Mobile phone terminal multi-source fusion 3D visual high-precision indoor navigation method

Abstract

According to the mobile phone terminal multi-source fusion 3D visual high-precision indoor navigation method, firstly, an improved robust PDR scheme based on a built-in MEMS sensor of an intelligent terminal is constructed, a dynamic optimal filtering measurement model is constructed by taking INS as core optimization and utilizing multilayer constraint data as observation data, sensor errors are estimated and compensated in real time, and robustness and positioning precision are improved; secondly, establishing a shortest path constraint improved magnetic field matching algorithm, adding the magnetic declination as a magnetic field fingerprint of one dimension into the magnetic field matching algorithm, respectively solving a shortest path constraint result by utilizing the magnetic field fingerprints of multiple dimensions, solving an overall matching result, and reducing a mismatching probability; and finally, by utilizing complementary characteristics of different positioning data sources, an algorithm fusing PDR, magnetic field and Bluetooth multi-source data is established, a quality control mechanism is set, the precision and the robustness of the positioning system are improved, the visual indoor positioning system based on the intelligent terminal is realized, and the integrated navigation positioning precision and the robustness are good.

Description

Technical Field

[0001] This application relates to a mobile visual high-precision indoor navigation method, and particularly to a mobile phone multi-source fusion 3D visual high-precision indoor navigation method, belonging to the technical field of indoor navigation. Background Art

[0002] The public's demand for location-based services (LBS) is becoming increasingly strong, and LBS has been widely used in all aspects of people's daily lives. LBS navigation and positioning services have become an indispensable part of travel. At the same time, based on LBS, it actively expands to other surrounding services. For example, in the commercial field, LBS provides more efficient advertising placement services; in the transportation aspect, LBS reduces traffic pressure through real-time traffic condition reports. In addition, LBS will also play a more active role in emergency services, public safety and other fields; in the construction of smart cities and the era of the Internet of Everything, LBS also plays an important role.

[0003] The penetration rate of smart terminals is already very high. In particular, mobile phones are integrated with rich sensors, such as accelerometers, gyroscopes, magnetometers, barometers, etc., as well as signal receiving modules such as Wi-Fi and Bluetooth, giving mobile phones powerful sensing capabilities. People are already accustomed to using their smartphones to obtain and share location data.

[0004] At present, satellite navigation and positioning technology has developed maturely and has been widely applied. The accuracy of satellite navigation and positioning systems represented by GPS and Beidou basically meets the needs of users for outdoor positioning. However, in urban canyon areas with high-rise buildings and indoor environments where people mostly move, the signal is blocked, the signal strength is severely attenuated, and even the signal cannot be received at all, resulting in neither GPS nor base station positioning technology being able to achieve an ideal positioning effect and unable to meet the needs of indoor high-precision location services. Most people spend nearly 90% of their time indoors, and the need for indoor pedestrian navigation has long emerged. There are many large shopping malls, entertainment venues, airports, railway stations, underground parking lots, etc. Moreover, due to the interference of indoor reinforced concrete and other structures on the environmental magnetic field, the usability of compasses is poor, resulting in people being easily lost in large indoor venues. For example, in large underground parking lots, it is very difficult to find parking spaces and locate cars; in large shopping malls, it is difficult for people to find the items they want. Approximately two-thirds of customers cannot find their favorite products in physical stores in a timely manner. In crowded places such as hospitals and railway stations, it is necessary to reach the destination in a timely manner, saving time costs and improving the user experience. Indoor positioning has shifted from a "false demand" to a rigid demand, and the strong market demand has brought huge business opportunities for indoor positioning technology.

[0005] The problems that need to be solved by the existing mobile visual indoor navigation and the key technical difficulties of this application include:

[0006] (1) The difficulties of existing indoor positioning technologies include: for wireless positioning methods such as TOA and AOA, due to the complex indoor environment, there are serious non-line-of-sight (NLoS) phenomena, severe multipath effect interference, etc., which make outdoor positioning methods difficult to work indoors; the built-in sensors of smart terminals, mobile phones, smart bracelets, etc. have low accuracy; the motion patterns of pedestrians are complex and difficult to model, making it difficult to use motion constraints similar to those in vehicles or ships; in addition, the position of smart terminals relative to pedestrians often changes, making it difficult to estimate the installation angle. The indoor magnetic field is severely interfered, resulting in no reliable heading observation values. The results of existing indoor positioning systems are not satisfactory. If a positioning service similar to outdoor GNSS can be implemented on smart terminals, it will surely bring great convenience to people's lives. Currently, consumer-grade indoor positioning based on smart terminals is still based on WiFi, Bluetooth, magnetic fields, PDR, etc. Due to the influence of the complex indoor environment, different positioning technologies have their own advantages and disadvantages in terms of accuracy, stability, cost, etc. Therefore, it is urgent to use multiple positioning methods for multi-source fusion to achieve complementary advantages, improve the accuracy and robustness of the positioning system, and balance performance and cost.

[0007] (2) The traditional PDR algorithm based on step detection in existing technologies has a simple structure and can utilize very few constraint data. This algorithm has low accuracy, a short available time, and can only provide two-dimensional position data. Existing technologies lack a pedestrian dead reckoning algorithm that can retain the rich data of traditional strapdown inertial navigation without losing the accuracy of the PDR algorithm, cannot extend the available time of the PDR algorithm, cannot provide rich real-time data including three-dimensional position, speed, etc., lack real-time estimation and compensation of sensor errors, and the robustness and positioning accuracy of existing PDR algorithms are poor, and the algorithm generality is weak. Repeated tests are carried out indoors and outdoors using different models of smartphones, and the closed-loop error of the system's positioning results is large, and the positioning accuracy and robustness are poor.

[0008] (3) In existing technologies, the magnetic field matching magnetic field database has a low dimension, the magnetometer calibration is difficult, and the magnetic field matching calculation is large. The probability of magnetic field matching misalignment is high. Regarding the phenomenon of easy misalignment in current indoor positioning based on magnetic field matching, there is a lack of a method to add magnetic declination to the magnetic field database as a matching quantity, and by increasing the dimension, obtain the shortest neighboring path results of multiple dimension matches to reduce the misalignment efficiency. In addition, there is a lack of a method to use the shortest neighboring path results of multiple dimensions, according to the map topology constraint data, and calculate the overall matching result based on the shortest path to further reduce the probability of misalignment. Existing technologies use single modulus matching or modulus + shortest path constraint matching, and the probability of misalignment is very high. Regarding the requirements of current consumer-grade and popular indoor positioning application scenarios, there is a lack of a combined navigation indoor positioning scheme that combines Bluetooth, magnetic fields, and PDR for multi-source fusion.

[0009] (4) In the multi-source fusion integrated navigation algorithm, the prior art fails to fully utilize the complementary characteristics of different positioning data sources. There is a lack of an algorithm for fusing multi-source data of PDR, magnetic field, and Bluetooth, and a lack of a quality control mechanism to improve the accuracy and robustness of the positioning system. The prior art of single magnetic field matching or Bluetooth positioning has significantly poor accuracy and robustness. In a real-time positioning system, the prior art lacks a real-time positioning APP developed based on the Android platform, cannot use the Mapbox map engine to draw indoor 3D maps, and cannot provide a visualization display and analysis evaluation interface for positioning results. Summary of the Invention

[0010] This application improves pedestrian dead reckoning positioning, magnetic field matching positioning, and the multi-source fusion method of PDR, magnetic field, and Bluetooth. First, an improved robust PDR scheme based on MEMS sensors built into smart terminals is constructed. With INS as the core, multi-layer constraint data is optimally utilized, and a dynamic optimal filtering measurement model is constructed using the data as observation data. By real-time estimating and compensating for sensor errors, the robustness and positioning accuracy of the PDR algorithm are improved. Second, a magnetic field matching algorithm improved by shortest path constraints is established. The magnetic declination is added as a dimension of the magnetic field fingerprint to the magnetic field matching algorithm. The shortest path constraint results are obtained separately using magnetic field fingerprints in multiple dimensions, and they are converted into vertices of a graph. Then, through the shortest path constraint, the overall matching result is obtained, reducing the probability of false matching. Finally, in the multi-source fusion integrated navigation algorithm, by utilizing the complementary characteristics of different positioning data sources, an algorithm for fusing multi-source data of PDR, magnetic field, and Bluetooth is established, and a quality control mechanism is set up to improve the accuracy and robustness of the positioning system, realizing a visualization indoor positioning system based on smart terminals. Aiming at the requirements of current consumer-grade and popular indoor positioning application scenarios, the integrated navigation indoor positioning scheme of multi-source fusion of Bluetooth, magnetic field, and PDR has good positioning accuracy and robustness.

[0011] To achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0012] A multi-source fusion 3D visual high-precision indoor navigation method for mobile devices, which improves pedestrian dead reckoning positioning, magnetic field matching positioning, and PDR, magnetic field, and Bluetooth multi-source fusion methods. First, an improved robust PDR scheme based on MEMS sensors built into intelligent terminals is constructed. With INS as the core, multi-layer constraint data is optimized and utilized, including human motion model constraints, quasi-static, quasi-static magnetic fields, turning detection, etc. As observation data, a dynamic optimal filtering measurement model is constructed. By estimating and compensating sensor errors in real time, the robustness and positioning accuracy of the PDR algorithm are improved. Second, a magnetic field matching algorithm improved by shortest path constraints is established. The magnetic declination is added as a dimension of the magnetic field fingerprint to the magnetic field matching algorithm. The shortest path constraint results are obtained separately using magnetic field fingerprints in multiple dimensions, including modulus, horizontal component, vertical component, and magnetic declination. They are converted into vertices of a graph, and then the overall matching result is obtained through shortest path constraints, reducing the probability of false matching. Finally, in the multi-source fusion integrated navigation algorithm, using the complementary characteristics of different positioning data sources, an algorithm that fuses PDR, magnetic field, and Bluetooth multi-source data is established, and a quality control mechanism is set up to improve the accuracy and robustness of the positioning system, realizing a visualization indoor positioning system based on intelligent terminals;

[0013] 1) Improved pedestrian indoor dead reckoning positioning with multi-layer constraints: An observation equation constructed with multi-layer constraint data is established. By using multi-layer constraint data as the dynamic optimal filtering observation signal, the available time of the existing PDR is increased through effective error feedback;

[0014] 2) Improved magnetic field feature matching algorithm: Through data collection and database construction, the magnetic declination is used as a dimension in the magnetic field fingerprint database, increasing the dimensional data of the magnetic field fingerprint. The overall matching result is obtained through shortest path constraints;

[0015] 3) Indoor positioning system based on dead reckoning / Bluetooth / magnetic field: A strategy for multi-source data fusion indoor positioning is constructed, fusing multiple data sources to make up for each other's deficiencies. Based on Android terminals, a real-time visualization indoor positioning APP is developed using mapbox.

[0016] Preferably, for improved pedestrian indoor dead reckoning positioning with multi-layer constraints: An improved pedestrian dead reckoning algorithm guided by INS prior is constructed. Using multi-layer constraint data as the observation equation of dynamic optimal filtering, the positioning accuracy of the system is improved, thus maximizing the available time of the PDR algorithm;

[0017] The multi-layer constraint data used includes ZUPT, ZARU, gravity vector constraint, NHC, and magnetic field vector constraint. The specific steps to improve the PDR algorithm are as follows: First, perform sensor error compensation on the initial data of the accelerometer, gyroscope, and magnetometer, and then conduct strapdown inertial navigation mechanical arrangement. At the same time, detect the state of the user, including foot points, zero speed, low dynamics, and quasi-static magnetic field. If the constraint conditions are met, trigger the corresponding measurement update, and at the same time, perform error feedback to correct the current navigation state data. Finally, return to the step of inputting the initial sensor signal.

[0018] Preferably, the observation equation constructed by multi-layer constraint data: The multi-layer constraint data used includes: ZUPT, ZARU, NHC, gravity vector, magnetic field vector;

[0019] 1 - Quasi-static constraint: Based on the ZUPT detection method, combine the critical value and the hypothesis test item, and use the weighted method of the gyroscope and accelerometer to judge whether the pedestrian is in the quasi-static state. When it is detected that the pedestrian meets the quasi-static condition, it is determined that the pedestrian speed is zero, and the pseudo velocity observation value v n =[0 0 0] T is calculated based on the hypothesis condition, and the corresponding error observation equation is:

[0020]

[0021] where is the speed at the current moment calculated by inertial navigation in the n system;

[0022] At the same time, when it is determined that the pedestrian is in the quasi-static state, the heading remains unchanged, the output of the heading gyroscope is 0, and all heading changes are caused by the error of the heading gyroscope. Calculate the pseudo observation value ω z =0, and the corresponding error observation equation is:

[0023]

[0024] where is the output of the heading gyroscope, and the design matrix under the quasi-static condition is:

[0025]

[0026] 2 - Gravity vector constraint: Under the condition of determining that there is no external acceleration, using the gravity vector obtained by acceleration observation to update the navigation state can ensure the horizontal angle accuracy. Whether to use the acceleration observation value is measured according to the magnitude of A in Equation 4:

[0027] A = |norm(f b ) - g| Equation 4

[0028] where g is the local gravitational acceleration. Under the condition of small external input acceleration, A ≤ |THacc If so, the standard deviation of the accelerometer measurement noise is σ a , representing the noise level of the accelerometer; under the condition of a large external acceleration input, A > |TH acc |, then the accelerometer is not used to assist the horizontal attitude angle, and TH acc is the critical value;

[0029] Under static and low-dynamic conditions, based on the accelerometer measurement data to assist the attitude angle, a tightly coupled method is used to construct the acceleration observation equation to ensure that the carrier does not have the ±90deg Euler angle singularity problem under various motion states. The acceleration error equation is:

[0030]

[0031] where f n =-g n =[0 0 g] T , ψ is the attitude error, n is the measurement noise, and the design matrix under low-dynamic conditions is:

[0032] H = |0 3×3 0 3×3 g n ×0 3×3 0 3×3 | Equation 6

[0033] 3 - Pedestrian motion constraint: When the vehicle body does not skid or jump, it is assumed that the vehicle body only has a speed in the forward or backward direction. If the pedestrian does not show an abnormal walking state, they will all move forward along the corridor. At this time, there is only a forward speed, satisfying the NHC condition. Under turning conditions, there is a skidding phenomenon similar to that in vehicle navigation, and it is necessary to turn off the NHC in time to ensure the stability of the system. The turning detection is used as the basis for whether to use the NHC. The heading gyro observation value is combined with the sliding window critical value to judge whether the pedestrian is in a turning state:

[0034]

[0035] is the observation value of the current gyro, min represents the minimum value of a time window, and TH gyro is the critical value. When it is judged that the pedestrian is in a straight walking state, the pseudo-velocity observation value v b =[v vx 0 0] T , and the error equation is constructed according to the NHC hypothesis:

[0036]

[0037] where v nThe current moment velocity calculated by inertial navigation in the n coordinate system is the direction cosine matrix for the rotation from the n coordinate system to the b coordinate system, and v vx represents the forward velocity. stepL and Δt are the estimated step length and the footstep time interval respectively:

[0038] v vx = stepL / Δt Equation 9

[0039] where a zmax and a zmin represent the maximum and minimum values of the acceleration modulus within the sliding time window. K represents a coefficient. The design matrix under non-integrity constraints is:

[0040]

[0041] 4 - Quasi-Static Magnetic Field (QSMF) Constraint: In a local area, especially in an empty indoor area, the environmental magnetic field remains stable. To fully utilize the magnetometer correction data without reducing the robustness of the system, by detecting the surrounding environmental magnetic field, the magnetometer vector constraint is used only when the environmental magnetic field is QSMF. The judgment of quasi-static conditions is as follows: Project the current epoch magnetometer observation value onto the n coordinate system. By comparing the difference between the current epoch and the environmental magnetic field in the buffer, determine whether the current epoch is a quasi-static magnetic field. Specifically:

[0042]

[0043] where N is the number of buffered epochs, and TH mag is the quasi-static magnetic field critical value;

[0044] When it is determined that the current epoch is in a quasi-static magnetic field, use the magnetic field vector to assist the current attitude angle, and construct the magnetometer observation equation in a tightly coupled manner. The magnetic field vector error equation is:

[0045]

[0046] where m n is the calibrated environmental magnetic field, is the magnetometer observation value. The design matrix under quasi-static magnetic field conditions is:

[0047]

[0048] Based on UPT, ZARU, NHC, the gravity vector, and the magnetic field vector, construct the observation equation through multi-layer constrained data.

[0049] Preferably, for data acquisition and database construction with improved magnetic field feature matching: Based on the visualization of indoor magnetic field fingerprint acquisition by IndoorAtlas, for any area included in the indoor map, plan the trajectory by dotting. The collector walks along the predetermined trajectory, keeping the intelligent terminal consistent with the forward direction of the pedestrian. When the user reaches the specified road point, obtain the road point data by clicking the button. During the preprocessing of the database, interpolation is performed according to the step detection and road point data. Project the collected magnetic field and other sensor data onto the corresponding positions on the map, project the collected magnetometer observation values onto the horizontal and vertical planes, and calculate the magnetic heading by projecting onto the horizontal plane at the same time;

[0050] During the processing of the database construction data, first perform indoor pedestrian dead reckoning positioning with multi-layer constraints improvement to obtain more accurate horizontal angle and heading angle. Project the collected magnetic field signals onto the horizontal and vertical planes by projection. In addition, calculate the magnetic heading angle during the walking process, and calculate the difference with the heading angle obtained by PDR to calculate the magnetic declination data corresponding to a certain point, which is used as a dimension data in the magnetic field database. When using the magnetic declination for matching, subtract the mean value from the two waveform sequences to remove the influence of the unknown initial heading. Establish the database using the traveling distance, and at the same time consider that the fingerprints may walk in the opposite direction. Flip the segments of fingerprints to obtain fingerprints in the opposite walking direction.

[0051] Preferably, the magnetic field matching algorithm process: Determine the similarity between the observed data and the fingerprints in the database. Use the shortest neighbor path to obtain N optimal matching results for each dimension of the magnetic field signal respectively, convert them into the nodes of the graph, and use the improved Dijkstra to obtain the matching result of the entire traveling process by using the shortest path constraint.

[0052] Preferably, the magnetic field matching compares the similarity of two waveforms. Assume the observed value sequence Seq observe =(B o1 …B om ), where m is the sequence length, and a certain section of the database fingerprint Seq fingerprint =(B f1 …B fn ), n is the sequence length. Find the similar parts in the two sequences, compare the similarity of the two sequences by moving them to the same horizontal plane, and through the function g(o i , f iIt is determined whether two epochs are similar by checking if it is less than the critical value. If two epochs are similar, the scoring factor is assigned as 3, otherwise the penalty factor is set to -3. The scoring matrix is calculated, and parameters are set to determine whether two sequences are similar. First, to avoid the maximum value of the scores in the scoring matrix being too small (the matching part is too short), the lowest score critical value Th1 is set. Second, the parameter Th2 is set. When the number of deleted epochs is greater than this critical value, the two sequences are considered unmatched. Similarly, the parameter Th3 is set to avoid inserting too many epochs.

[0053] Preferably, for the analysis of the multi-dimensional shortest neighbor path results, the self-conformity is calculated based on the comparison of the matching results of multiple dimensions in the magnetic field matching. When the maximum matches of the modulus value, vertical component, horizontal component, and magnetic declination all correspond to a certain segment in the database, it is determined that the matching result has a high credibility, and the number of nodes corresponding to this observation data segment is merged to reduce the computational amount of the shortest path.

[0054] Preferably, for the shortest path constraint, each segment of magnetic field observation values may be similar to multiple fingerprints s match ={B fi …B fj}. The algorithm is designed to connect the segments of observation data to form a complete trajectory, converting this problem into a shortest path problem. All the fingerprint data segments determined to be matched are converted into the vertices of graph G. For any two vertices in the graph, using their corresponding coordinates in the fingerprint database and based on the topological relationship in the map, the distance between the two vertices is calculated according to Dijkstra's algorithm. The map constraint data is adopted. If the distance between two vertices is greater than the critical value Thdis1, where Thdis1 = 10m, the corresponding infinite value INT_MAX is set in the adjacency matrix. Then the starting point and ending point of the shortest path are set. Two new points START and END are added. Among them, START is connected to any fingerprint that matches the first segment of observation data, and END is connected to any fingerprint that matches the last segment of observation data. The lengths of the newly added edges are all set to 0. The overall matching result is obtained by finding the shortest path from START to END;

[0055] Using the results of the shortest adjacent path matching in different dimensions of the magnetic field as the vertices of graph G, when there is no correct matching in terms of the magnetic field modulus, vertical component, horizontal component, and magnetic declination, to ensure the spatial continuity of the overall matching result obtained from the shortest path, parameters are set to eliminate the spatial discontinuity phenomenon. If a certain distance in the obtained shortest path is greater than the critical value Thdis2 (let Thdis2 = 10m), it is determined that the matching result of this section is spatially discontinuous; a new end point END1 is added before this section of the matching, and a new start point START1 is added after this section of the matching. The problem of the shortest path from START to END is converted into the sum of the two shortest distances from START to END1 and from START1 to END. If there is also spatial discontinuity from START1 to END, the above steps are repeated.

[0056] Preferably, Bluetooth fingerprint recognition: A Bluetooth indoor positioning system is constructed based on the fingerprint method, including two stages: database establishment and matching positioning. In the database establishment stage, a Bluetooth fingerprint database is generated by means of walking surveying and mapping. By interpolating the map coordinates corresponding to the dead reckoning of pedestrians, the map coordinates at the moment of Bluetooth signal reception are obtained. Then, within the range of fingerprint collection, a 2m square grid is established, and the Bluetooth signals during walking are mapped to the nearest grid as the fingerprint signal of this grid point. The initially collected RSS signals are processed to improve the quality of the database. First, the RSS signals below the critical value are eliminated; second, when there are multiple initial signals mapped to a certain grid point, the mean value is taken to reduce the noise in the database. Finally, the established Bluetooth fingerprint signals are as follows:

[0057] DB i ={Pos i , (mac1, RSS1), …(mac n , RSS n )} Equation 14

[0058] where Pos i is the coordinate of the fingerprint point, mac is the mac address of the AP, and RSS is the signal strength of the corresponding AP. The finally established Bluetooth fingerprint;

[0059] After collecting the initial Bluetooth signal, preprocess the data. First, delete the Bluetooth RSS data below the critical value; use the average of the latest 3 Bluetooth observed RSS values as the RSS value of the current epoch for Bluetooth observation. For the preprocessed Bluetooth signal, if the number of APs is sufficient, compare the observed data with the fingerprint points in the database one by one, and select K fingerprint points with the closest distances. Then, process the selected K RPs. The K fingerprint points are close to each other, and the RPs with a distance greater than a certain critical value from the mean of all RPs are removed. Use the mean of the RP coordinates that meet the conditions as the positioning result of fingerprint matching. Finally, check the positioning result. If the matching result is similar to the historical positioning result, it is considered that the positioning result meets the requirements; otherwise, the current matching result is treated as a mis-match.

[0060] Preferably, dead reckoning / Bluetooth / magnetic field integrated navigation: Use dynamic optimal filtering to implement Bluetooth / magnetic field / PDR integrated navigation, and adopt a loose combination method for the combination of Bluetooth and PDR. When the pedestrian walks a certain distance, obtain the current user position observation value through magnetic field matching, and use the magnetic field matching result as the observation data of dynamic optimal filtering to achieve the integrated navigation of magnetic field and PDR. PDR provides continuous navigation results, and the results of Bluetooth matching and magnetic field matching jump back and forth. By comparing the PDR result with the results of Bluetooth and magnetic field matching, when their distance is greater than a certain critical value, it is considered that the result of fingerprint matching has low credibility, and the quality control of the observation data is realized.

[0061] The improved PDR algorithm takes INS as the core, ensures the performance of the positioning system through multi-layer constraint data, uses PDR as the system framework to provide continuous positioning results, and Bluetooth and magnetic field are used as position observation data to provide system position correction. Under the condition of sufficient BLE deployment, it provides a high-precision positioning result. The positioning result of magnetic field matching is used as position reference data, but the magnetic field positioning accuracy is related to the specific position. Bluetooth and magnetic field are used as supplementary means to improve the accuracy of the integrated navigation system by correcting the positioning result.

[0062] Use the position obtained by magnetic field matching or Bluetooth fingerprint recognition to construct the dynamic optimal filtering observation equation:

[0063] p PDR -p obs = δp + n Equation 15

[0064] Among them, P PDR is the position data predicted by PDR, P obs is the result of Bluetooth or magnetic field matching, n is the noise of the observed signal, and δp is the position error; the design matrix of dynamic optimal filtering:

[0065] H = |I 3×3 O 3×3 O3×3 O 3×3 O 3×3 | Formula 16

[0066] Q matrix for dynamic optimal filtering:

[0067]

[0068] wherein, VRW represents velocity random walk, ARW is angular random walk, σ bg is the gyro bias noise variance, and σ ba is the accelerometer bias noise variance;

[0069] Quality control includes: judging the credibility of Bluetooth and magnetic field matching results by using the continuity of PDR positioning results; using Bluetooth positioning results or current integrated navigation results to reduce the search space of magnetic field matching.

[0070] Compared with the prior art, the innovation points and advantages of the present application are as follows:

[0071] (1) The present application improves pedestrian dead reckoning positioning, magnetic field matching positioning, and the multi-source fusion method of PDR, magnetic field and Bluetooth. First, an improved robust PDR scheme based on MEMS sensors built in smart terminals is constructed. With INS as the core, multi-layer constraint data is optimally utilized, and an observation data-based dynamic optimal filtering measurement model is constructed. By real-time estimating and compensating sensor errors, the robustness and positioning accuracy of the PDR algorithm are improved. Second, a magnetic field matching algorithm improved by shortest path constraints is established. The magnetic declination is added to the magnetic field matching algorithm as a dimension of magnetic field fingerprint. The shortest path constraint results are obtained respectively by using magnetic field fingerprints in multiple dimensions, and they are converted into vertices of a graph. Then, through shortest path constraints, the overall matching result is obtained, and the probability of mis-matching is reduced. Finally, in the integrated navigation algorithm of multi-source fusion, by using the complementary characteristics of different positioning data sources, an algorithm for fusing multi-source data of PDR, magnetic field and Bluetooth is established, and a quality control mechanism is set up to improve the accuracy and robustness of the positioning system, and a visualization indoor positioning system based on smart terminals is realized; aiming at the requirements of current consumer-grade and popular indoor positioning application scenarios, the integrated navigation indoor positioning scheme of multi-source fusion of Bluetooth, magnetic field and PDR has good positioning accuracy and robustness.

[0072] (2) This application designs a combined navigation algorithm that fuses Bluetooth, magnetic field, and PDR on a smartphone terminal, and develops a real-time visualization indoor positioning APP based on the Android system to improve the accuracy and robustness of existing PDR and magnetic field matching algorithms. In the improved PDR algorithm, this application uses multi-layer constrained data as the dynamic optimal filtering observation signal and improves the available time of the existing PDR through effective error feedback. According to the experimental results, the improved PDR algorithm platform of this application has strong universality and is applicable to smart terminals of different brands. At the same time, it has good performance in indoor and outdoor tests, indicating that the algorithm of this application can effectively improve the performance of pedestrian dead reckoning based on smart terminals, the accuracy and robustness of the positioning system, and achieve a perfect balance between performance and cost.

[0073] (3) In the magnetic field matching algorithm, aiming at the phenomenon of easy mis-matching in magnetic field matching, this application takes the magnetic declination as a dimension in the magnetic field fingerprint database to increase the dimensional data of the magnetic field fingerprint, which is easy to distinguish between fingerprints. At the same time, for different dimensions of magnetic field signals, this application proposes to use their KNN results to convert them into vertices of a graph, and obtain the overall matching result through the shortest path constraint. The experimental results show that the mis-matching rate of the magnetic field matching method proposed in this application is only 0.8%, which greatly reduces the mis-matching rate of magnetic field matching.

[0074] (4) In multi-source fusion, this application uses the complementary characteristics of various data and improves the accuracy and robustness of the positioning system through a quality control mechanism. The experimental results show that the positioning accuracy and robustness of multi-source fusion are significantly better than those of single magnetic field matching or Bluetooth fingerprint positioning. According to the algorithm design of this application, a real-time positioning system based on Android terminals is developed, and using the mapbox map engine, a detailed map production process is given to facilitate the visualization of positioning results, providing a visualization display and analysis evaluation interface for positioning results. The indoor positioning speed is fast and the accuracy is high. Description of the Drawings

[0075] Figure 1 It is a flowchart of dynamic optimal filtering of multi-layer constrained data.

[0076] Figure 2 It is an experimental diagram of indoor pedestrian dead reckoning positioning library improved by multi-layer constraints.

[0077] Figure 3 It is an experimental diagram of indoor pedestrian dead reckoning positioning playground improved by multi-layer constraints.

[0078] Figure 4 It is an example diagram of the map data collection App.

[0079] Figure 5 It is a schematic diagram of Bluetooth fingerprint recognition experiment.

[0080] Figure 6 It is a combined navigation algorithm diagram for effectively integrating multi-source data.

[0081] Figure 7 It is a schematic diagram of the positioning trajectories of different strategies in Experiment 1.

[0082] Figure 8 It is a schematic diagram of the positioning trajectories of different strategies in Experiment 2.

[0083] Figure 9 It is a schematic diagram of the positioning trajectories of different strategies in Experiment 3.

[0084] Figure 10 It is the cumulative error distribution diagram of multiple experiments of this application.

[0085] Figure 11 It is the overall effect diagram of the multi-source fusion indoor positioning APP based on Android. Specific implementation manners

[0086] Next, in combination with the accompanying drawings, the technical solutions of the multi-source fusion 3D visual high-precision indoor navigation method provided by this application are further described, so that those skilled in the art can better understand this application and be able to implement it.

[0087] With the continuous development of the Internet and the Internet of Things, the demand for positioning in location-based services (LBS) has gradually expanded from outdoor to indoor. The global satellite navigation positioning system (GNSS) is widely used outdoors, but it cannot be used indoors due to factors such as severe signal occlusion, non-line-of-sight propagation (NOLS), and multipath. Currently, indoor positioning technologies include Wi-Fi, Bluetooth, magnetic field, ultra-wideband (UWB), and pedestrian dead reckoning (PDR) and other technologies. In recent years, due to the high popularity, rich integrated sensors, and convenient carrying of smartphones, they have provided an experimental platform for the implementation of indoor positioning technologies based on intelligent terminals, making low-cost indoor positioning technologies based on smartphones a research hotspot in recent years. Aiming at the requirements of current consumer-level and popular indoor positioning application scenarios, this application designs a combined navigation indoor positioning scheme that fuses Bluetooth, magnetic field, and PDR.

[0088] This application improves pedestrian dead reckoning positioning, magnetic field matching positioning, and multi-source fusion methods of PDR, magnetic field, and Bluetooth to implement a visualization indoor positioning system based on intelligent terminals.

[0089] First, this application designs a robust PDR solution based on the MEMS (Micro-Electromechanical Systems) sensors built into smart terminals. This method takes the strap-down inertial navigation system (INS) as the core and uses multi-layer constrained data, including human motion model constraints, quasi-static, quasi-static magnetic fields, turning detection, etc., as observation data to construct a dynamic optimal filtering measurement model. By estimating and compensating sensor errors in real time, the robustness and positioning accuracy of existing PDR algorithms are significantly improved. The PDR algorithm designed in this application has strong versatility and performs well both indoors and outdoors. In actual tests, with the handheld device in a flat state, multiple repeated tests were carried out indoors and outdoors using different models of smartphones, and the closure error of the positioning results of the system was better than 1.6% of the travel distance.

[0090] Secondly, this application designs a magnetic field matching algorithm improved based on the K-Nearest Neighbour (KNN) algorithm and the shortest path constraint. Aiming at the problems of low distinctiveness of magnetic field fingerprints and easy occurrence of mis-matching in current magnetic field matching positioning, this application adds the magnetic declination as a dimension of the magnetic field fingerprint to the magnetic field matching algorithm, and uses magnetic field fingerprints in multiple dimensions (including modulus, horizontal component, vertical component, and magnetic declination) to obtain KNN results respectively, converts them into vertices of a graph, and then obtains the overall matching result through the shortest path constraint, reducing the probability of mis-matching. In actual tests, the mis-matching rate of the improved magnetic field matching algorithm proposed in this application has been greatly reduced.

[0091] Finally, in the multi-source fusion integrated navigation algorithm, this application makes full use of the complementary characteristics of different positioning data sources and designs an algorithm that fuses multi-source data of PDR, magnetic field, and Bluetooth. Through a quality control mechanism, the accuracy and robustness of the positioning system are improved. Experiments show that the integrated navigation algorithm that fuses multi-source data has significantly improved accuracy and robustness compared with single magnetic field matching or Bluetooth positioning. In the real-time positioning system, this application develops a real-time positioning APP based on the Android platform and uses the Mapbox map engine to draw a 3D indoor map, providing a visual display and analysis evaluation interface for the positioning results.

[0092] I. Indoor Pedestrian Trajectory Estimation and Positioning Improved by Multi-Layer Constraints

[0093] This application constructs a system observation equation through the complementary relationship among multi-sensors including a speedometer, a gyroscope, and a magnetometer, as well as human kinematic model data. By using an extended dynamic optimal filter to estimate the accelerometer bias and gyro bias online, the divergence rate of the heading error over time is suppressed, and the available time of the pedestrian dead reckoning algorithm for intelligent terminals is increased. The method includes the following steps: By preprocessing the initial sensor signals, it is judged whether the current is in quasi-static, quasi-static magnetic field, and turning constraint conditions to ensure the accuracy of the observation data, so as to perform correct error feedback, and the improved pedestrian dead reckoning algorithm proposed in this application is analyzed and verified through the actual positioning results.

[0094] (1) Overall algorithm design

[0095] The traditional PDR algorithm based on foot detection has a simple structure and can utilize very few constraint data. Therefore, this algorithm has low accuracy, short available time, and can only give two-dimensional position data. On the premise of retaining the rich data of traditional strapdown inertial navigation and not losing the accuracy of the existing PDR algorithm, this application constructs an improved pedestrian dead reckoning algorithm guided by INS prior, using multi-layer constraint data as the observation equation of the dynamic optimal filter to improve the positioning accuracy of the system, thereby maximizing the available time of the PDR algorithm. Compared with the existing PDR algorithms, it can provide richer real-time data, including three-dimensional position, speed, etc. The implementation of this algorithm is as Figure 1 shown.

[0096] The multi-layer constraint data used in this application includes ZUPT (Zero Velocity Update), ZARU (Zero Angular Rate Update), gravity vector constraint, NHC (Non-Holonomic Constraints), and magnetic field vector constraint. The specific steps to improve the PDR algorithm include: First, perform sensor error compensation on the initial data of the accelerometer, gyroscope, and magnetometer, then perform strapdown inertial navigation mechanical arrangement. At the same time, detect the states of the user including foot points, zero speed, low dynamics, and quasi-static magnetic field. If the constraint conditions are met, trigger the corresponding measurement update, and at the same time perform error feedback to correct the current navigation state data. Finally, return to the step of inputting the initial sensor signal.

[0097] (2) Observation equation constructed by multi-layer constraint data

[0098] The multi-layer constraint data adopted includes: ZUPT, ZARU, NHC, gravity vector, and magnetic field vector;

[0099] 1 - Quasi - static constraint: Based on the ZUPT detection method, a method combining a critical value and a hypothesis test item, and weighting the gyroscope and accelerometer is used to determine whether a pedestrian is in a quasi - static state. When it is detected that the pedestrian meets the quasi - static conditions, the pedestrian speed is determined to be zero, and the pseudo - speed observation value v is calculated based on the hypothesis conditions n =[0 0 0] T , corresponding to the error observation equation:

[0100]

[0101] where is the speed at the current moment deduced by the inertial navigation in the n - system;

[0102] Meanwhile, when it is determined that the pedestrian is in a quasi - static state, the heading remains unchanged, the output of the heading gyroscope is 0, and all heading changes are caused by the heading gyroscope error. The pseudo - observation value of the heading angular rate ω is calculated z =0, corresponding to the error observation equation:

[0103]

[0104] where is the output of the heading gyroscope, and the design matrix under quasi - static conditions is:

[0105]

[0106] 2 - Gravity vector constraint: Under the condition of determining that there is no external acceleration, using the gravity vector obtained from the acceleration observation to update the navigation state can ensure the horizontal angle accuracy. Whether to use the acceleration observation value is measured according to the magnitude of A in Equation 4:

[0107] A = |norm(f b ) - g| Equation 4

[0108] where g is the local gravitational acceleration. Under the condition of a small external input acceleration, if A ≤ |TH acc |, then the standard deviation of the accelerometer measurement noise is σ a , representing the accelerometer noise level; under the condition of a large external acceleration input, if A > |TH acc |, then the accelerometer is not used to assist the horizontal attitude angle, and TH acc is the critical value;

[0109] Under static and low - dynamic conditions, based on the accelerometer measurement data to assist the attitude angle, a tightly - coupled method is used to construct the acceleration observation equation to ensure that the carrier does not have the ±90deg Euler angle singularity problem under various motion states. The acceleration error equation:

[0110]

[0111] where f n = -g n = [0 0 g] T , ψ is the attitude error, n is the measurement noise, and the design matrix under low-dynamic conditions is:

[0112] H = |0 3×3 0 3×3 g n ×0 3×3 0 3×3 | Equation 6

[0113] 3 - Pedestrian motion constraint: When the vehicle body does not skid or jump, it is considered that the vehicle body only has a speed in the forward or backward direction. If the pedestrian does not show abnormal walking states, they will all move forward along the corridor. At this time, there is only a forward speed, satisfying the NHC condition. Under turning conditions, there is a skidding phenomenon similar to that in vehicle navigation, and it is necessary to turn off NHC in a timely manner to ensure the stability of the system. The turning detection is used as the basis for whether to use NHC, and the heading gyro observation value is combined with the sliding window critical value to judge whether the pedestrian is in a turning state:

[0114]

[0115] is the observation value of the current gyro, min represents the minimum value of a time window, and TH gyro is the critical value. When it is judged that the pedestrian is in a straight walking state, the pseudo-velocity observation value v b = |v vx 0 0] T , and the error equation is constructed according to the NHC hypothesis:

[0116]

[0117] where v n is the speed at the current moment deduced by inertial navigation in the n system, is the direction cosine matrix for rotating the n system to the b system, and v vx represents the forward speed, and stepL and △t are the estimated step length and the footstep time interval respectively:

[0118] v vx = stepL / Δt Equation 9

[0119] where, a zmax , a zmin represent the maximum and minimum values of the acceleration modulus within the sliding time window, K represents the coefficient, and the design matrix under nonholonomic constraints is:

[0120]

[0121] 4 - Quasi - Static Magnetic Field (QSMF) Constraint: In a local area, especially in an open indoor area, the environmental magnetic field remains stable. While fully utilizing the magnetometer correction data without reducing the system's robustness, by detecting the surrounding environmental magnetic field, the magnetometer vector constraint is only used when the environmental magnetic field is QSMF. The quasi - static condition is judged as follows: Project the magnetometer observation value of the current epoch onto the n - system. By comparing the difference between the environmental magnetic field of the current epoch and that in the buffer, it is determined whether the current epoch is a quasi - static magnetic field. Specifically as follows:

[0122]

[0123] where N is the number of cached epochs, and TH mag is the critical value of the quasi - static magnetic field;

[0124] When it is determined that the current epoch is in a quasi - static magnetic field, the magnetic field vector is used to assist the current attitude angle, and the magnetometer observation equation is constructed using a tightly - coupled method. The magnetic field vector error equation is:

[0125]

[0126] where m n is the calibrated environmental magnetic field, is the magnetometer observation value, and the design matrix under the quasi - static magnetic field condition is:

[0127]

[0128] Based on UPT, ZARU, NHC, the gravity vector, and the magnetic field vector, the observation equation is constructed by multi - layer constrained data.

[0129] (III) Experimental Results and Analysis

[0130] To verify that the algorithm designed in this application is applicable to multiple terminals and various indoor and outdoor scenarios, we used Huawei Mate60 and Xiaomi 14 smartphones in total. Only the built - in sensors of the mobile phones (such as three - axis accelerometers, gyroscopes, magnetometers) were used to collect initial data, and the sampling rate was 50Hz for all. Typical indoor and outdoor scenarios were covered: The library of the Data Department of a certain university and the playground of the Data Department of a certain university were used as experimental sites. Pedestrians held the smartphones and started from a fixed point along a fixed direction, walked along the planned trajectory of the site, and finally returned to the starting point to form a closed loop. A total of 6 outdoor trajectories were collected, 2 for each mobile phone, and the total length of a single trajectory was 390 meters; 6 indoor trajectories were collected, 2 for each mobile phone, and the total length of a single trajectory was 205 meters. The robust PDR algorithm designed in this application was used for processing, and the walking trajectories of the pedestrians were calculated (as shown in Figure 2 and Figure 3 ).

[0131] From Figure 2 andFigure 3 It can be seen that the multiple tests of different brands on the same trajectory have good repeatability, indicating that the algorithm designed in this application is not sensitive to the sensor differences of different terminals and has good generality. The closure errors of the two navigation and positioning results of the three brands of smart phones are all small, with the maximum being 1.6%, which is better than the navigation accuracy of the closure error of 3% to 5% of the ordinary PDR algorithm described in the prior art.

[0132] The method of this application can effectively suppress the speed of the heading divergence over time. This application compares the PDR solution results of three processing schemes: Scheme 1, without using any constraint data; Scheme 2, using the accelerometer and magnetometer as constraint data; Scheme 3, using the method designed in this application. The method of this application, due to using multi-layer constraint data to online estimate the gyro zero bias and heading error, compared with Scheme 1 and Scheme 2, the speed of the heading divergence over time is slow, increasing the available time of the existing PDR algorithm.

[0133] In addition, the algorithm of this application has the ability to online estimate the accelerometer zero bias in real time. Compared with the traditional PDR algorithm that relies on the accelerometer leveling, it has more robust performance in terms of data source. At the same time, this algorithm can provide richer navigation data (such as three-dimensional velocity data) and can utilize more constraint data (such as motion model constraints).

[0134] The traditional PDR algorithm updates the position based on the pedestrian's foot points, with a low update frequency (1 to 2 Hz). When fusing other positioning sources (such as absolute positioning means like Wi-fi and Bluetooth), time alignment interpolation processing is required, and the error propagation lacks a rigorous and reliable theoretical basis. The algorithm of this application outputs navigation data at a high frequency of 50 Hz. Using the nearest point data to fuse other positioning sources can solve the problem of time misalignment in multi-source data fusion; at the same time, this algorithm uses the INS core, which is the same as the common strapdown INS. In the multi-source fusion process, the traditional INS integrated navigation method can be directly applied, greatly reducing the algorithm complexity and workload of data fusion.

[0135] The improved PDR algorithm proposed in this application uses multi-layer constraint data to ensure the navigation performance of the system. The test results show that this algorithm is suitable for smart terminals of different brands and has good platform generality. In the state of holding the mobile phone flat, the closure errors of the repeated test results are all less than 1.6% of the travel distance. The method of this application can improve the accuracy and available time of the positioning system based on MEMS sensors. At the same time, the high-frequency navigation data output by this application is easy to perform multi-source fusion with other positioning data sources.

[0136] II. Improved Magnetic Field Feature Matching Algorithm

[0137] Due to the presence of a large amount of reinforced concrete, cables, and large motors in the indoor environment, the environmental magnetic field is disordered, complex, and unpredictable, and it is impossible to provide effective heading observation values through a magnetometer. However, by utilizing the spatial difference of the indoor magnetic field, positioning can be achieved through fingerprint matching. Compared with other indoor positioning technologies, magnetic field matching has the following advantages: First, the magnetic field does not require any layout, and the magnetic field fingerprint can be used in indoor scenarios; Second, compared with other fingerprint signals (such as Wi-Fi), the magnetic field is more stable. The disorder of the indoor magnetic field mainly comes from the reinforced concrete structure indoors, and the activities of small and medium-sized electrical appliances and people indoors have little impact on the magnetic field, and the magnetic field characteristics change slowly over time; Third, the magnetic field is sensitive to position changes and it is easy to achieve high-precision positioning; Fourth, in terms of the power consumption of intelligent terminals, the energy consumption of magnetic field sensors is much smaller than that of frequent Bluetooth and Wi-Fi scans.

[0138] The quality of the magnetic field matching result depends to a large extent on the environmental magnetic field. The greater the spatial difference of the environmental magnetic field, the better the matching effect. On the contrary, in areas where the environmental magnetic field fluctuates less, the matching result is often poor and mis-matching is likely to occur.

[0139] Deficiencies of existing magnetic field matching technologies:

[0140] 1. Low dimension of the magnetic field database. The output of the intelligent terminal magnetometer is a three-dimensional vector B = <Bx, By, Bz> with the magnetometer itself as the reference. At the same location, there are significant differences in the magnetometer observation values in different directions. The simplest method is to collect database fingerprints with the mobile phone facing different directions at a certain location, but this greatly increases the workload of collection and the difficulty of matching, and it is basically infeasible in practical applications. The currently commonly used method is to project the magnetic field into a specified reference coordinate system through the projection of the coordinate system during collection. Since the attitude accuracy of the intelligent terminal diverges rapidly over time and there is no observed data for the heading, it is very difficult to project it into the n system. Currently, during the collection process, it is commonly used to project the magnetic field observation values onto the horizontal plane using an accelerometer. As the indoor space increases in the vertical plane, the magnetic field matching dimension is too small, which easily increases the probability of mis-matching.

[0141] 2. Difficulty in calibrating the magnetometer. Due to the soft magnetic and hard magnetic effects caused by the presence of ferromagnetic materials in the intelligent terminal itself, the magnetometer has errors such as zero bias, scale factor, and cross-axis coupling, and calibration is required before use. There are significant differences in the magnetometers of intelligent terminals from different manufacturers, and even for the same intelligent terminal, the calibration parameters of the magnetometer are not exactly the same at different times. Currently, in intelligent terminals, the magnetometer is calibrated by rotating around three axes and performing ellipsoid fitting to calibrate the error, but this brings great trouble to actual use.

[0142] 3. The magnetic field matching involves a large amount of computation. When estimating the user's current position, particle filtering requires a great deal of calculation. Even when using server-based magnetic field matching positioning, each user needs to start a separate positioning thread, resulting in a still large overall computational load. The best way to reduce particle filtering is to decrease the number of particles, but this will increase the probability of incorrect matching and it is impossible to strike a balance between computational efficiency and positioning performance.

[0143] To reduce the probability of incorrect magnetic field matching, this application establishes an improved magnetic field matching algorithm for the shortest adjacent path. By separately calculating the shortest adjacent factors of different magnetic field components (modulus value, vertical component, etc.), the magnetic field matching result during the entire movement process is obtained according to the shortest path constraint.

[0144] (1) Data acquisition and database construction

[0145] Based on the visualization of IndoorAtlas indoor magnetic field fingerprint acquisition, for any area included in the indoor map, the trajectory is planned by punching points. The collector walks along the predetermined trajectory, keeping the intelligent terminal consistent with the pedestrian's forward direction. When the user reaches the specified waypoint, the waypoint data is obtained by clicking the button. During the preprocessing of the database, interpolation is performed according to the step detection and waypoint data, and the collected magnetic field and other sensor data are projected onto the corresponding positions on the map. The collected magnetometer observation values are projected onto the horizontal and vertical planes, and at the same time, the magnetic heading is calculated by projecting onto the horizontal plane.

[0146] The roll and pitch angles are calculated from the acceleration. During the process of pedestrian movement mapping, in addition to being sensitive to the earth's gravity, the accelerometer also has additional motion input acceleration, resulting in low horizontal angle calculation accuracy. Therefore, to obtain a higher-precision horizontal angle, in this application, during the database construction data processing, first, multi-layer constraint improved pedestrian indoor dead reckoning positioning is performed to obtain higher-precision horizontal and heading angles. The collected magnetic field signals are projected onto the horizontal and vertical planes through projection. In addition, to increase the dimensional data of the magnetic field database, the magnetic heading angle during the movement is calculated, and the difference is taken with the heading angle obtained by PDR to calculate the magnetic declination data corresponding to a certain point, which is used as a dimensional data in the magnetic field database. Since the initial heading is unknown, when using the magnetic declination for matching, the two waveform sequences are shifted and the mean value is subtracted to eliminate the influence of the unknown initial heading. A database is established using the travel distance, as Figure 4 shown. At the same time, considering that the fingerprints may walk in the opposite direction, each segment of the fingerprint is flipped to obtain fingerprints in the opposite walking direction.

[0147] (2) Magnetic field matching algorithm process

[0148] Determine the fingerprint similarity between the observed data and the database. Use the shortest neighboring path to obtain the N optimal matching results for each dimension of the magnetic field signal, convert them into nodes of a graph, and use the improved Dijkstra algorithm with the shortest path constraint to obtain the matching results of the entire marching process.

[0149] 1) The magnetic field matching compares the similarity of two waveforms. Assume the observed value sequence Seq observe =(B o1 …B om ), where m is the sequence length. A certain section of the database fingerprint Seq fingerprint =(B f1 …B fn ), n is the sequence length. Find the similar parts in the two sequences. By moving the two sequences to the same horizontal plane to compare their similarity, determine whether two epochs are similar by whether the function g(o i , f i ) is less than the critical value. If two epochs are similar, assign a score factor of 3; otherwise, set a penalty factor of -3. Calculate the score matrix and determine whether the two sequences are similar by setting parameters. First, to avoid the maximum value of the score matrix being too small (the matching part is too short), set the lowest critical value of the score Th1. Second, set the parameter Th2. When the number of deleted epochs is greater than this critical value, it is considered that the two sequences do not match. Similarly, set the parameter Th3 to avoid inserting too many epochs.

[0150] 2) Analysis of the multi-dimensional shortest neighboring path results. According to the matching results of each dimension in the magnetic field matching, calculate the self-conformity. When the maximum matches of the modulus value, vertical component, horizontal component, and magnetic declination all correspond to a certain section in the database, it is considered that this matching result has a high credibility, and the node numbers corresponding to this observed data section are merged to reduce the calculation amount of the shortest path;

[0151] 3) Shortest path constraint. Each section of the magnetic field observation value may match multiple fingerprints s match ={B fi …B fj}Similarly, an algorithm is designed to connect segments of observation data to form a complete trajectory, and this problem is converted into a shortest path problem. All identified matching fingerprint data segments are converted into vertices of graph G. For any two vertices of the graph, their corresponding coordinates in the fingerprint database are used. Based on the topological relationship in the map, the distance between the two vertices is calculated according to the Dijkstra algorithm. The map constraint data is used. If the distance between the two vertices is greater than the critical value Thdis1, Thdis1=10m is taken, and the corresponding adjacency matrix infinite value INT_MAX is set. Then, the starting point and end point of the shortest path are set, and two new points START and END are added. Among them, START is connected to any fingerprint that matches the first segment of observation data, and END is connected to any fingerprint that matches the last segment of observation data. The length of the newly added edge is set to 0, and the overall matching result is obtained by finding the shortest path from START to END.

[0152] Use the results of matching the shortest adjacent paths in different dimensions of the magnetic field as the vertices of graph G. When the magnetic field modulus, vertical component, horizontal component, and magnetic declination do not have a correct match, ensure that the overall matching result obtained by the shortest path has spatial continuity, set parameters to eliminate spatial discontinuity, and if a certain distance in the obtained shortest path is greater than the critical value Thdis2, take Thdis2=10m, and determine that the matching result of this segment is spatially discontinuous; add a new end point END1 before matching this segment, and add a new starting point START1 after matching this segment, and convert the shortest path problem from START to END into the sum of the shortest distances from START to END1 and from START1 to END. If there is also spatial discontinuity from START1 to END, repeat the above steps.

[0153] 3. Indoor positioning system based on dead reckoning / Bluetooth / magnetic field

[0154] 1. Bluetooth fingerprint recognition

[0155] The construction of Bluetooth indoor positioning system based on fingerprint method includes two stages: database establishment and matching positioning. In the database establishment stage, the Bluetooth fingerprint database is generated by walking surveying. The map coordinates corresponding to the pedestrian track are interpolated to obtain the map coordinates at the time of Bluetooth signal reception. Then, within the range of fingerprint collection, a 2m square grid is established, and the Bluetooth signal during walking is planned to the nearest grid as the fingerprint signal of the grid point. The collected initial RSS signal is processed to improve the database quality. In the first step, the RSS signal below the critical value is eliminated; in the second step, when there are multiple initial signals planned to a certain grid point, the average is used to reduce the noise in the database. Finally, the established Bluetooth fingerprint signal is as follows:

[0156] DBi = {Pos i , (mac1, RSS1), … (mac n , RSS n )} Equation 14

[0157] where Pos i is the coordinate of the fingerprint point, mac is the mac address of the AP, RSS is the signal strength of the corresponding AP, and the finally established Bluetooth fingerprint is as Figure 5 .

[0158] After collecting the initial Bluetooth signal, preprocess the data. First, delete the Bluetooth RSS data below the critical value; use the average of the latest 3 Bluetooth observed RSS values as the RSS value of the current epoch Bluetooth observation. For the preprocessed Bluetooth signal, if the number of APs is sufficient, use the observed data to compare with the fingerprint points in the database one by one, and select K (4 in this application) fingerprint points with the closest distance (Euclidean distance). Then, process the selected K RPs. The K fingerprint points approach each other, and eliminate the RPs whose distance from the mean value of all RPs is greater than a certain critical value (15m in this application). Use the mean value of the coordinates of the RPs that meet the conditions as the positioning result of fingerprint matching. Finally, check the positioning result. If the matching result is close to the historical positioning result, it is considered that the positioning result meets the requirements; otherwise, the matching result of this time is treated as a mis-match.

[0159] (II) Dead Reckoning / Bluetooth / Magnetic Field Integrated Navigation

[0160] The indoor positioning based on Bluetooth fingerprint, pedestrian dead reckoning, and magnetic field matching each has its own advantages and disadvantages. The indoor positioning accuracy based on Bluetooth fingerprint is greatly affected by the deployment density. In places where APs are sparse, the positioning accuracy is poor, and the positioning result fluctuates greatly. The positioning based on pedestrian dead reckoning is basically not affected by the environment, the positioning result is continuous and has small fluctuations, making up for the defects of Bluetooth fingerprint positioning. However, pedestrian dead reckoning is a relative positioning that requires obtaining the initial position and direction, and the PDR positioning error accumulates continuously over time. The indoor positioning based on magnetic field matching is related to the quality of the magnetic field fingerprint. The higher the distinguishability between magnetic field fingerprints, the better the matching result. On the contrary, if the magnetic field fingerprints are not clearly distinguishable from each other, it is very difficult to obtain a good positioning result. For pure magnetic field matching, as the database continues to increase, the probability of mis-match also correspondingly increases.

[0161] Implement Bluetooth / magnetic field / PDR integrated navigation using dynamic optimal filtering. Adopt a loose integration method for the combination of Bluetooth and PDR. When the pedestrian walks a certain distance, obtain the current user position observation value through magnetic field matching, and use the magnetic field matching result as the observation data of dynamic optimal filtering to achieve the integrated navigation of magnetic field and PDR. PDR provides continuous navigation results, and the results of Bluetooth matching and magnetic field matching jump back and forth. By comparing the PDR result with the results of Bluetooth and magnetic field matching, when their distance is greater than a certain critical value, it is determined that the result of fingerprint matching has low credibility, realizing the quality control of observation data.

[0162] To effectively integrate multi-source data and give full play to their respective advantages, this application adopts Figure 6 the integrated navigation algorithm shown below.

[0163] The improved PDR algorithm takes INS as the core, ensures the performance of the positioning system through multi-layer constrained data, uses PDR as the system framework to provide continuous positioning results, and Bluetooth and magnetic field are used as position observation data to provide system position correction. Under the condition of sufficient BLE deployment, it provides high-precision positioning results. The positioning result of magnetic field matching is used as position reference data. However, the magnetic field positioning accuracy is related to the specific position. Bluetooth and magnetic field are used as supplementary means to improve the accuracy of the integrated navigation system by correcting the positioning result.

[0164] Construct a dynamic optimal filtering observation equation using the position obtained by magnetic field matching or Bluetooth fingerprint recognition:

[0165] p PDR -p obs = δp + n Equation 15

[0166] where, P PDR is the position data predicted by PDR, P obs is the result of Bluetooth or magnetic field matching, n is the noise of the observation signal, and δp is the position error; the design matrix of dynamic optimal filtering:

[0167] H = |I 3×3 O 3×3 O 3×3 O 3×3 O 3×3 | Equation 16

[0168] The Q matrix of dynamic optimal filtering:

[0169]

[0170] where, VRW represents velocity random walk, ARW is angular random walk, σ bg is the gyro bias noise variance, σ ba is the accelerometer bias noise variance;

[0171] The quality control in this application includes: judging the credibility of Bluetooth and magnetic field matching results by using the continuity of PDR positioning results; using Bluetooth positioning results or current integrated navigation results to reduce the search space of magnetic field matching.

[0172] (3) Experimental results and analysis

[0173] In order to verify the multi-source fusion positioning strategy of this application, its positioning results are compared with magnetic field matching and Bluetooth fingerprint recognition. The experiment is repeated multiple times. During the movement of pedestrians, timestamps are marked at waypoints to obtain the reference positions during the walking process of pedestrians. The positioning performances of different strategies are analyzed by comparing the results of magnetic field matching, Bluetooth fingerprint positioning and multi-source fusion. Figure 7 、 Figure 8 、 Figure 9 The trajectory results of different positioning strategies for three groups of experimental data are shown. It can be seen that the magnetic field matching and multi-source fusion positioning basically coincide with the real trajectory. Due to the K-means algorithm, the Bluetooth positioning results tend to move towards the middle of the trajectory.

[0174] The advantage of magnetic field matching is that it has high positioning performance when the matching is correct, but it is easy to find mis-matches. At this time, the multi-source fusion strategy needs to play its role. In order to quantitatively analyze the improvement of the multi-source fusion strategy of this application on the performance of the positioning system, this application conducts repeated experiments and compares the RMS values of different strategies. The positioning accuracy of pure Bluetooth matching is about 4m, and the positioning accuracy is poor. The positioning accuracy of the multi-source fusion strategy proposed in this application is better than 2m, basically meeting the needs of consumer-grade pedestrian indoor navigation. The magnetic field positioning accuracy in Experiments 1 and 2 is greater than two meters because of mis-matching phenomena. As shown in the figure, by fusing PDR and Bluetooth positioning, the influence of mis-matches on the positioning performance can be suppressed. Compared with MM, the positioning performances of the fusion strategies in this application are improved by 33% and 18% respectively.

[0175] Figure 10 This is the cumulative error distribution diagram of multiple experiments in this application. When the magnetic field matching is correct, due to the high positioning accuracy of magnetic field matching, the multi-source fusion strategy does not improve the positioning performance. When the Bluetooth positioning results are poor, the positioning performance is even slightly lower than that of pure magnetic field matching. Generally speaking, the multi-source fusion positioning performance is basically equivalent to that of magnetic field matching. When there are mis-matches in the magnetic field, the robustness of the overall multi-source fusion positioning performance is obvious, and the positioning error is basically better than 5m. The overall performance of Bluetooth matching positioning is poor, but when there are mis-matches in the magnetic field or it is difficult to use the magnetic field in indoor open areas, the positioning results are used to assist PDR to ensure the availability of the positioning system. The following conclusions are obtained through the above analysis: The multi-source fusion strategy has higher positioning accuracy and higher stability than using Bluetooth fingerprint or magnetic field matching alone.

[0176] (4) Real-time positioning system architecture

[0177] The APP designed in this application is still in the experimental testing stage, and based on the server system architecture, a public network IP is required. However, there is a lack of a public network IP in the experiment. Therefore, this application selects a client-based system architecture as the implementation of the positioning system in this application.

[0178] The indoor positioning system designed in this application includes the following modules: a multi-source data acquisition module, a map display module, a Bluetooth fingerprint matching module, a PDR solution module, and a multi-source fusion module. Due to the inconsistent calibration parameters of magnetometers in different intelligent terminals, it is difficult to calibrate magnetometers accurately and reliably. When the magnetometer is not calibrated or the calibration result is poor, it will instead contaminate the positioning result. At the same time, the real-time performance of magnetic field matching based on the client is greatly affected by its hardware performance. When the hardware computing power is insufficient, real-time positioning results cannot be obtained. Therefore, the real-time program in this application does not include a magnetic field matching module.

[0179] The data collected by the multi-source data acquisition module includes Bluetooth, accelerometer, gyroscope, and magnetometer. Taking the Android system as an example, the operating system provides APIs for these sensors, which facilitates users to directly call and obtain sensor data.

[0180] The Bluetooth fingerprint matching module uses the scanned Bluetooth signals and the corresponding signal strength values to form observation data in a certain format. This observation data is compared with the RP in the database, and the Bluetooth fingerprint matching result is obtained using WKNN and the robust strategy proposed in this application.

[0181] The PDR positioning module collects the initial observation values of the accelerometer, gyroscope, and magnetometer. The accelerometer data is used for step detection and step length estimation, and at the same time, the horizontal attitude angle is corrected in a tightly coupled manner. The gyroscope is used for integral calculation of the heading and state detection. The magnetometer observation data is used to correct the heading angle. The relative position of the pedestrian is calculated through the PDR algorithm.

[0182] The multi-source fusion module obtains the final positioning result according to the Bluetooth fingerprint matching result and the PDR calculation result using the multi-source fusion method.

[0183] The map display module uses the mapbox map SDK to display the three-dimensional indoor map of the test site, and shows the user's current location by adding marker points. At the same time, the current positioning result is optimized according to the map data. Unreasonable positioning results such as passing through walls are avoided by constraining the positioning result to the path.

[0184] (V) Overall Effect of the APP

[0185] The overall effect of the multi-source fusion indoor positioning APP based on Android is as Figure 11As shown. The main interface is used for map display, floor switching, and 2D and 3D view switching. Whether the user's position is at the center of the map and the map rotates along with the forward direction.

[0186] This application analyzes the advantages and disadvantages of PDR, magnetic field matching, and Bluetooth fingerprint positioning, and proposes a multi-source data fusion indoor positioning strategy. By fusing multiple data sources, the deficiencies are mutually compensated to improve the performance of the positioning system. Through experimental comparison, the multi-source fusion positioning results are compared with the Bluetooth fingerprint and magnetic field matching results. The experimental results show that the accuracy and stability of the multi-source fusion positioning results are higher than those of the individual magnetic field and Bluetooth positioning, and it has higher stability, verifying the feasibility of the multi-source fusion algorithm of this application. Finally, a real-time visualization indoor positioning APP is developed based on the Android terminal.

Claims

1. A method for multi-source fusion 3D visual high-precision indoor navigation on a mobile device, characterized in that, Improve pedestrian dead reckoning positioning, magnetic field matching positioning, and the multi-source fusion method of PDR, magnetic field, and Bluetooth. First, construct an improved robust PDR scheme based on the MEMS sensors built into intelligent terminals. Optimize and utilize multi-layer constraint data with INS as the core, including human motion model constraints, quasi-static, quasi-static magnetic fields, turning detection, etc., and use them as observation data to construct a dynamic optimal filtering measurement model. By estimating and compensating sensor errors in real time, improve the robustness and positioning accuracy of the PDR algorithm. Second, establish a magnetic field matching algorithm improved by the shortest path constraint. Add the magnetic declination as a dimension of the magnetic field fingerprint to the magnetic field matching algorithm. Use the magnetic field fingerprints of multiple dimensions including modulus, horizontal component, vertical component, and magnetic declination to obtain the shortest path constraint results respectively, convert them into vertices of a graph, and then obtain the overall matching result through the shortest path constraint to reduce the probability of mis-matching. Finally, in the multi-source fusion integrated navigation algorithm, utilize the complementary characteristics of different positioning data sources to establish an algorithm that fuses multi-source data of PDR, magnetic field, and Bluetooth, set up a quality control mechanism, improve the accuracy and robustness of the positioning system, and realize a visualization indoor positioning system based on intelligent terminals; 1) Improved pedestrian indoor dead reckoning positioning with multi-layer constraints: Establish an observation equation constructed from multi-layer constraint data. Use the multi-layer constraint data as the dynamic optimal filtering observation signal to extend the available time of the existing PDR through effective error feedback; 2) Improved magnetic field feature matching algorithm: Through data collection and database construction, add the magnetic declination as a dimension in the magnetic field fingerprint database to increase the dimensional data of the magnetic field fingerprint, and obtain the overall matching result through the shortest path constraint; 3) Indoor positioning system based on dead reckoning / Bluetooth / magnetic field: Construct a strategy for multi-source data fusion indoor positioning, fuse multiple data sources to make up for each other's deficiencies, and develop a real-time visualization indoor positioning APP based on Android terminals using Mapbox; 2. The mobile - end multi - source fusion 3D visual high - precision indoor navigation method according to claim 1, wherein, Improved pedestrian indoor dead reckoning positioning with multi-layer constraints: Construct an improved pedestrian dead reckoning algorithm guided by INS prior. Use multi-layer constraint data as the observation equation of dynamic optimal filtering to improve the positioning accuracy of the system, thereby maximizing the available time of the PDR algorithm; The multi-layer constraint data used includes ZUPT, ZARU, gravity vector constraint, NHC, and magnetic field vector constraint. The specific steps to improve the PDR algorithm are as follows: First, perform sensor error compensation on the initial data of the accelerometer, gyroscope, and magnetometer, and then perform strapdown inertial navigation mechanical scheduling. At the same time, detect the states of the user including foot points, zero speed, low dynamics, and quasi-static magnetic fields. If the constraint conditions are met, trigger the corresponding measurement update, and at the same time perform error feedback to correct the current navigation state data. Finally, return to the step of inputting the initial sensor signal.

3. The mobile - end multi - source fusion 3D visual high - precision indoor navigation method according to claim 1, characterized in that, Observation equation constructed from multi-layer constraint data: The multi-layer constraint data used includes: ZUPT, ZARU, NHC, gravity vector, magnetic field vector; 1 - Quasi - static constraint: Based on the ZUPT detection method, a method combining a critical value and a hypothesis test item, and weighting the gyroscope and accelerometer is used to determine whether a pedestrian is in a quasi - static state. When it is detected that the pedestrian satisfies the quasi - static condition, the pedestrian speed is determined to be zero, and the pseudo - speed observation value v is calculated based on the assumed conditions n = [0 0 0] T , corresponding to the error observation equation: Among them, is the current velocity calculated by inertial navigation extrapolation in the n system; Meanwhile, it is assumed that the pedestrian's heading remains unchanged in the quasi-static state, the output of the heading gyroscope is 0, and all heading changes are caused by the error of the heading gyroscope. Calculate the pseudo-observation value ω of the heading angular rate z = 0, corresponding to the error observation equation: Among them is the output of the heading gyroscope, and the design matrix under quasi-static conditions is as follows: 2 - Gravity vector constraint: Under the condition of determining no external acceleration, updating the navigation state using the gravity vector obtained from acceleration observations can ensure the horizontal angle accuracy. Whether to use the acceleration observation value is measured according to the magnitude of A in Equation 4: A = |norm(f b ) - g| Equation 4 where g is the local acceleration due to gravity. Under the condition of a small external input acceleration, if A ≤ |TH acc |, the standard deviation of the acceleration measurement noise is σ a , representing the noise level of the accelerometer. Under the condition of a large external acceleration input, if A > |TH acc |, the accelerometer is not used to assist in the horizontal attitude angle, and TH acc is the critical value; Under static and low - dynamic conditions, the attitude angle is assisted based on the acceleration measurement data, and the acceleration observation equation is constructed in a tightly - coupled manner to ensure that the vehicle does not have the ±90deg Euler angle singularity problem in various motion states. The acceleration error equation is: where f n = -g n = [0 0 g] T , ψ is the attitude error, n is the measurement noise, and the design matrix under low-dynamic conditions is: H = |0 3×3 0 3×3 g n × 0 3×3 0 3×3 | Equation 6 3 - Pedestrian motion constraint: When the vehicle body does not skid or jump, it is considered that the vehicle body only has a speed in the forward or backward direction. If the pedestrian does not show an abnormal walking state, they will move forward along the corridor. At this time, there is only a forward speed, meeting the NHC condition. Under turning conditions, there is a skidding phenomenon similar to that in vehicle navigation, and it is necessary to turn off NHC in a timely manner to ensure the stability of the system. Whether to use NHC is based on turning detection. The heading gyro observation value combined with the sliding - window threshold is used to judge whether the pedestrian is in a turning state: is the observed value of the current gyroscope, min represents the minimum value of a time window, and TH gyro is the critical value. When it is determined that the pedestrian is in a straight walking state, the pseudo-velocity observed value v b =[[v vx 0 0] T , and an error equation is constructed according to the NHC hypothesis: where v n is the current velocity calculated by inertial navigation in the n-frame, is the direction cosine matrix for rotating the n-frame to the b-frame, and v vx represents the forward velocity, stepL and △t are the estimated step length and footstep time interval respectively: v vx = stepL / Δt Equation 9 where a zmax , a zmin represents the maximum and minimum values of the acceleration modulus within the sliding time window, K represents a coefficient, and the design matrix under non-integrity constraints is: 4 - Quasi - static magnetic field (QSMF) constraint: In a local area, especially in an indoor open area, the environmental magnetic field remains stable. The magnetometer correction data is fully utilized without reducing the robustness of the system. By detecting the surrounding environmental magnetic field, the magnetometer vector constraint is used only when the environmental magnetic field is QSMF. The judgment of the quasi - static condition is as follows: Project the current - epoch magnetometer observation value onto the n - system. By comparing the difference in the environmental magnetic field between the current epoch and the buffer, it is judged whether the current epoch is a quasi - static magnetic field. Specifically: where N is the number of cached epochs, and TH mag is the critical value of the quasi-static magnetic field; When it is judged that the current epoch is in a quasi - static magnetic field, the magnetic field vector is used to assist the current attitude angle, and the magnetometer observation equation is constructed in a tightly - coupled manner. The magnetic field vector error equation is: where m n is the calibrated ambient magnetic field, is the magnetometer observation value, and the design matrix under the quasi-static magnetic field condition is: Based on UPT, ZARU, NHC, the gravity vector, and the magnetic field vector, the observation equation is constructed through multi - layer constraint data.

4. The mobile - side multi - source fusion 3D visual high - precision indoor navigation method according to claim 1, wherein, Improved data acquisition and database construction for magnetic field feature matching: Based on the visualization of IndoorAtlas indoor magnetic - field fingerprint acquisition, for any area included in the indoor map, the trajectory is planned by dotting. The collector walks along the predetermined trajectory, keeping the intelligent terminal consistent with the forward direction of the pedestrian. When the user reaches the specified waypoint, the waypoint data is obtained by clicking the button. During the pre - processing of the database, interpolation is performed according to the footstep detection and waypoint data, and the collected magnetic field and other sensor data are projected onto the corresponding positions on the map. The collected magnetometer observation values are projected onto the horizontal and vertical planes, and at the same time, the magnetic heading is calculated by projecting onto the horizontal plane; When processing the data for building the database, first perform multi-layer constraint improvement on pedestrian indoor dead reckoning positioning to obtain more accurate horizontal and heading angles. Project the collected magnetic field signals onto the horizontal and vertical planes through projection. In addition, calculate the magnetic heading angle during walking, and calculate the difference from the heading angle obtained by PDR to calculate the magnetic declination data corresponding to a certain point, which is used as a dimensional data in the magnetic field database. When using the magnetic declination for matching, subtract the mean from the two waveform sequences to eliminate the influence of the unknown initial heading. Establish the database using the travel distance, and at the same time consider that the fingerprint may walk in the opposite direction. Flip the segments of the fingerprint to obtain the fingerprint in the opposite walking direction.

5. The mobile - end multi - source fusion 3D visual high - precision indoor navigation method according to claim 1, characterized in that, Magnetic field matching algorithm process: Determine the similarity between the observed data and the fingerprints in the database. Use the shortest neighbor path to obtain the N optimal matching results for each dimension of the magnetic field signal respectively, and convert them into the nodes of the graph. Use the improved Dijkstra algorithm with the shortest path constraint to obtain the matching result of the entire walking process.

6. The mobile - end multi - source fusion 3D visual high - precision indoor navigation method according to claim 5, characterized in that: The magnetic field matching compares the similarity of two waveforms. Assume the observed value sequence Seq observe =(B o1 …B om ), where m is the sequence length. A certain section of the database fingerprint Seq fingerprint =(B f1 …B fn ), n is the sequence length. Find the similar parts in the two sequences. By moving the two sequences to the same horizontal plane to compare their similarity, and judge whether two epochs are similar by whether the function g(o i , f i ) is less than the critical value. If two epochs are similar, assign a score factor of 3, otherwise set the penalty factor to -3. Calculate the score matrix, and judge whether two sequences are similar by setting parameters. First, to avoid the maximum value of the score matrix being too small (the matching part is too short), set the lowest critical value of the score Th1. Second, set the parameter Th2. When the number of deleted epochs is greater than this critical value, it is determined that the two sequences do not match. Similarly, set the parameter Th3 to avoid inserting too many epochs.

7. The mobile - side multi - source fusion 3D visual high - precision indoor navigation method according to claim 5, wherein: Analysis of the multi-dimensional shortest neighbor path results. Calculate the self-conformity according to the comparison of the matching results of each dimension in the magnetic field matching. When the maximum matches of the modulus, vertical component, horizontal component, and magnetic declination all correspond to a certain segment in the database, it is considered that the matching result has a high credibility. Merge the number of nodes corresponding to the observed data segment to reduce the calculation amount of the shortest path.

8. The mobile terminal multi-source fusion 3D visual high-precision indoor navigation method according to claim 5, characterized in that: Shortest path constraint, each magnetic field observation value may be similar to multiple fingerprints s in the database match ={B fi …B fj}. It is necessary to design an algorithm to connect the observation data segments to form a complete trajectory, convert this problem into a shortest path problem, convert all recognized matching fingerprint data segments into vertices of graph G. For any two vertices in the graph, use their corresponding coordinates in the fingerprint database, based on the topological relationship in the map, calculate the distance between the two vertices according to the Dijkstra algorithm, and adopt the map constraint data. If the distance between two vertices is greater than the critical value Thdis1, take Thdis1 = 10m, set the corresponding adjacency matrix infinite value INT_MAX, and then set the starting point and ending point of the shortest path. Add two new points START and END. Among them, START is connected to any fingerprint that matches the first observation data segment, and END is connected to any fingerprint that matches the last observation data segment. The lengths of the newly added edges are all set to 0, and the overall matching result is obtained by finding the shortest path from START to END; Use the shortest neighbor path results of different dimensions of the magnetic field as the vertices of graph G. When there is no correct match for the magnetic field modulus, vertical component, horizontal component, and magnetic declination, ensure that the overall matching result obtained by the shortest path has spatial continuity. Set parameters to eliminate spatial discontinuity phenomena. If the distance of a certain segment in the obtained shortest path is greater than the critical value Thdis2 (take Thdis2 = 10m), it is considered that the matching result of this segment is spatially discontinuous; add a new end point END1 before this segment of the match, and add a new start point START1 after this segment of the match. Convert the shortest path problem from START to END into the sum of the two shortest distances from START to END1 and from START1 to END. If there is also spatial discontinuity from START1 to END, repeat the above steps.

9. The mobile - side multi - source fusion 3D visual high - precision indoor navigation method according to claim 1, wherein, Bluetooth fingerprint recognition: Build a Bluetooth indoor positioning system based on the fingerprint method, including two stages: building the database and matching for positioning. In the stage of building the database, generate the Bluetooth fingerprint database by means of walking mapping. Interpolate the map coordinates corresponding to the pedestrian dead reckoning to obtain the map coordinates at the moment of Bluetooth signal reception, and then establish a 2m square grid within the fingerprint collection range. Plan the Bluetooth signals during walking to the nearest grid as the fingerprint signal of this grid point. Process the collected initial RSS signals to improve the quality of the database. First step, eliminate the RSS signals below the critical value; second step, when there are multiple initial signals planned to a certain grid point, use the method of taking the mean to reduce the noise in the database. Finally, the established Bluetooth fingerprint signal is as follows: DB i = {Pos i , (mac1, RSS1), … (mac n , RSS n )} Equation 14 Among them, Pos i is the coordinate of the fingerprint point, mac is the mac address of the AP, RSS is the signal strength of the corresponding AP, and the finally established Bluetooth fingerprint; After the initial Bluetooth signal is collected, pre - processing of the data is carried out. First, Bluetooth RSS data below the critical value is deleted; The average of the latest 3 Bluetooth observed RSS values is used as the RSS value of the current epoch for Bluetooth observation. For the pre - processed Bluetooth signal, if the number of APs is sufficient, the observed data is compared with the fingerprint points in the database one by one, and K fingerprint points with the closest distances are selected. Then, the selected K RPs are processed. The K fingerprint points are made to approach each other, and the RPs with distances greater than a certain critical value from the mean of all RPs are removed. The mean of the RP coordinates that meet the conditions is used as the positioning result of fingerprint matching. Finally, the positioning result is verified. If the matching result is similar to the historical positioning result, it is determined that the positioning result meets the requirements; otherwise, this matching result is treated as a false match.

10. The mobile - side multi - source fusion 3D visual high - precision indoor navigation method according to claim 1, wherein, Dead - reckoning / Bluetooth / magnetic field integrated navigation: Dynamic optimal filtering is used to achieve Bluetooth / magnetic field / PDR integrated navigation. A loose - coupling method is adopted for the combination of Bluetooth and PDR. When the pedestrian walks a certain distance, the current user position observation value is obtained through magnetic field matching. The magnetic field matching result is used as the observation data for dynamic optimal filtering to achieve the integrated navigation of magnetic field and PDR. PDR provides continuous navigation results, and the results of Bluetooth matching and magnetic field matching jump back and forth. By comparing the PDR result with the Bluetooth and magnetic field matching results, when their distances are greater than a certain critical value, it is determined that the result of fingerprint matching has low credibility, realizing the quality control of the observation data; The improved PDR algorithm takes INS as the core and ensures the performance of the positioning system through multi - layer constrained data. PDR is used as the system framework to provide continuous positioning results, and Bluetooth and magnetic field are used as position observation data to provide system position correction. Under the condition of sufficient BLE deployment, high - precision positioning results are provided. The positioning result of magnetic field matching is used as position reference data, but the magnetic field positioning accuracy is related to the specific position. Bluetooth and magnetic field are used as supplementary means to improve the accuracy of the integrated navigation system by correcting the positioning result; Construct a dynamic optimal filtering observation equation using the position obtained by magnetic field matching or Bluetooth fingerprint recognition: p PDR -p obs = δp + n Equation 15 where P PDR is the position data predicted by PDR, and P obs is the result of Bluetooth or magnetic field matching, n is the noise of the observed signal, and δp is the position error; the design matrix of the dynamic optimal filter: H = |I 3×3 0 3×3 0 3×3 0 3×3 0 3×3 | Equation 16 The Q matrix of dynamic optimal filtering: Among them, VRW represents velocity random walk, ARW is angular random walk, and σ bg is the gyro bias noise variance, and σ ba is the accelerometer bias noise variance; Quality control includes: judging the credibility of Bluetooth and magnetic field matching results using the continuity of the PDR positioning result; using the Bluetooth positioning result or the current integrated navigation result to reduce the search space of magnetic field matching.

Images