Positioning method, system and electronic equipment based on Bayesian filtering neural network

Abstract

This invention relates to the field of environmental perception fusion positioning technology, and provides a positioning method, system, and electronic device based on a Bayesian filtering neural network. The method includes: acquiring an observation waveform matrix generated by a base station after receiving a waveform signal transmitted by a target object at each sampling time; inputting the observation waveform matrix into a preset positioning model to obtain the positioning information of the target object; fitting the Bayesian filtering result to the positioning model using a recurrent neural network, which is trained using a cost function including motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm. This invention addresses the shortcomings of existing technologies where the application of Bayesian filtering algorithms in environmental perception fusion positioning requires approximation algorithms and complex derivations to obtain positioning results, thus affecting positioning speed and accuracy.

Description

Technical Field

[0001] This invention relates to the field of environmental perception fusion positioning technology, and in particular to a positioning method, system and electronic device based on a Bayesian filter neural network. Background Technology

[0002] Currently, when performing environmental perception fusion positioning, Bayesian filtering algorithms are introduced to link user information from multiple moments to overcome the problem that multipath information cannot improve the estimation accuracy of user location when there is no prior environmental information.

[0003] Traditional Bayesian filtering algorithms and their derivatives are based on Hidden Markov Models (HMMs). Given the motion and observation equations of the latent variable states, they can iteratively infer the posterior distribution of the latent variable states at a certain time, thereby estimating the latent variable. When both the motion and observation equations are linear equations, Bayesian filtering is equivalent to Kalman filtering. However, because the motion and observation equations are high-dimensional nonlinear equations in complex environmental fusion localization scenarios, it is difficult to represent the results of Bayesian filtering in a closed-form expression. Therefore, several approximation algorithms have emerged, such as the Extended Kalman Filter (EPF) algorithm and the Volumetric Kalman Filter (VKF) algorithm. For example, the Particle Filter algorithm is based on Monte Carlo estimation and uses importance sampling to fit the Bayesian filter.

[0004] Therefore, existing Bayesian filtering algorithms have the following problems when applied to environmental perception fusion and localization tasks:

[0005] In terms of accuracy, existing Bayesian filtering algorithms cannot obtain closed-form solutions for this task. Therefore, it is necessary to rely on approximation algorithms such as extended Kalman filtering and particle filtering. However, approximation algorithms in environment fusion localization problems depend on the adjustment of hyperparameters. Hyperparameters have a significant impact on the convergence speed and accuracy of Bayesian filtering algorithms and depend on precise tuning by technicians proficient in this task. In addition, regarding model simplicity, since environment fusion localization algorithms need to deal with complex problems such as multipath generation and extinction and matching, in order to integrate such complex observation models into Bayesian filtering algorithms in practical systems, complex derivations of the filtering algorithm are required. This limits the scalability of Bayesian filtering algorithms for such problems, thus affecting the speed and accuracy of localization. Summary of the Invention

[0006] This invention provides a positioning method, system, and electronic device based on a Bayesian filter neural network, which solves the problem that in the prior art, when applying the Bayesian filter algorithm in environmental perception fusion positioning, it is necessary to rely on approximate algorithms and complex derivations to obtain positioning results, thus affecting positioning speed and positioning accuracy.

[0007] This invention provides a localization method based on a Bayesian filter neural network, comprising:

[0008] At each sampling time, the base station generates the observed waveform matrix after receiving the waveform signal transmitted by the target object.

[0009] The observed waveform matrix is ​​input into a preset positioning model to obtain the positioning information of the target object output by the positioning model;

[0010] The localization model uses a recurrent neural network to fit the Bayesian filtering results. The recurrent neural network is trained using a cost function that includes motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm.

[0011] According to the localization method based on Bayesian filtering neural network of the present invention, the recurrent neural network is constructed based on long short-term memory network.

[0012] According to the localization method based on a Bayesian filter neural network of the present invention, the cost function further includes:

[0013] Prior constraints constructed based on other prior information during the training process of the recurrent neural network;

[0014] The other prior information includes prior information generated based on loop closure detection, initial value constraints, map constraints, and birth-death processes.

[0015] According to the positioning method based on Bayesian filter neural network of the present invention, the step of inputting the observed waveform matrix into a preset positioning model to obtain the positioning information of the target object output by the positioning model includes:

[0016] The observed waveform matrix is ​​input into the result fitting layer of the localization model. The result fitting layer uses the recurrent neural network to fit the Bayesian filtering result to obtain the posterior distribution of the target object output by the result fitting layer.

[0017] The posterior distribution is input to the positioning confirmation layer of the positioning model, and the positioning confirmation layer determines the positioning information of the target object based on the posterior distribution, thus obtaining the positioning information of the target object output by the positioning confirmation layer.

[0018] The present invention also provides a positioning system based on a Bayesian filter neural network, comprising:

[0019] The acquisition module is used to acquire the observation waveform matrix generated by the base station after receiving the waveform signal transmitted by the target object at each sampling time;

[0020] The positioning module is used to input the observed waveform matrix into a preset positioning model to obtain the positioning information of the target object output by the positioning model;

[0021] The localization model uses a recurrent neural network to fit the Bayesian filtering results. The recurrent neural network is trained using a cost function that includes motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm.

[0022] According to the positioning system based on Bayesian filtering neural network of the present invention, the recurrent neural network is constructed based on long short-term memory network.

[0023] According to the localization system based on a Bayesian filter neural network of the present invention, the cost function further includes:

[0024] Prior constraints constructed based on other prior information during the training process of the recurrent neural network;

[0025] The other prior information includes prior information generated based on loop closure detection, initial value constraints, map constraints, and birth-death processes.

[0026] According to the positioning system based on Bayesian filter neural network of the present invention, the positioning module is specifically used for:

[0027] The observed waveform matrix is ​​input into the result fitting layer of the localization model. The result fitting layer uses the recurrent neural network to fit the Bayesian filtering result to obtain the posterior distribution of the target object output by the result fitting layer.

[0028] The posterior distribution is input to the positioning confirmation layer of the positioning model, and the positioning confirmation layer determines the positioning information of the target object based on the posterior distribution, thus obtaining the positioning information of the target object output by the positioning confirmation layer.

[0029] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the localization method based on a Bayesian filter neural network as described above.

[0030] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the localization method based on a Bayesian filter neural network as described above.

[0031] The present invention provides a positioning method, system, and electronic device based on a Bayesian filtering neural network. This method acquires the observed waveform matrix generated by the base station after receiving the waveform signal transmitted by the target object at each sampling time. The observed waveform matrix is ​​then input into a preset positioning model to obtain the positioning information of the target object output by the positioning model. The positioning model uses a recurrent neural network to fit the Bayesian filtering result. The recurrent neural network is trained using a cost function including motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm. This method achieves the fitting of the Bayesian filtering result using a recurrent neural network, thus solving the problem of the impact on positioning speed and accuracy when applying the Bayesian filtering algorithm in environmental perception fusion positioning. This significantly improves the scalability of traditional environmental perception fusion positioning algorithms and enhances their convergence speed and positioning accuracy. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating a localization method based on a Bayesian filter neural network provided in an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of a hidden Markov model provided in an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of a Bayesian filter network structure provided in an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of the structure of a positioning system based on a Bayesian filter neural network provided in an embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0039] First, to facilitate understanding of the localization method based on Bayesian filtering neural network provided by this invention, the technical solution provided by the embodiments of this invention will be described below using ultra-wideband (UWB) signals as an example.

[0040] It should be noted that the ultra-wideband signal is only one example of the signal types to which the positioning method based on Bayesian filtering neural network provided in this embodiment of the invention can be applied. The positioning method based on Bayesian filtering neural network provided in this embodiment of the invention can also be applied to other types of signals, such as 5G NR, WIFI, Bluetooth, geomagnetic sensing, etc., which will not be elaborated here.

[0041] It is understandable that a single-base station ultra-wideband positioning system includes one The antenna is located at the base station and the user equipment (UE), and the center position of the base station array is known. ,exist At any given time, the location of the unknown UE is defined as... Considering the environment has The location of the reflective surface is unknown, and the base station can be used relative to the first... The image of the first reflecting surface (defined as the virtual base station VA) is used to describe the first reflecting surface. The parameters of each reflector relative to the base station will then be The time is unknown The position of each VA is represented as ,and Meanwhile, for the sake of model simplicity, the 0th VA is assumed to be the known base station location, i.e. .

[0042] Because, through virtual base stations Base station The location of the virtual tag can be calculated based on the UE's location. Furthermore, based on Fresnel's theorem, all MPC (Multipath Component) signals received by the base station can be considered as virtual tags. The transmitted signal. Therefore, at every moment... It can be estimated by receiving waveforms. Dually, the position of the UE can be estimated from the waveform. and the location of all virtual base stations (VAs) .

[0043] Furthermore, after receiving the UWB waveform transmitted by the UE, the base station can record each received CIR (Channel Impulse Response) as... Its signal model is as follows:

[0044] (1)

[0045] in, It is an echo signal. It is the first The guide vector of each MPC, It is Gaussian white noise. It is the first The echo signal received at each sampling time.

[0046] By All the unknown variables at time 1 are connected together and represented as: ,Right now:

[0047] (2)

[0048] At each sampling time ,by The sampling frequency can be collected Each snapshot is accumulated to form an observation waveform matrix, i.e.:

[0049] (3)

[0050] It is known that, through parameter estimation algorithms, it is possible to directly obtain... The ToA (Time of Arrival) and AoA (Angle of Arrival) parameters of each MPC are estimated. Therefore, given Able to directly use geometric mapping functions Calculate the location of the UE to be estimated and the VA ,Right now:

[0051] (4)

[0052] Then we have:

[0053] (5)

[0054] It is evident that by estimating MPC, we can obtain... Parameters to be estimated at time posterior distribution This allows us to obtain estimates of the user's location and the location of the virtual base station.

[0055] However, without prior environmental information, multipath information (MPC) cannot improve the accuracy of user location estimation. Therefore, it is necessary to use a Bayesian filtering algorithm to combine information from multiple time points, i.e., in... k Time calculation .

[0056] However, existing Bayesian filtering algorithms have the following problems when applied to environmental perception fusion and localization tasks:

[0057] In terms of accuracy, existing Bayesian filtering algorithms cannot obtain closed-form solutions for this task, thus requiring the reliance on approximation algorithms such as extended Kalman filtering and particle filtering. However, these approximation algorithms depend on hyperparameter tuning in environment fusion localization problems. Hyperparameters significantly impact the convergence speed and accuracy of Bayesian filtering algorithms and require precise tuning by skilled technicians proficient in the task. Furthermore, regarding model simplicity, environment fusion localization algorithms need to address complex issues such as multipath generation and destruction, and matching. In practical systems, integrating such complex observation models into Bayesian filtering algorithms requires complex derivations, limiting the scalability of Bayesian filtering algorithms for these types of problems.

[0058] Based on this, embodiments of the present invention provide a localization method that uses a recurrent neural network to fit the Bayesian filtering results, thereby solving the convergence and scalability problems faced by the current Bayesian filtering algorithm when applied to environmental fusion localization tasks, and thus improving the localization speed and accuracy.

[0059] The following is combined Figures 1 to 3 This invention describes a localization method based on a Bayesian filter neural network, which can be executed by software and / or hardware in electronic devices such as computers, tablets, and mobile phones. Figure 1 As shown, the method includes the following steps:

[0060] 101. Obtain the observation waveform matrix generated by the base station after receiving the waveform signal transmitted by the target object at each sampling time.

[0061] It is understandable that, taking the single-base station ultra-wideband positioning system, as described above, which includes a base station with a P-antenna and a user equipment (UE), the target objects refer to the UE and the virtual base station. k At a given moment, after receiving the waveform signal transmitted by the target object, the base station generates the observed waveform matrix as shown in formula (3) above. .

[0062] 102. Input the observed waveform matrix into the preset positioning model to obtain the positioning information of the target object output by the positioning model;

[0063] The localization model uses a recurrent neural network to fit the Bayesian filtering results. The recurrent neural network is trained using a cost function that includes motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm.

[0064] In this embodiment, a recurrent neural network is trained using a cost function that includes motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm. This allows the recurrent neural network to adaptively fit the Bayesian filtering results, thereby solving the convergence and scalability problems faced by existing Bayesian filtering algorithms when applied to environmental perception fusion and localization tasks.

[0065] The following details the network structure and training process of the Bayesian filtering network (BFN) used in the localization model provided in this embodiment of the invention, to demonstrate that, based on the network structure and corresponding cost function described in this embodiment of the invention, the output of the recurrent neural network is a maximum likelihood estimate of the output of the Bayesian filtering algorithm.

[0066] First, such as Figure 2 The example shown uses a hidden Markov model to... Modeling is performed, among which This indicates that we are in a relatively stable environment. The discussion will proceed as follows, assuming no drastic changes in the environment during the exercise. Then, as... Figure 2 The diagram shows the derivation of the fitting of the Bayesian filtering algorithm using a recurrent neural network under the Hidden Markov Model.

[0067] In one specific embodiment, the parameters are used. Defined recurrent neural network distribution right Fitting is then performed. This involves adjusting the parameters of the recurrent neural network. To estimate the parameters using maximum likelihood estimation, we have:

[0068] (6)

[0069] in, It is a pre-defined model of uniform linear motion, that is, it assumes... Time and Throughout the time interval, the user maintains a uniform linear motion. This is because, in indoor positioning scenarios with high sampling frequencies and slow user movement speeds, the motion model can be considered approximately satisfied. This is the likelihood function given by the parameter estimation algorithm at a single time step. Therefore, in order to fit the Bayesian filtering algorithm under the Hidden Markov Model, it is necessary to use the distribution... Generate a series of samples Then, gradient descent is performed on these samples using the cost function shown in Equation (7). After training is complete, The Bayesian filtering result can then be output.

[0070] (7)

[0071] Therefore, recurrent neural networks can be designed. Structure to meet the requirements Efficient learning under spatiotemporal constraints. Among these constraints is the UE location. There is a motion correlation, therefore it is assumed that the user is... and It maintains uniform linear motion at any given moment. These are the coordinates of the virtual base station (VA). In typical indoor scenarios, the VA, relative to the communication frequency between the UE and the base station, will only undergo unpredictable and slow migrations as the environment evolves. Therefore, it can be considered... It follows a Brownian motion process, that is:

[0072] (8)

[0073] Therefore, based on the above embodiments, the recurrent neural network is constructed based on the long short-term memory network.

[0074] Even when modeling with Long Short-Term Memory (LSTM) networks , and using additional A virtual base station is represented by a random variable (one base station requires two coordinates to be represented in a two-dimensional map) that can be gradient descent. .

[0075] Furthermore, additional prior information generated based on loop closure detection, initial value constraints, map constraints, birth and death conditions, etc., can be added to the cost function to train the recurrent neural network, thereby improving its ability to constrain map information.

[0076] Therefore, based on the above embodiments, the cost function further includes:

[0077] Prior constraints are constructed based on other prior information during the training process of the recurrent neural network.

[0078] Other prior information includes prior information based on loop closure detection, initial value constraints, map constraints, and birth-death processes.

[0079] In summary, the network structure provided by the embodiments of the present invention can be as follows: Figure 3 As shown. And based on Figure 3 The training process of the localization model that uses a recurrent neural network to fit the Bayesian filtering results, as shown in the network structure of this embodiment of the invention, specifically includes:

[0080] First, in BFN, the parameters that can be gradient descent include the parameters of the LSTM. and the coordinates of all virtual base stations ,Right now:

[0081]

[0082] like Figure 3 As shown, LSTM at each time step... and the memory passed on from the previous moment For input, use a Gaussian distribution. Approach At each time step, the LSTM outputs... The mean and variance of the variables are calculated using a random variable that follows a standard normal distribution. Resample the distribution to obtain the sampling results. Each sampling result can be viewed as a node in a Hidden Markov Model. For each node... All of the following three constraints are applied:

[0083] 1. Motion Constraints: Combine the motion model to generate motion likelihood. .

[0084] 2. Observation Constraints: Based on the parameter estimation results, construct the observation likelihood. .

[0085] 3. Prior Constraints: Based on additional prior information such as loop closure detection, initial value constraints, map constraints, birth and death conditions, etc., additional prior information can be applied to each node. .

[0086] After training is complete, the output of LSTM is This is a recurrent neural network. The result of fitting the Bayesian filter. It can be seen that, through methods such as... Figure 3The BFN shown not only stabilizes the movement of the virtual base station and imposes strong constraints on map information, but also enhances the temporal correlation of user states using a recurrent neural network, thereby improving the neural network's learning ability regarding temporal correlation properties. Furthermore, by modeling unknown variables using BFN and nodalizing them, not only can Bayesian filtering algorithms be easily implemented, but complex constraints in traditional SLAM algorithms, such as the birth-death process, map constraints, and the generation of matching information, which require redesigning the filtering algorithm, can also be quickly incorporated. Thus for It offers greater potential for deployment in more complex scenarios.

[0087] The above method, based on a Hidden Markov Model (HMM), represents the nodes to be estimated in the HMM using the output of a recurrent neural process. It constructs a cost function using the likelihoods represented by the edges of the motion model, observation model, and prior model to train the recurrent neural network, achieving the goal of deriving a Bayesian filtering algorithm for environmental perception fusion localization based on a recurrent neural network. The Bayesian filtering network provided in this embodiment can perform maximum likelihood estimation on the Bayesian filtering results. Furthermore, its node and edge design allows for strong scalability, making it suitable for complex scenarios such as multipath generation / death, false alarms, and matching. This significantly improves the scalability of traditional environmental perception fusion localization algorithms and enhances their convergence speed and localization accuracy.

[0088] Based on the above embodiments, the observed waveform matrix is ​​input into a preset positioning model to obtain the positioning information of the target object output by the positioning model, including:

[0089] The observed waveform matrix is ​​input into the result fitting layer of the localization model. The result fitting layer uses a recurrent neural network to fit the Bayesian filtering result, and obtains the posterior distribution of the target object output by the result fitting layer.

[0090] The posterior distribution is input into the positioning confirmation layer of the positioning model. The positioning confirmation layer determines the positioning information of the target object based on the posterior distribution, and the positioning information of the target object is output by the positioning confirmation layer.

[0091] In this embodiment, by setting a result fitting layer in the positioning model that uses a recurrent neural network to fit the Bayesian filtering results, the positioning speed and accuracy are effectively accelerated.

[0092] The positioning method based on a Bayesian filter neural network provided in the above embodiments of the present invention can be used in application scenarios requiring positioning and navigation, such as smart factories and emergency search and rescue where satellite rejection or weak signals are present. Furthermore, the Bayesian filter network used in the positioning model, as a Bayesian filtering algorithm, is applicable to various situations including positioning and navigation, satellite tracking, and remote sensing transmission.

[0093] Based on the same general inventive concept, this invention also protects a positioning system based on a Bayesian filtering neural network. The positioning system based on a Bayesian filtering neural network provided by this invention will be described below. The positioning system based on a Bayesian filtering neural network described below can be referred to in correspondence with the positioning method based on a Bayesian filtering neural network described above.

[0094] Figure 4 This is a schematic diagram of the positioning system based on a Bayesian filter neural network provided by the present invention. Figure 4 As shown, it includes: an acquisition module 410 and a positioning module 420; wherein,

[0095] The acquisition module 410 is used to acquire the observation waveform matrix generated by the base station after receiving the waveform signal transmitted by the target object at each sampling time;

[0096] The positioning module 420 is used to input the observed waveform matrix into the preset positioning model to obtain the positioning information of the target object output by the positioning model;

[0097] The localization model uses a recurrent neural network to fit the Bayesian filtering results. The recurrent neural network is trained using a cost function that includes motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm.

[0098] The positioning system based on a Bayesian filtering neural network provided in this invention obtains the observed waveform matrix generated by the base station after receiving the waveform signal transmitted by the target object at each sampling time. This observed waveform matrix is ​​then input into a preset positioning model to obtain the positioning information of the target object output by the positioning model. The positioning model uses a recurrent neural network to fit the Bayesian filtering result. The recurrent neural network is trained using a cost function including motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm. This system achieves the fitting of the Bayesian filtering result using a recurrent neural network, thus solving the problem of the impact on positioning speed and accuracy when applying the Bayesian filtering algorithm in environmental perception fusion positioning. This significantly improves the scalability of traditional environmental perception fusion positioning algorithms and enhances their convergence speed and positioning accuracy.

[0099] Based on the above embodiments, the recurrent neural network is constructed based on the long short-term memory network.

[0100] Based on the above embodiments, the cost function further includes:

[0101] Prior constraints are constructed based on other prior information during the training process of recurrent neural networks;

[0102] Other prior information includes prior information based on loop closure detection, initial value constraints, map constraints, and birth-death processes.

[0103] Based on the above embodiments, the positioning module 420 is specifically used for:

[0104] The observed waveform matrix is ​​input into the result fitting layer of the localization model. The result fitting layer uses a recurrent neural network to fit the Bayesian filtering result, and obtains the posterior distribution of the target object output by the result fitting layer.

[0105] The posterior distribution is input into the positioning confirmation layer of the positioning model. The positioning confirmation layer determines the positioning information of the target object based on the posterior distribution, and the positioning information of the target object is output by the positioning confirmation layer.

[0106] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communication interface 520, a memory 530, and a communication bus 540. The processor 510, communication interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a positioning method based on a Bayesian filtering neural network. This method includes: acquiring an observation waveform matrix generated by the base station after receiving the waveform signal transmitted by the target object at each sampling time; inputting the observation waveform matrix into a preset positioning model to obtain the positioning information of the target object output by the positioning model; wherein the positioning model uses a recurrent neural network to fit the Bayesian filtering result, the recurrent neural network is trained using a cost function including motion likelihood and observation likelihood, the motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm.

[0107] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the positioning method based on a Bayesian filtering neural network provided by the above methods. The method includes: acquiring an observation waveform matrix generated by a base station after receiving a waveform signal transmitted by a target object at each sampling time; inputting the observation waveform matrix into a preset positioning model to obtain the positioning information of the target object output by the positioning model; wherein the positioning model uses a recurrent neural network to fit the Bayesian filtering result, the recurrent neural network is trained by a cost function including motion likelihood and observation likelihood, the motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm.

[0109] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the positioning method based on a Bayesian filtering neural network provided by the above methods. The method includes: acquiring an observation waveform matrix generated by a base station after receiving a waveform signal transmitted by a target object at each sampling time; inputting the observation waveform matrix into a preset positioning model to obtain the positioning information of the target object output by the positioning model; wherein the positioning model uses a recurrent neural network to fit the Bayesian filtering result, the recurrent neural network is trained through a cost function including motion likelihood and observation likelihood, the motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm.

[0110] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A localization method based on a Bayesian filter neural network, characterized in that, include: At each sampling time, the base station generates the observed waveform matrix after receiving the waveform signal transmitted by the target object. The observed waveform matrix is ​​input into a preset positioning model to obtain the positioning information of the target object output by the positioning model; The localization model uses a recurrent neural network to fit the Bayesian filtering results. The recurrent neural network is trained using a cost function that includes motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm. The cost function further includes: Prior constraints constructed based on other prior information during the training process of the recurrent neural network; The other prior information includes prior information generated based on loop closure detection, initial value constraints, map constraints, and birth-death processes.

2. The localization method based on a Bayesian filter neural network according to claim 1, characterized in that, The recurrent neural network is constructed based on a long short-term memory network.

3. The localization method based on a Bayesian filter neural network according to any one of claims 1 to 2, characterized in that, The step of inputting the observed waveform matrix into a preset positioning model to obtain the positioning information of the target object output by the positioning model includes: The observed waveform matrix is ​​input into the result fitting layer of the localization model. The result fitting layer uses the recurrent neural network to fit the Bayesian filtering result to obtain the posterior distribution of the target object output by the result fitting layer. The posterior distribution is input to the positioning confirmation layer of the positioning model, and the positioning confirmation layer determines the positioning information of the target object based on the posterior distribution, thus obtaining the positioning information of the target object output by the positioning confirmation layer.

4. A positioning system based on a Bayesian filter neural network, characterized in that, include: The acquisition module is used to acquire the observation waveform matrix generated by the base station after receiving the waveform signal transmitted by the target object at each sampling time; The positioning module is used to input the observed waveform matrix into a preset positioning model to obtain the positioning information of the target object output by the positioning model; The localization model uses a recurrent neural network to fit the Bayesian filtering results. The recurrent neural network is trained using a cost function that includes motion likelihood and observation likelihood. The motion likelihood is generated based on the motion model of the target object sample, and the observation likelihood is generated based on the position information of the target object sample determined by a parameter estimation algorithm. The cost function further includes: Prior constraints constructed based on other prior information during the training process of the recurrent neural network; The other prior information includes prior information generated based on loop closure detection, initial value constraints, map constraints, and birth-death processes.

5. The positioning system based on a Bayesian filter neural network according to claim 4, characterized in that, The recurrent neural network is constructed based on a long short-term memory network.

6. The positioning system based on a Bayesian filter neural network according to any one of claims 4 to 5, characterized in that, The positioning module is specifically used for: The observed waveform matrix is ​​input into the result fitting layer of the localization model. The result fitting layer uses the recurrent neural network to fit the Bayesian filtering result to obtain the posterior distribution of the target object output by the result fitting layer. The posterior distribution is input to the positioning confirmation layer of the positioning model, and the positioning confirmation layer determines the positioning information of the target object based on the posterior distribution, thus obtaining the positioning information of the target object output by the positioning confirmation layer.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the localization method based on a Bayesian filter neural network as described in any one of claims 1 to 3.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the localization method based on a Bayesian filter neural network as described in any one of claims 1 to 3.

Images