represents the variance of the positioning error in the Y direction; ψ(t) represents the time decay weight; T represents the total number of time steps; represents the sum of the absolute errors between the true position and the predicted position at time step t; α1, α2 and α3 are weight coefficients.

4. The high-confidence large-space positioning method in the passive domain according to claim 1, characterized in that: The method for training the positioning model includes: using multi-objective optimization to measure the accuracy, stability and dynamic adaptability of the model by combining absolute error, variance and time decay weight; at the same time, introducing time decay weight ψ(t) to adapt to dynamic environments: Among them, z t Represents the original input data of the target domain; Represents the target domain data after noise enhancement; ∈ i represents the i-th noise type; represents the indicator function, indicating whether the i-th noise type is selected, 1 represents selected, and 0 represents unselected; N represents the total number of noise types; p i represents the probability that the i-th noise type is selected.

5. The high-confidence large-space positioning method in passive domain according to claim 1, characterized in that: The position localization model adds feature consistency loss, noise robustness loss, and prediction confidence loss on the basis of knowledge distillation loss: in, represents the knowledge distillation loss; represents feature consistency loss; ||·|| 2 represents the Euclidean norm; represents the noise robust loss; Represents the prediction confidence loss; v1, v2, and v3 represent weight coefficients, which are used to control the weights of different items in the total loss.

6. The high-confidence large-space positioning method in passive domain according to claim 5, characterized in that: Knowledge Distillation Loss The expression is as follows: in, is the student network loss function; θ s are the parameters of the student network; Represents the expected value of all possible target domain samples; where X′ t and Y′ t is the pseudo label generated by the teacher network, indicating the true location of the target domain sample, and is the predicted position of the student network; B represents the batch size; in and is the pseudo label of the b-th sample generated by the teacher network, and is the predicted position of the bth sample of the target domain data by the student network.

7. A large-space positioning system with high confidence in a passive domain, the system being used to implement the method according to any one of claims 1 to 6, characterized in that: include: Receiving module, building module and positioning module; The receiving module is used to use a receiver to receive the position signal of the object to be located in the VR environment and perform normalization processing; The building module is used to build a location positioning model based on the knowledge distillation network; The positioning module is used to input the processed position signal into the position positioning model to complete large-space positioning in the passive domain.

8. The large-space positioning system with high confidence in the passive domain according to claim 7, characterized in that: The workflow of the construction module includes: constructing the position positioning model based on the knowledge distillation network; wherein the knowledge distillation network realizes self-supervised learning through a teacher-student network structure; the teacher network generates pseudo labels to guide the training of the student network; and the student network uses these pseudo labels for training to adapt to the target domain data.

Images