A wireless SLAM method based on multi-base station information fusion

Abstract

The present invention discloses a wireless SLAM method based on multi-base station information fusion, in which user positioning and environment mapping are performed respectively by utilizing the different characteristics of line-of-sight (LoS) and non-line-of-sight (NLoS) channel parameters. Based on the extended Kalman filter (EKF) framework, the user position is predicted by fusing the LoS path angle parameters of multiple base stations. In terms of environment mapping, the present invention proposes to use virtual reference points instead of virtual anchor points to model the reflection surface, so as to form a consistent environment description between different base stations, and construct a unified map fusion framework for different scenarios. The present invention uses the user position predicted by EKF fusion of LoS path channel parameters as the prior of the PHD-SLAM particle filter, which effectively improves the particle efficiency. Finally, the method realizes the fusion perception of multi-base station LoS and NLoS information, effectively improves the perception performance, and significantly reduces the computational burden.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication perception integration and wireless positioning, and in particular relates to a wireless SLAM method based on multi-base station information fusion. Background Art

[0002] Future wireless communication technologies will be very similar to radar technology in terms of hardware architecture, frequency bands, and signal processing methods, enabling them to provide ubiquitous sensing services. Furthermore, Integrated Sensing and Communication (ISAC), which integrates communication and sensing functions, can significantly improve spectrum and energy efficiency.

[0003] Current mainstream SLAM technology solutions rely primarily on various sensors, such as LiDAR, cameras, and wireless signals. While LiDAR-based SLAM is renowned for its high accuracy and stability, it also suffers from significant drawbacks: high cost and susceptibility to inclement weather. While visual SLAM can provide rich environmental information at a low cost, its performance is significantly compromised in situations with drastic lighting variations or sparse scene textures. In contrast, radio SLAM leverages ubiquitous wireless signals (such as Wi-Fi, Bluetooth, and 5G) for positioning and mapping. This is not only cost-effective but also unaffected by lighting and weather conditions, making it more adaptable in complex environments. However, relying solely on data from a single sensor or device for perception has inherent limitations. For example, the positioning accuracy of a single wireless device can be affected by multipath effects and environmental noise. To overcome these limitations, information fusion technology has become an important research direction.

[0004] Currently, many SLAM systems based on multi-sensor fusion or multi-device collaboration have been proposed, such as SLAM systems that integrate inertial measurement units (IMUs) and wireless signals, and systems that use multiple mobile devices for collaborative map construction. These methods effectively improve the accuracy and stability of positioning and map construction. However, existing research mainly focuses on multi-sensor or multi-device fusion within the coverage area of ​​a single base station. In contrast, there is relatively little research on wireless SLAM systems that use multiple base stations to work together. Multi-base station collaboration can significantly expand the coverage area of ​​wireless SLAM systems and provide richer spatial information, thereby further improving the performance of positioning and map construction, especially in large and complex environments. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a wireless SLAM method based on multi-base station information fusion to achieve user positioning and environmental perception during beam scanning, improve user position estimation results by using multi-base station LoS path fusion results, reduce particle filter complexity by using LoS path position estimation as NLoS map fusion prior, and improve map estimation by fusing NLoS paths using information fusion algorithm, ultimately achieving fusion perception in multi-base station scenarios with low complexity.

[0006] Technical solution: A wireless SLAM method based on multi-base station information fusion of the present invention comprises the following steps:

[0007] Step 1: Multiple base stations and users use full-angle beam scanning to obtain the current beam heat map and extract several candidate beam angle pairs from the heat map;

[0008] Step 2: Multiple base stations filter out multi-base line-of-sight (LoS) path information from candidate beam pairs. Combining the user position estimate from the previous moment's SLAM algorithm and the pose prior information output by the IMU, the user's current position is predicted. An extended Kalman filter is used to fuse the user position prediction and the LoS path channel parameters to obtain a partial a posteriori estimate of the user's position.

[0009] Step 3: For each base station, the non-line-of-sight (NLoS) path information is filtered from the candidate beam pairs. The user position output by the extended Kalman filter is used as the user position prior of the SLAM algorithm. The particle filter is used to scatter the prior to approximate the user distribution.

[0010] Step 4: For each base station, the user's possible NLoS path channel parameters are predicted using the SLAM map estimate prior and the user's location points at the previous moment. The SLAM algorithm is then used to associate the NLoS path channel parameters with the map estimate prior, thereby completing a single-base station posterior estimate of the radio map.

[0011] Step 5: Upload the local map estimation results of multiple base stations to the data fusion center, perform map fusion according to different fusion scenarios, and complete the posterior estimation of the map of multiple base stations;

[0012] Step 6: Based on the SLAM radio map feature posterior estimation and the multi-base station NLoS path channel parameters, the weight of each user particle is calculated. Based on the weight, the user particles are re-scattered using the importance resampling algorithm to complete the probabilistic scattering of user locations.

[0013] Furthermore, step 1 is specifically as follows: multiple base stations and users use uplink full-angle beam scanning, the user uses the DFT codebook to form beams in different directions, the base station uses the DFT matrix to receive incident signals from different directions, and forms a beam heat map of the entire space. Using a continuous elimination scheme, the strongest path is sequentially extracted from the beam heat map and deleted until the energy stop threshold is met. The path with the strongest energy among the extracted paths is used as the LoS path, and the remaining paths are used as NLoS paths. The information output during the full-angle beam scanning stage is:

[0014]

[0015] in, represents the estimated channel parameter pair, represents the estimated AoA of the lth path, represents the estimated AoD of the lth path, Indicates the estimated number of multipaths.

[0016] Furthermore, step 2 specifically includes the following steps:

[0017] Step 2.1: Extract the LoS path channel parameters from the channel parameters of each base station-user channel using a LoS path identification method. The LoS path identification method includes a method based on received signal energy, a method based on received signal delay, and a method based on a priori geometric models.

[0018] Step 2.2: Analyze the geometric constraints satisfied by the user position and the base station position to obtain the relationship between the LoS path channel parameters and the user position. Use the user position output by the SLAM algorithm at the previous moment as prior information, combined with the established assumptions made about the user motion model, which includes the uniform motion model, uniform acceleration motion model, segmented uniform acceleration motion model, and random walk model; and use the pose information output by the IMU to predict the user position at the current moment.

[0019] Step 2.3: Use Taylor expansion to expand the constraint relationship between the LoS path channel parameters and the user position at the user's predicted position, retain the first-order terms and discard high-order small quantities to obtain a linear approximation constraint relationship. Use a single-target parameter estimation method to obtain a posterior estimate of the user's current position based on the LoS channel parameters. The single-target parameter estimation method includes LS and extended Kalman filter.

[0020] Furthermore, step 3 is specifically as follows: for the local SLAM algorithm of each base station, the user position posterior information is used as the point scattering prior of the particle filter. The particle filter is based on the distribution of user positions, including Gaussian distribution, uniform distribution, and Poisson distribution; the user position is scattered according to probability, and the accuracy of user position prediction and the efficiency of particle scattering are improved by the fusion result of LoS path.

[0021] Furthermore, step 4 is specifically as follows: for the local SLAM algorithm of each base station, the map posterior distribution of the base station at the current moment is estimated through the multi-base station LoS path user position scattering and the SLAM map estimation result of the previous moment, which specifically includes two parts: the legacy part of the previous moment and the newly observed part of the current moment; for a static environment, the legacy map features of the previous moment remain unchanged; for the newly observed feature part, the feature prediction of the current moment is generated through the empty set method, the single feature method, and the observation-based method, and the radio map feature estimation is updated by using the filter fusion NLoS path channel parameters, including the confidence propagation filter and the probability hypothesis strength filter, to complete the local map posterior estimation of a single base station based on the NLoS path parameters.

[0022] Furthermore, step 5 specifically includes the following steps:

[0023] Step 5.1: Upload the local map posterior estimates of each base station to the data fusion center via optical cables for fusion. Different information fusion strategies are adopted according to the different characteristics of the test scenario.

[0024] Step 5.2: When the fields of view of each base station in the environment do not overlap or only partially overlap, the weighted averaging method is used to complete map fusion by weighted averaging the map features of each base station, retaining the features observed by all base stations. This method is suitable for scenarios with high missed detection rates.

[0025] Step 5.3: When the fields of view of each base station in the environment completely overlap, the geometric mean method is used. This method converts the exponential geometric mean of the Gaussian mixture into a geometric mean Gaussian mixture. This method preserves the Gaussian mixture characteristics of the map and the common characteristics observed by each base station, making it suitable for scenarios with high false alarm rates.

[0026] Step 5.4: For scenarios with high complexity requirements, a sequential fusion approach is used. This approach does not require additional fusion steps and sequentially completes map estimation between base stations. This approach is suitable for fast fusion scenarios with overlapping visual fields.

[0027] Step 5.5: In view of the fact that the user's field of view changes with location, a cloud-based fusion algorithm based on weighted average is used to screen out convergence components whose weights and variances meet the threshold from the local map of each base station and upload them to the cloud. The uploaded converged map is fused with the cloud map. The cloud only fuses the map components without pruning them, and sends the fused map to each base station as the prior at the next moment. In this way, the estimation of the local map and the global map is completed at the base station and cloud respectively.

[0028] Furthermore, step 6 is specifically as follows: based on the constraint relationship between the virtual reference anchor point position, the user position, and the NLoS path channel parameters, the NLoS path channel parameters at the current moment are predicted; by correlating the observed and predicted channel parameters at the current moment, including implicit and explicit correlation methods, the relative probability of each user particle is calculated, and the user positions are re-scattered based on the importance resampling algorithm to achieve a result that conforms to the probability distribution, with more points scattered at high-probability locations and fewer points scattered at low-probability locations, avoiding particle degradation, and averaging the resampled results to complete the global posterior estimate of the user position.

[0029] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0030] 1. Existing SLAM algorithms based on virtual anchors do not have a unified wireless map representation across different base stations. This method maps virtual anchors to a unified virtual reference point and uses it as a new infinite map feature, forming a consistent representation of the wireless environment across multiple base stations.

[0031] 2. The number of particles required for convergence of existing particle-based SLAM algorithms is too high, which causes the number of features in the fusion algorithm to increase exponentially, making the fusion complexity too high to meet real-time requirements. By designing an extended Kalman filter to fuse multiple base station LoS paths, the user position prediction accuracy is improved, and the number of particles required for SLAM algorithm convergence and the complexity of the fusion algorithm are greatly reduced;

[0032] 3. Existing fusion algorithms are often targeted at single scenarios and struggle to adapt to complex and changing wireless environments. By implementing weighted averaging, geometric averaging, sequential fusion, and cloud-based fusion strategies within a unified framework, fusion algorithms can be flexibly adapted to different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Flowchart of the present invention.

[0034] Figure 2 Performance comparison of different multi-base station information fusion SLAM schemes in fully overlapping FoV scenarios, including (a) positioning error, (b) mapping error, and (c) complexity performance.

[0035] Figure 3 Comparison of Cloud fusion performance in partially overlapping FoV scenarios, including (a) positioning error and (b) mapping error. DETAILED DESCRIPTION

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] like Figure 1 As shown, the present invention provides a wireless SLAM method based on multi-base station information fusion, comprising the following steps:

[0038] Step 1: Multiple base stations and users use uplink full-angle beam scanning to obtain the channel parameters at the current user location. For each beam angle pair, the RSRP value of the main synchronization signal of the received signal SSB block is used as the received signal energy estimate on the direction pair. The full-angle beam heat map is obtained by combining the RSRP values ​​on different angle pairs. Through angle parameter extraction methods, including but not limited to maximum value extraction method, continuous elimination method, deep learning feature extraction method, etc., the channel parameter angle estimation values ​​of several LoS and NLoS are obtained. Therefore, the information output in the full-angle beam scanning stage is

[0039]

[0040] where represents the estimated channel parameter pair, represents the estimated AoA of the lth path, represents the estimated AoD of the lth path, Indicates the estimated number of multipaths.

[0041] Step 2: From the channel parameters output in step 1, use LoS path identification methods, including but not limited to methods based on received signal energy, received signal delay, and geometric model priors, to extract the LoS path channel parameters from each base station-user channel parameter. By analyzing the geometric constraints satisfied by the user position and the base station position, the relationship between the LoS path channel parameters and the user position is obtained. Assume that the LoS path channel parameters estimated by the kth base station are The constraint relationship can be expressed as:

[0042]

[0043] Among them, x t =[x t ,y t ] T and are the user location at time t and the location of the kth base station respectively. Then, the results obtained from different base stations can be integrated into the following matrix equation:

[0044] Ax t =b

[0045] Among them, the coefficient matrix and constant term Its elements are defined as:

[0046]

[0047] The user location estimation result of LS can be obtained as x t =(A T A) -1 A T b. It should be noted that although each base station can obtain two equations based on the arrival angle and departure angle, the two equations are linearly related due to the complementary relationship between the arrival angle and the departure angle under LoS conditions. Therefore, to perform position estimation in two-dimensional space, at least two base stations are required to use the LS estimation method.

[0048] The user's motion is modeled as piecewise constant acceleration motion, which means that the user's acceleration remains unchanged between two adjacent observation moments. The user's posture is represented by a 4-dimensional state vector s t =[x t ,y t ,v x,t ,v y,t ] T , including position coordinates and velocity components, the acceleration information output by IMU is u t , then the user's motion model is:

[0049] s t∣t-1 =F1s t-1 +F2(u t-1 +ε)

[0050] Among them, ε is the Gaussian white noise of the IMU sensor, and the specific expressions of the coefficient matrices F1 and F2 are:

[0051]

[0052] Where ΔT is the observation time interval. The LoS channel parameter observation and user position coordinates are simply expressed as And the first-order Taylor expansion is:

[0053]

[0054] Among them, x t∣t-1 represents the user's position predicted based on the motion model and IMU data, Represents the function h in x t∣t-1The Jacobian matrix at , and Denotes the Peano remainder of the Taylor expansion. Therefore, the steps of using the extended Kalman filter fusion estimation can be divided into prediction and update:

[0055] Prediction: Given the estimated value x of the state vector at the previous moment t∣t-1 And the corresponding covariance matrix P t-1 , then the current state information prediction is:

[0056] s t∣t-1 =F1s t-1 +F2u t-1 ,

[0057]

[0058] Where Q is the IMU control signal covariance matrix.

[0059] Update: When new angle parameters are estimated Upon arrival, the state vector and covariance matrix are updated as:

[0060]

[0061] P t =P t∣t-1 -K t H t P t∣t-1 ,

[0062] Where Kt represents the Kalman gain and R represents the covariance matrix of the LoS angle parameter estimates.

[0063] Step 3: For each base station's local SLAM algorithm, the user position posterior information output from step 2 is used as the particle filter's point scattering prior. The particle sample variance is used as the user position variance prior. The particle filter scatters the user position based on the probability distribution, including but not limited to Gaussian distribution, uniform distribution, Poisson distribution, etc. The fusion result of the LoS path improves the accuracy of user position prediction, thereby improving the efficiency of particle scattering.

[0064] Step 4: For each base station's local SLAM algorithm, estimate the posterior distribution of the base station's map at the current moment using the multi-base station LoS path user position points output in step 3 and the SLAM map estimation result at the previous moment. A set of N particles is used to represent the user position and the RF map:

[0065]

[0066] in, is the weight of the i-th particle, The estimated user position for this particle, is the first-order moment of the posterior multi-target RFS radio frequency map, also known as PHD. For the particle filter implementation of PHD-SLAM, is approximated by the following GM model:

[0067]

[0068] in, is the number of GM components, and are the corresponding means and covariances, is the weight of the jth GM component. The main difference between the VRP-based PHD-SLAM and the VA-based scheme is the way the new features are generated. Specifically, the channel parameters The relationship between it and VRP is simply as follows:

[0069]

[0070] in, is the measurement error. Perform a first-order Taylor expansion at , and after some simple transformations we can get:

[0071]

[0072] in is the Jacobian matrix. Using the map prior of the previous moment, the user's position prediction and the new observation information, the PHD map of the current moment is predicted:

[0073]

[0074] in, is the generated new PHD component. When the new channel parameter estimate arrives, the particle weight of the user position and the map PHD can be updated using the particle filter and PHD filter respectively:

[0075]

[0076] in, is the transition probability of the state transition function, is the detection probability, is the observation likelihood function. This completes the posterior estimation of the local map of a single base station based on the NLoS path parameters;

[0077] Step 5: Upload the local map posterior estimates of each base station via optical cable to the data fusion center for fusion. Different information fusion strategies are used based on the characteristics of the test scenario. When the field of view of each base station in the environment does not overlap or only partially overlaps, the weighted average method is used to complete the map fusion by weighted averaging the map features of each base station. This method tends to retain the features observed by all base stations and is suitable for scenarios with high missed detection rates. It can be specifically expressed as:

[0078]

[0079] Among them, ρ k It is used to adjust the fusion weight ratio of the kth PA and satisfy ∑ρ k = 1. When the fields of view of each base station in the environment completely overlap, the geometric mean method is used. By converting the exponential geometric mean of the Gaussian mixture into a geometric mean Gaussian mixture, the Gaussian mixture characteristics of the map are preserved. This method tends to retain the common characteristics observed by each base station and is suitable for scenarios with high false alarm rates. It can be specifically expressed as:

[0080]

[0081] Among them, the geometric mean will change the Gaussian mixture characteristics of the fusion result. By ignoring the cross product of different Gaussian components, the fractional exponential power of the Gaussian mixture can be approximated as follows (for simplicity, the time and particle index are omitted):

[0082]

[0083] For the exponential power of Gaussian mixtures, the following results are obtained:

[0084]

[0085] Among them,∈(ρ k ,∑ (k , j) ) is in the form of:

[0086]

[0087] After obtaining the above approximate results, the geometric mean fusion degenerates to N PA Multiply the Gaussian mixtures and use the Gaussian multiplication lemma to get the fusion result.

[0088] For scenarios with high complexity requirements, a sequential fusion approach is used This solution does not require additional fusion steps and completes map estimation sequentially between base stations. It is suitable for fast fusion scenarios with overlapping fields of view. In view of the situation where the user's field of view changes with location, a cloud-based fusion algorithm based on weighted average is adopted. In order to retain the observed global historical wireless map features, the cloud-based fusion center only fuses but does not prune. In order to prevent the features in the cloud from being contaminated by false alarms, it is first necessary to make a convergence judgment on the local map features of each base station. Only when the weight of the Gaussian component is greater than the set threshold and the trace of the variance is less than the set threshold, the map component is considered to be a converged feature and is allowed to be uploaded to the cloud for fusion. The fused map is sent to each base station as a priori for the next moment, thereby completing the estimation of the local map and the global map at the base station and the cloud respectively;

[0089] Step 6: Based on the constraint relationship between the virtual reference anchor point location, user location, and NLoS path channel parameters, the current NLoS path channel parameters are predicted. By correlating the observed and predicted channel parameters at the current moment, including methods such as implicit and explicit correlation, the relative probability of each user particle is calculated. The user locations are then re-scattered using an importance resampling algorithm to achieve a probability distribution with more particles at high-probability locations and fewer particles at low-probability locations, effectively avoiding particle degradation. The re-sampling results are averaged to complete the global posterior estimate of the user location.

[0090] Example

[0091] Figure 2 The performance comparison of different multi-base station information fusion SLAM schemes in completely overlapping FoV scenarios is shown. For the non-fusion and SE fusion schemes, the number of particles is set to 300. In contrast, in the LoS-EKF assisted AA and GCI schemes, the number of particles is significantly reduced to 30. Figure 2 -(a) and Figure 2 -As shown in (b), the SLAM algorithm using AA or GCI fusion can successfully converge even with only 30 particles, confirming the feasibility and effectiveness of using LoS-based user position prediction to improve particle efficiency. In terms of positioning error and map optimal sub-pattern allocation (OSPA), GCI fusion shows superior performance, which is consistent with its applicability to scenarios with overlapping FoVs between multiple base stations, while AA fusion shows slightly lower accuracy than GCI. Consistent with the previous analysis, GCI fusion incurs higher computational cost than AA fusion. However, since the number of particles in both schemes is significantly reduced, their overall complexity is still lower than that of non-fusion methods. Although the SE scheme shows relatively low positioning accuracy, its performance is only slightly inferior to GCI, with a difference of about 0.5 meters.

[0092] For scenarios where the FoVs of multiple base stations partially overlap, assume that all base station antenna arrays face the center of the room and can only receive signals within ±90° of the normal direction. This means that signals from the back of the antenna array cannot be received. Consider two base stations placed at (2,2) and (-2,-2). Figure 3 (a) shows a comparison of global OSPA performance. From the results for a single base station, we can see that global OSPA drops to its lowest value at the 40th time step. This is because the convergence judgment module ensures that features are estimated accurately enough before uploading. In addition, by comparing the global OSPA of a single base station and dual base stations, we can see that increasing the number of uploading base stations can significantly improve convergence speed and estimation accuracy. At the same time, features that are no longer within the FoV are retained. Figure 3 -(b) shows the local map OSPA of base station 1. Compared with the result without fusion, the map OSPA fused by Cloud is more accurate and stable.

Claims

1. A wireless SLAM method based on multi-base station information fusion, characterized in that: The steps include: Step 1: Multiple base stations and users use full-angle beam scanning to obtain the current beam heat map and extract several candidate beam angle pairs from the heat map; Step 2: Multiple base stations filter out multi-base station LoS path information from the candidate beam pairs. Combining the user position estimate from the previous moment’s SLAM algorithm and the pose prior information output by the IMU, the user position at the current moment is predicted. The extended Kalman filter is used to fuse the user position prediction and the LoS path channel parameters to obtain a partial posterior estimate of the user position. Step 3: For each base station, the NLoS path information is filtered from the candidate beam pairs. The user position output by the extended Kalman filter is used as the user position prior of the SLAM algorithm. The particle filter is used to scatter the prior to approximate the user distribution. Step 4: For each base station, the user's possible NLoS path channel parameters are predicted using the SLAM map estimate prior and the user's location points at the previous moment. The SLAM algorithm is then used to associate the NLoS path channel parameters with the map estimate prior, thereby completing a single-base station posterior estimate of the radio map. Step 5: Upload the local map estimation results of multiple base stations to the data fusion center, perform map fusion according to different fusion scenarios, and complete the posterior estimation of the map of multiple base stations; Step 6: Based on the SLAM radio map feature posterior estimation and the multi-base station NLoS path channel parameters, the weight of each user particle is calculated. Based on the weight, the user particles are re-scattered using the importance resampling algorithm to complete the probabilistic scattering of user locations.

2. A wireless SLAM method based on multi-base station information fusion according to claim 1, characterized in that, Step 1 is as follows: multiple base stations and users use uplink full-angle beam scanning. Users use DFT codebooks to form beams in different directions. The base station uses the DFT matrix to receive incident signals from different directions, forming a full-space beam heat map. Using a continuous elimination scheme, the strongest paths are sequentially extracted from the beam heat map and deleted until the energy stop threshold is met. The path with the strongest energy among the extracted paths is used as the LoS path, and the remaining paths are used as NLoS paths. The information output during the full-angle beam scanning phase is: in, represents the estimated channel parameter pair, represents the estimated AoA of the lth path, represents the estimated AoD of the lth path, Indicates the estimated number of multipaths.

3. A wireless SLAM method based on multi-base station information fusion according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Extract the LoS path channel parameters from the channel parameters of each base station-user channel using a LoS path identification method. The LoS path identification method includes a method based on received signal energy, a method based on received signal delay, and a method based on a priori geometric models. Step 2.2: Analyze the geometric constraints satisfied by the user position and the base station position to obtain the relationship between the LoS path channel parameters and the user position. Use the user position output by the SLAM algorithm at the previous moment as prior information, combined with the established assumptions made about the user motion model, which includes the uniform motion model, uniform acceleration motion model, segmented uniform acceleration motion model, and random walk model; and use the pose information output by the IMU to predict the user position at the current moment. Step 2.3: Use Taylor expansion to expand the constraint relationship between the LoS path channel parameters and the user position at the user's predicted position, retain the first-order terms and discard high-order small quantities to obtain a linear approximation constraint relationship. Use a single-target parameter estimation method to obtain a posterior estimate of the user's current position based on the LoS channel parameters. The single-target parameter estimation method includes LS and extended Kalman filter.

4. A wireless SLAM method based on multi-base station information fusion according to claim 1, characterized in that, Step 3 is as follows: for each base station's local SLAM algorithm, the user's position posterior information is used as the particle filter's point scattering prior. The particle filter is based on the distribution of user positions, including Gaussian distribution, uniform distribution, and Poisson distribution; the user position is scattered based on probability, and the accuracy of user position prediction and the efficiency of particle scattering are improved by fusing the LoS path results.

5. A wireless SLAM method based on multi-base station information fusion according to claim 1, characterized in that, Step 4 is as follows: for each base station's local SLAM algorithm, the posterior distribution of the base station's map at the current moment is estimated by scattering the user positions of the multi-base station LoS path and the SLAM map estimation results at the previous moment. This distribution includes two parts: the legacy part at the previous moment and the newly observed part at the current moment. For static environments, the legacy map features at the previous moment remain unchanged. For the newly observed feature part, the feature prediction at the current moment is generated through the empty set method, the single feature method, and the observation-based method. The radio map feature estimate is updated using filters that fuse the NLoS path channel parameters, including the confidence propagation filter and the probability hypothesis strength filter, to complete the local map posterior estimation of a single base station based on the NLoS path parameters.

6. A wireless SLAM method based on multi-base station information fusion according to claim 1, characterized in that, Step 5 specifically includes the following steps: Step 5.1: Upload the local map posterior estimates of each base station to the data fusion center via optical cables for fusion. Different information fusion strategies are adopted according to the different characteristics of the test scenario. Step 5.2: When the fields of view of each base station in the environment do not overlap or only partially overlap, the weighted averaging method is used to complete map fusion by weighted averaging the map features of each base station, retaining the features observed by all base stations. This method is suitable for scenarios with high missed detection rates. Step 5.3: When the fields of view of each base station in the environment completely overlap, the geometric mean method is used. This method converts the exponential geometric mean of the Gaussian mixture into a geometric mean Gaussian mixture. This method preserves the Gaussian mixture characteristics of the map and the common characteristics observed by each base station, making it suitable for scenarios with high false alarm rates. Step 5.4: For scenarios with high complexity requirements, a sequential fusion approach is used. This approach does not require additional fusion steps and sequentially completes map estimation between base stations. This approach is suitable for fast fusion scenarios with overlapping visual fields. Step 5.5: In view of the fact that the user's field of view changes with location, a cloud-based fusion algorithm based on weighted average is used to screen out convergence components whose weights and variances meet the threshold from the local map of each base station and upload them to the cloud. The uploaded converged map is fused with the cloud map. The cloud only fuses the map components without pruning them, and sends the fused map to each base station as the prior at the next moment. In this way, the estimation of the local map and the global map is completed at the base station and cloud respectively.

7. A wireless SLAM method based on multi-base station information fusion according to claim 1, characterized in that, Step 6 specifically involves predicting the current NLoS path channel parameters based on the constraint relationship between the virtual reference anchor point position, user location, and NLoS path channel parameters. By correlating the observed and predicted channel parameters at the current moment, including implicit and explicit correlation methods, the relative probability of each user particle is calculated. The user locations are then re-scattered based on the importance resampling algorithm to achieve a probability distribution with more points at high-probability locations and fewer points at low-probability locations, avoiding particle degradation. The re-sampled results are averaged to complete the global posterior estimate of the user location.

Images