A Landmark Matching PDR Positioning Method in Multi-Intersection Scenarios

Abstract

This invention discloses a landmark matching PDR positioning method in multi-intersection scenarios, belonging to the field of indoor positioning technology. The method enhances landmark credibility by constructing landmarks through multi-dimensional information perception. Simultaneously, by adaptively extracting landmarks with different spatial features, this method is applicable to various landmarks, such as some unconventional intersections. Furthermore, this application uses different thresholds to define the radius of influence of different landmarks, solving the localization system mismatch problem caused by traditional single-location detection of landmarks, thus improving positioning accuracy to a certain extent. This application also further improves positioning accuracy by incorporating heading weights into trajectory similarity metrics, achieving dual correction of the position and heading of moving targets.

Description

Technical Field

[0001] This invention relates to a landmark matching PDR positioning method in multi-intersection scenarios, belonging to the field of indoor positioning technology. Background Technology

[0002] The Global Positioning System (GPS) is widely used for outdoor location services, but signal loss due to building obstruction prevents it from functioning effectively indoors. With the increasing capabilities of mobile computing and sensing, pedestrian dead reckoning (PDR) technology based on smart terminals, especially mobile phones, is attracting growing attention from location service researchers. As an autonomous positioning technology, PDR uses a triple vector of accelerometers, gyroscopes, and magnetometers to estimate pedestrian positions, achieving accurate positioning and target tracking over short time and distances, thus making it suitable for indoor positioning requirements. However, due to the inevitable cumulative errors of mobile sensors, PDR information deviates from the true value over time, necessitating periodic benchmark updates to clear these accumulated errors. Indoor landmarks are typically used as reference locations for clearing these errors. Once a landmark is accurately detected, the original cumulative error in the PDR trajectory estimation can be eliminated based on it, but accurate acquisition of the landmark's location is crucial.

[0003] In another indoor positioning technology, WiFi indoor positioning, fingerprint matching offers higher location estimation accuracy and better adaptability to complex environments. Researchers select WiFi fingerprint points in specific indoor areas such as corners and staircases, fusing the received signal strength (RSS) data at these fingerprint points with WiFi access point (AP) data to create a landmark database. Landmarks are constructed by detecting peaks and troughs in the WiFi received signal strength. An adaptive weighted fusion algorithm improves the location accuracy of the landmarks, defining and processing signal patterns of the same level to address inconsistencies in WiFi signal observations, resulting in WiFi landmarks with higher location reliability. However, this method typically relies on a single WiFi signal to obtain landmark locations. The fluctuations in WiFi signals, multipath interference, and variations in peak position reduce the reliability of the landmarks.

[0004] To improve the reliability of landmarks, existing methods employ active map matching (AMM) to correct PDR trajectories and reduce accumulated PDR errors, thus mitigating the instability of WiFi landmarks to some extent. For example, accelerometer feature values ​​are used to assist in identifying user behaviors such as walking, climbing stairs, and riding elevators, constructing corresponding landmarks, and correcting indoor pedestrian positions. User behavior landmarks are obtained from crowdsourced data, and stable indoor landmarks are obtained through clustering techniques, which are then used to construct indoor maps. However, the performance of AMM-based localization solutions largely depends on the accuracy of activity recognition and map matching, which can easily lead to localized detection of pedestrian activities in specific indoor locations, potentially resulting in incorrect position matching. This is because there is no constant correspondence between user activity behavior and indoor spatial location, making landmark features indistinct. Therefore, establishing landmarks with significant features and a larger recognition area is particularly important. Researchers have proposed a structural landmark approach that uses the structural features of indoor spaces to correct the cumulative error of pedestrian directional tracking (PDR). This method integrates visual and inertial sensor information to detect structural landmarks (such as intersections, T-junctions, and corners) to correct user trajectories. This approach directly reflects and perceives the spatial characteristics of the indoor environment and can effectively correct pedestrian heading errors. However, the proposed structural landmark features are limited and cannot be applied to unconventional intersections, such as Y-junctions or multi-way intersections. Figure 1 As shown; and the visual sensor has strong requirements for the phone's pose, it must face forward to ensure the adaptability of the acquired images. Summary of the Invention

[0005] To address the existing problems, this invention provides a landmark matching PDR positioning method for multi-intersection scenarios, including:

[0006] Step 1: Build a system model for the multi-intersection scenario to be located;

[0007] Step 2: Obtain multi-dimensional information of each landmark in the multi-intersection scene to be located to form the multi-dimensional features of the landmark, including the center location information of the landmark, the radius of influence, information of all branch roads of the landmark, and Wi-Fi feature fingerprint;

[0008] Step 3: Match PDR positioning based on the multidimensional features of the landmark. During the positioning process, the heading weight is incorporated into the trajectory similarity metric to achieve dual correction of the position and heading of the moving target, which is equipped with an inertial sensor.

[0009] Optionally, step 1 includes:

[0010] Assuming the multi-intersection scenario to be located includes A landmark, ; detectable in the described scenario Each AP base station, then each landmark ;

[0011] in, Indicates the first Central location information of each landmark Indicates the first The radius of influence of each landmark; Indicates landmark The Location of each branch feature point and heading , and ; Indicates landmark Wi-Fi signature fingerprints will soon be used in landmarks. The mean RSS vector of each AP base station collected at the turning termination point is used as a feature variable for landmark range detection;

[0012] express Within the time period Inertial sensors collect raw data at all times Set, containing the original data Indicates the first The data collected by the inertial sensor at all times On-axis acceleration and angular velocity .

[0013] Optionally, step 2, obtaining all branch road information of the landmark, includes:

[0014] Define the turning endpoint: if the moving target At any given time, a turning action occurs within a certain landmark area, and... If the pedestrian changes from turning to going straight at a certain moment, then the trajectory formed by the turning point is represented as follows: Turning point This refers to a turning point within the landmark area;

[0015] Behavioral data is collected using inertial sensors carried by the moving target. The behavioral state of the moving target at each moment is identified based on the behavioral data. The behavioral state includes straight-line movement and turning. The behavioral data includes acceleration information and angular velocity information.

[0016] Determine the location of each turning termination point based on the behavior state of the moving target;

[0017] Data was collected at each steering endpoint. The RSS vector of each AP;

[0018] Based on the position of the turning termination point, heading, and collected data The RSS vectors of each AP together constitute the sample set. The heading is calculated based on the angular velocity information in the behavior data, and the RSS vector of the AP is obtained through acquisition.

[0019] Gaussian mixture clustering algorithm is used to obtain Clusters of turning termination points ;

[0020] Location of the cluster center Its corresponding heading information Together they form the cluster center point set of turning termination points ,in , Representing the cluster center points respectively Axis coordinates;

[0021] If the turning point terminates at the cluster center point In the neighborhood radius There is another point within the range. and Connected, then for The set of connected points Inner element; at the same time, if This indicates Inner elements and Indirect connection, all with The set of points that are directly or indirectly connected This represents a locally connected space, and all locally connected spaces constitute a set of landmark branch feature points with sequence identifiers. .

[0022] Optionally, identifying the behavioral state of the moving target at each moment based on the behavioral data includes:

[0023] Define behavior tags ,for consecutive integers within the interval, where The number of all possible behavior state categories, including going straight and turning;

[0024] Regarding the current moment Sampled by inertial sensor Before the moment Continuous data at each moment constitutes the observation data. ;

[0025] Using the Long Short-Term Memory (LSTM) model to give Time-of-action label vector and through The function extracts the index of the maximum value in a vector to obtain the label of the current row, as shown in the following formula:

[0026] (1)

[0027] in, Indicates going straight. Indicates a change of direction.

[0028] Optionally, step 2, obtaining the center location information of the landmark, includes:

[0029] Targeting landmarks Let its landmark branch feature point set be set. The center point of the landmark Landmark branch road feature points Intersection with its reverse extension line The calculation formula is as follows:

[0030] (2)

[0031] when At that time, the landmark only contains two branch feature points, whose backward extensions intersect at one point. There is a unique solution;

[0032] when When a landmark contains multiple branch feature points, and the intersection of their backward extensions is not unique, the real physical point is represented by the least squares solution of the system of equations. The approximate solution is, The solution.

[0033] Optionally, step 2, obtaining the influence radius and Wi-Fi feature fingerprint, includes:

[0034] For different landmarks, the radius of influence of the landmark is defined using different thresholds. :

[0035] Targeting landmarks ,definition The set of all clusters to which the cluster centers of all turning termination points belong is ,but It is expressed as follows:

[0036] (3)

[0037] In the formula, ;

[0038] Correspondingly, for:

[0039] (4)

[0040] In the formula, for Number of elements in the set for Collect data at each point within the set. RSS vectors of APs.

[0041] Optionally, step 3 includes:

[0042] Define adaptive threshold for:

[0043] (6)

[0044] in, as a landmark The mean of the similarity measure, This indicates the turning termination point determined in real time during the matching phase. The calculation formula is as follows:

[0045] (5)

[0046] when At that time, it was detected that the pedestrian's current turning termination point was located at the landmark. Location, and then based on landmarks Correct its position and course.

[0047] Optionally, step 3 further includes extracting the steering feature trajectory:

[0048] Assumption and as a landmark Any two landmark branch feature points within the area;

[0049] The formula for determining road segments is defined as follows:

[0050] (7)

[0051] In the formula, and They are respectively and The course, if If the two landmark branch feature points are located in different road segments, then the landmarks can be extracted. Set of turning feature trajectories within the range Including the turning characteristic trajectory of the same branch road and different branch turning characteristic trajectories ,in For the turning feature point, express Unique identifier within the set;

[0052] Turning feature points are defined as: given landmarks If it exists Then in Inside On the bisector of the line, there must exist a point. and the circle centered at that point and the line segment and The shorter one is tangent to , At the same time with Hand over , That is, the landmark A turning feature point;

[0053] like , , for Vector angle, satisfying ,but and The position can be calculated using equation (8):

[0054] (8)

[0055] when hour, It conforms to the actual turning characteristic trajectory trend.

[0056] Optionally, step 3 further includes:

[0057] Incorporating heading weights into the trajectory similarity metric, based on the pedestrian's turning termination point time. Guided by this, and then assess pedestrians. Turning trajectory at all times and detected landmarks Turning characteristic trajectory within range Similarity results between The calculation method is as follows:

[0058] (9)

[0059] In the formula, The Frescher distance between two trajectories is the minimum value of the elements in the set of the maximum Euclidean distance between the sequence pairs of points in the two discrete trajectories.

[0060] Representing two trajectories and Heading difference weights:

[0061] (10)

[0062] Therefore, by judging the similarity between the pedestrian's turning trajectory and the current landmark turning feature trajectory, the location of the pedestrian's turning termination point and heading information are updated.

[0063] Optionally, step 3 further includes:

[0064] A dual-threshold binarization peak-valley detection algorithm is used to obtain the number of steps and an LSTM-based behavior recognition algorithm is used to dynamically estimate the step size.

[0065] like The pedestrian's position is constantly detected as the turning termination point, with coordinates as follows: The heading is Matching the location of landmark branch feature points on the turning feature trajectory and heading The formulas for updating pedestrian position and heading are then expressed as:

[0066] (11)

[0067] In the formula, for Pedestrian stride length at all times and These are the updated x and y coordinates of the pedestrian's location.

[0068] The beneficial effects of this invention are:

[0069] By constructing landmarks using multi-dimensional information perception, the credibility of landmarks is enhanced. Furthermore, by adaptively extracting landmarks with different spatial features, the method can be applied to various landmarks, such as unconventional intersections. Moreover, this application uses different thresholds to define the radius of influence of different landmarks, solving the localization system mismatch problem caused by traditional single-location landmark detection and improving positioning accuracy to some extent. Additionally, by incorporating heading weights into trajectory similarity metrics, this application achieves dual correction of the position and heading of moving targets, further improving positioning accuracy. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 Define a landmark map for multidimensional information perception.

[0072] Figure 2This is a flowchart of the method of the present invention (PDR positioning algorithm based on Multi-dimensional Information Perception Landmark Matching, MIPLM-PDR).

[0073] Figure 3A A schematic diagram for constructing the landmark center point when the number of elements in the landmark branch feature point set is equal to 2;

[0074] Figure 3B A schematic diagram for constructing the center point of a landmark when the number of feature point set elements of a landmark branch is greater than 2.

[0075] Figure 4 This is a schematic diagram of the experimental scenario and crowdsourcing trajectory scheme.

[0076] Figure 5 shows the probability distribution of the horizontal and vertical coordinates of the elements in the set of turning termination points and the simulation diagram of the fitted Gaussian function.

[0077] Figure 6 shows the relationship between neighborhood radius and local connectivity clustering accuracy.

[0078] Figure 7 shows the experimental route.

[0079] Figure 8 This is a comparison chart of the walking distance of the method in this application and two existing methods.

[0080] Figure 9A This is a comparison chart of the cumulative positioning error function of the method in this application and two existing methods under the condition of a known starting point.

[0081] Figure 9B This is a comparison chart of the cumulative positioning error function of the method in this application and two existing methods under the condition of an unknown starting point. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0083] Example 1:

[0084] This embodiment provides a landmark matching PDR positioning method in a multi-intersection scenario. See [link to relevant documentation] Figure 2 The method includes:

[0085] Step 1: Construct a system model;

[0086] Assuming a two-dimensional indoor environment, containing One intersection (landmark): It can be detected One AP base station. Landmark. ,in, Indicates the first Central location information of each landmark Indicates the first The radius of influence of a landmark refers to the range within which the landmark can collect surrounding trajectories. Indicates landmark The Location of each branch feature point and heading , and Branch road feature points refer to the starting points of each branch road at the intersection; Indicates landmark WiFi fingerprint feature, soon to be a landmark The mean RSS vector of each AP is collected at the turning termination point and used as a feature variable for landmark range detection; express Within the time period Inertial sensors collect raw data at all times Set, containing the original data Indicates the first The data collected by the inertial sensor at all times On-axis acceleration and angular velocity .

[0087] Step 2: Construct multi-dimensional information-sensing landmarks;

[0088] Landmarks are typically prominent objects and locations in an environment, facilitating spatial cognition and navigation applications. Based on an indoor multi-intersection environment, the proposed multi-dimensional information-sensing landmark not only describes specific locations within the indoor space but also characterizes the spatial structure of the landmark and effectively integrates WiFi fingerprint information.

[0089] like Figure 1 As shown, landmarks typically include conventional and unconventional landmarks. Conventional landmarks, such as L-shaped, T-shaped, and cross-shaped landmarks, have a headway difference between adjacent branch roads. ( (For positive integers), and for unconventional landmarks such as Y-shaped intersections, the difference in heading between adjacent branch roads exists. Non-integer multiples. Each landmark is determined by its center point. Landmark branch road feature point set Radius of the landmark's influence and landmark WiFi feature fingerprint Definition. This representation method avoids the single location definition of landmarks, while being able to adaptively extract structural information from both conventional and unconventional landmarks.

[0090] Step 2.1, Construct a set of landmark branch feature points

[0091] Based on crowdsourced trajectory data, when a pedestrian makes a turning action at a landmark, they will leave at least one trajectory point. In order to extract the landmark branch feature points in the crowdsourced trajectory, the turning termination point is defined and given by Definition 1.

[0092] Definition 1: If pedestrians At any given time, a turning action occurs within a certain landmark area, and... If the pedestrian changes from turning to going straight at a certain moment, then the trajectory formed by the turning point is represented as follows: Turning point This refers to a turning point within the area defined by the landmark.

[0093] By using inertial sensor data collected from pedestrians' smartphones (including accelerometers and gyroscopes) to identify their behavior at each moment, the system can effectively extract pedestrian turning termination points. However, considering the temporal nature of inertial sensor sampling during pedestrian movement, and the fact that traditional machine learning classification algorithms ignore the dynamic trends hidden in time-series samples and fail to uncover the cross-information between different features, while also relying heavily on prior knowledge and requiring time and effort to extract scene features, this paper proposes a solution. To balance accuracy and efficiency in behavior recognition, a Long Short-Term Memory (LSTM) model is used to solve the temporal classification problem, achieving high accuracy even with near-raw data.

[0094] Define a behavior label for a given behavior state. ,for consecutive integers within the interval, where This represents the number of all possible behavioral state categories. To avoid errors in feature calculation caused by a single continuous numerical representation, [the following is omitted as it is not explicitly stated in the original text]. Perform one-hot vector encoding, using a length of The bit register encodes each behavior category, with each category value corresponding to an independent register bit. At any given time, only one register bit is 1, while the rest are 0, resulting in a behavior label vector. Regarding the current moment... Before the sensor samples Continuous data at different times constitute observation data LSTM provides Time-of-action label vector and through The function extracts the index of the maximum value in a vector to obtain the label of the current row, as shown in the following formula:

[0095] (1)

[0096] in, Indicates going straight. Indicates a change of direction.

[0097] Based on the characteristic that pedestrian trajectory points follow a Gaussian distribution on the corridor, the location of the crowdsourced trajectory turning and ending point, heading, and... The RSS vectors of each AP together constitute the sample set. Gaussian mixture clustering algorithm was used to obtain Clusters of turning termination points And the cluster center point location. Its corresponding heading information Together they form the cluster center point set of turning termination points ,in , The azimuth angle of a pedestrian in the horizontal direction is obtained using the quaternion method.

[0098] It should be noted that crowdsourced trajectory refers to trajectory data voluntarily uploaded by users to online platforms, typically generated through applications on smart mobile devices.

[0099] For the specific process of obtaining the horizontal azimuth angle of a pedestrian using the quaternion method, please refer to Zhou Xiaoren and Zhao Heming. Initial alignment algorithm for strapdown inertial navigation based on quaternion attitude estimation [J]. Journal of Sensor Technology, 2021, 34(09): 1182-1188.

[0100] Generally, cluster centers belonging to the same landmark turning point will cluster within that landmark area. If a turning point turning point cluster center is... In the neighborhood radius There is another point within the range. and Connected, then for The set of connected points Inner element. Also, if This indicates Inner elements and Indirect connection, all with The set of points that are directly or indirectly connected This represents a locally connected space, and all locally connected spaces constitute a set of landmark branch feature points with sequence identifiers. .

[0101] Step 2.2, Construct the landmark center point

[0102] Landmark Center As a physical point that truly exists in physical space, it satisfies the uniqueness of its physical spatial location, and its corresponding set of landmark branch feature points. Connections form the spatial structure of a landmark. Given a landmark... Landmark branch feature point set Landmark Center It can be determined by landmark branch road feature points The intersection point of its reverse extension with its heading indicates that, disregarding errors, they should intersect at a single point. The calculation formula is as follows:

[0103] (2)

[0104] Obviously, when At times, such as Figure 3A As shown, the landmark contains only two branch feature points, whose backward extensions intersect at a single point. There is a unique solution; when At times, such as Figure 3B As shown, the landmark contains multiple branch feature points. Due to spatial continuity and the introduction of measurement errors, the intersection of its backward extensions is not unique, i.e., equation (2) is an overdetermined system of equations with multiple solutions. In this case, the least squares solution of the system of equations is used to represent the actual physical point. The approximate solution is, The solution.

[0105] Step 2.3: Construct the radius of influence of the landmark and its WiFi feature fingerprint.

[0106] Considering the differences in the spatial structure of different landmarks, a single threshold cannot be used to define the radius of influence of a landmark. Based on the distribution characteristics of pedestrian turning termination points, all possible turning termination point locations should be included within the influence range of the landmark. Given a landmark... , Clustering centers based on turning termination points and with The set of points that are directly or indirectly connected is defined as follows: The set of all clusters to which the cluster centers of all turning termination points belong is ,but It is expressed as follows:

[0107] (3)

[0108] In the formula, At the same time, we can obtain for:

[0109] (4)

[0110] In the formula, for Number of elements in the set for Collect data at each point within the set. RSS vectors of APs.

[0111] Step 3: Landmark matching and PDR positioning;

[0112] Traditional landmark-assisted PDR (Progressive Directional Redirection) location updates typically match the current location with landmark points to adjust the current positioning status, which is effective when landmarks have strong geographical distinguishability. This application proposes a multi-dimensional information-based landmark perception method that uses landmark branch feature points to perform dual corrections on the pedestrian turning termination point position and heading, and uses the corrected turning termination point as the starting point for PDR positioning, intermittently eliminating accumulated PDR errors. However, because landmark branch feature points are closely adjacent and geographically ambiguous, single location information can easily cause localization mismatch problems.

[0113] To address the limitations of traditional landmark matching algorithms in specific scenarios and their strong dependence on location information, landmark area detection replaces specific point location detection in the landmark detection stage, adapting to landmarks with different spatial structures. In the landmark matching and localization stage, a turning feature trajectory geometric extraction method is used to construct turning feature trajectories based on landmark branch feature points and landmark center points. A similarity metric method with heading constraints is then used to determine the similarity between the pedestrian's current turning trajectory and the turning feature trajectory, thereby updating the position and heading of the pedestrian's turning termination point.

[0114] (1) Adaptive Landmark Detection Algorithm

[0115] In WiFi fingerprint-based localization, Euclidean distance is generally used to measure the similarity between online data and offline fingerprints. However, Euclidean distance does not consider the differences in the reliability of the distance represented by the AP signal in each dimension of the RSS vector. Therefore, the Euclidean distance can be corrected by weighting the AP signal. Similarly, given a landmark... Due to the instability of AP signals, the steering termination point is extracted during the offline phase. and Similarity metrics can be affected by outliers. In this case, by... and The Euclidean distances are weighted to obtain the landmarks. similarity measure mean :

[0116] (5)

[0117] Due to the radius of influence of the landmark The difference is used to determine the landmark matching stage. Pedestrian turning termination point and The Euclidean distance threshold is variable. Considering the logarithmic relationship between RSS value and physical distance, an adaptive threshold is defined. for:

[0118] (6)

[0119] when At that time, it was detected that the pedestrian's current turning termination point was located at the landmark. This allows for the correction of its position and course.

[0120] (2) Geometric extraction of turning feature trajectory

[0121] The turning characteristic trajectory represents the set of turning points left by a pedestrian as they pass through the approaching and exiting side roads within the landmark area. Assuming... and as a landmark Any two landmark branch feature points within the same road segment. Specifically, for landmark branch feature points of different branches within the same road segment. and There is no turning behavior, meaning there is no turning characteristic trajectory. Therefore, the road segment decision formula is defined as follows:

[0122] (7)

[0123] In the formula, and They are respectively and The course, if If the two landmark branch feature points are located in different road segments, then the landmarks can be extracted. Set of turning feature trajectories within the range Including the turning characteristic trajectory of the same branch road and different branch turning characteristic trajectories ,in The turning feature point is given by Definition 2. express A unique identifier within the set.

[0124] Definition 2: Given a landmark If it exists Then in Inside On the bisector of the line, there must exist a point. and the circle centered at that point and the line segment and The shorter one is tangent to , At the same time with Hand over , That is, the landmark A turning feature point.

[0125] like , , for Vector angle, satisfying ,but and The location can be calculated using equation (8).

[0126] (8)

[0127] when hour, It conforms to the actual turning characteristic trajectory trend.

[0128] (3) Heading-constrained trajectory similarity matching

[0129] The similarity of trajectories should not only characterize the physical spatial differences between them, but also take into account the directional trend of the trajectories. Calculating trajectories using only Euclidean distance can easily lead to incorrect similarity determinations. Therefore, this application incorporates directional weighting into the trajectory similarity metric, using the pedestrian turning termination point time... Guided by this, and then assess pedestrians. Turning trajectory at all times and detected landmarks Turning characteristic trajectory within range Similarity results between The calculation method is as follows:

[0130] (9)

[0131] In the formula, The Frescher distance between two trajectories is the minimum value of the elements in the set of the maximum Euclidean distance between the sequence pairs of two discrete trajectory points. Representing two trajectories and The heading difference weight is set as follows:

[0132] (10)

[0133] Therefore, by judging the similarity between the pedestrian's turning trajectory and the current landmark turning feature trajectory, the location of the pedestrian's turning termination point and heading information are updated.

[0134] For the specific calculation method of Fréchet distance, please refer to Bringmann K, Mulzer W. Approximability of the discrete Fréchet distance[J]. Journal of Computational Geometry, 2016,7(2): 46–76.

[0135] (4) Position and heading correction

[0136] PDR (Pedestrian Position Detection) estimates pedestrian position by recording the heading and velocity from a known starting location, providing continuous two-dimensional positioning information. Since PDR relies on the physiological characteristics of a person during the walking process, it can be used to estimate the number of steps and stride length, and thus the pedestrian's position. This application employs a dual-threshold binarization peak-valley detection algorithm to obtain the number of steps and LSTM-based behavior recognition to dynamically estimate the stride length. The pedestrian's position is constantly detected as the turning termination point, with coordinates as follows: The heading is Matching the location of landmark branch feature points on the turning feature trajectory and heading The formulas for updating pedestrian position and heading can be expressed as:

[0137] (11)

[0138] In the formula, for Pedestrian stride length at all times and These are the updated x and y coordinates of the pedestrian's location.

[0139] To evaluate the effectiveness of the method, the test scenario was selected as the corridor area within a double-ring floor plan. The internal corridor was 2m wide and approximately 180m long. Figure 4 As shown, the experimental area includes landmarks. That is, it includes 6 landmark center points 14 landmark branch road feature points. WiFi routers evenly deployed by the operator were selected as AP signal sources, respectively. .

[0140] The experiment collected data from four users (three men and one woman), with heights ranging from 158cm to 180cm, and trained a stride model for each user. To avoid the influence of phone differences on the results, all experimental data were collected using the same phone model at a sampling frequency of 5Hz. The experimenters placed the phone flat in front of them (face up) and... Figure 7 Starting from the indicated point, the path was traversed 80 times in different directions, resulting in a total of 640 trajectories. To ensure the reliability of the trajectories and reduce the impact of accumulated errors, each sampling round lasted 1 minute.

[0141] To verify that the turning termination points in the experimental scenario conform to a Gaussian distribution without loss of generality, trajectories on four different landmark branches were randomly selected, and a sample set of turning termination points was extracted. Gaussian fitting was performed on the distribution of the horizontal and vertical coordinates of the location information within the set. For example... Figure 5As shown, the sample data generally tends to follow a Gaussian distribution, which meets the application conditions of the method in this application. Based on the probability-first data processing method, this scenario sets a confidence interval of 0.95. Turning termination points with a confidence level lower than the lower limit of the confidence interval are considered outliers and removed, while turning termination points with a confidence level higher than the lower limit of the confidence interval are retained for landmark construction to achieve the purpose of filtering out biased data.

[0142] To optimize the performance of the positioning system, the neighborhood radius is... Perform testing and optimization. Domain radius. The settings have a significant impact on the identification of landmark spatial structures. Therefore, when constructing landmark branch feature point sets through local connectivity clustering, the clustering accuracy can effectively reflect... The applicability of the values. Experiments were conducted in... The values ​​were taken at intervals of 0.5 to obtain the clustering accuracy and... Relationship curves as follows Figure 6 As shown, based on the fitting of experimental data, when The range of values ​​is within Within the time frame, the clustering accuracy is 1, accurately identifying all landmark branch feature points within each landmark's area. Considering the special case of adjacent landmarks being close together, The value should not be too large, therefore the experiment was set to... .

[0143] Considering that the MIPLM-PDR method proposed in this application aims to construct highly reliable landmarks and landmark matching methods, thereby eliminating PDR cumulative errors, the localization performance of the MIPLM-PDR algorithm is compared with that of LRVCT and ALIMC, which are also based on crowdsourced trajectories for landmark construction, in the experiment.

[0144] For the LRVCT method, please refer to Sun G, Zhao J, Zhu C. A Novel Approach for LandmarkRecognition via Crowdsourced Trajectory[C]. Proceedings of the 3rdInternational Conference on Computer Science and Application Engineering.Sanya, China, 2019.

[0145] For the ALIMC method, please refer to Zhou B, Li Q, Mao Q, et al. ALIMC: Activity Landmark-Based Indoor Mapping via Crowdsourcing[J]. IEEE Transactions on IntelligentTransportation System, 2015, 16(5):2774-2785.).

[0146] These two comparison methods provide solutions for optimizing Wi-Fi landmark credibility and active landmark matching, respectively, and are representative to some extent.

[0147] To ensure fairness in the comparative experiments, it is assumed that all turning activities and Wi-Fi landmark construction are based at intersections or corners, and that all three methods are calibrated and positioned within the same PDR framework. The experimental route map is as follows: Figure 7 As shown.

[0148] Convergence speed is a crucial issue for landmark-based localization algorithms. It is defined as the distance the algorithm travels before converging to the ground truth point. Faster convergence means that successfully matched landmarks can be used more frequently to correct accumulated errors, which is essential for improving localization accuracy. Figure 8 The distance traveled by the proposed method and two comparative methods before algorithm convergence is given. This indicates that the algorithm cannot converge, by Figure 8 It is known that when the starting point is unknown, the initial error is large. Due to the accumulation of errors, the ALIMC method based on AMM cannot match pedestrian behavior with landmarks, and the corresponding algorithm cannot converge. However, the corresponding algorithms of LRVCT and the MIPLM-PDR method proposed in this application can converge quickly in the case of unknown and known starting points, and can be effectively applied to the elimination of accumulated errors in PDR.

[0149] Figure 9A and Figure 9BThe cumulative error probability distributions of the three methods are presented under known and unknown starting point conditions. It can be seen that when the starting point is known, the cumulative error probability of the MIPLM-PDR method proposed in this application is 68.6% when the positioning error is within 1m, and reaches 92.1% when the error is within 2m. This indicates that the average positioning error of the MIPLM-PDR method in this scenario can be controlled within 2m, demonstrating good positioning stability. In contrast, ALIMC's cumulative error probability is only 63.9% when the positioning error is less than 2m, which is 28.2% lower than that of the MIPLM-PDR method. The cumulative error probability curve of LRVCT consistently falls between the two. When the starting point is unknown, due to the large initial error, ALIMC cannot converge, and the error is generally above 2m. In contrast, LRVCT and the MIPLM-PDR method still maintain a cumulative error probability above 50% within 2m, exhibiting good positioning performance.

[0150] Tables 1 and 2 present the position estimation error analysis for the three methods.

[0151] Table 1. Location estimation error (given starting point)

[0152]

[0153] Table 2. Location estimation error (unknown starting point)

[0154]

[0155] As shown in Tables 1 and 2, when the starting point is known, the minimum system error is generated by the original PDR, and all three are consistent. The MIPLM-PDR method of this application outperforms LRVCT and ALIMC in terms of both average and maximum error. In the entire positioning environment, the average error of MIPLM-PDR is concentrated within a relatively small error range [0.011, 2.102], with error fluctuations controlled within 2.2m, indicating that its positioning stability is superior to the other two. When the starting point is unknown, the ALIMC algorithm cannot converge, the accumulated error cannot be eliminated, and the positioning error will increase. For LRVCT and the MIPLM-PDR algorithm of this application, the maximum error in this experimental scenario arises from the accumulation of the original error before algorithm convergence. Furthermore, because the MIPLM-PDR method of this application not only corrects the pedestrian position but also corrects the heading information, its average and minimum errors are reduced by 20.7% and 20.1% respectively compared to LRVCT. Experiments have shown that the MIPLM-PDR method proposed in this application can more effectively eliminate PDR cumulative errors in environments with multiple intersections, and the improvement in positioning results is comprehensive.

[0156] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0157] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A landmark matching PDR positioning method in a multi-intersection scenario, characterized in that, The method includes: Step 1: Build a system model for the multi-intersection scenario to be located; Step 2: Obtain multi-dimensional information of each landmark in the multi-intersection scene to be located to form the multi-dimensional features of the landmark, including the center location information of the landmark, the radius of influence, information of all branch roads of the landmark, and Wi-Fi feature fingerprint; Step 3: PDR positioning is performed based on the multidimensional features of the landmark. During the positioning process, the heading weight is incorporated into the trajectory similarity metric to achieve dual correction of the position and heading of the moving target. The moving target carries an inertial sensor. Step 1 includes: Assuming the multi-intersection scenario to be located includes A landmark, ; detectable in the described scenario Each AP base station, then each landmark ; in, Indicates the first Central location information of each landmark Indicates the first The radius of influence of each landmark; Indicates landmark The Location of each branch feature point and heading , and ; Indicates landmark Wi-Fi signature fingerprints will soon be used in landmarks. The mean RSS vector of each AP base station collected at the turning termination point is used as a feature variable when detecting the landmark range; express Within the time period Inertial sensors collect raw data at all times Set, containing the original data Indicates the first The data collected by the inertial sensor at all times On-axis acceleration and angular velocity ; Step 2, obtaining information on all branch roads of a landmark, includes: Define the turning endpoint: if the moving target At any given time, a turning action occurs within a certain landmark area, and... If the pedestrian changes from turning to going straight at a certain moment, then the trajectory formed by the turning point is represented as follows: Turning point This refers to a turning point within the landmark area; Behavioral data is collected using inertial sensors carried by the moving target. The behavioral state of the moving target at each moment is identified based on the behavioral data. The behavioral state includes straight-line movement and turning. The behavioral data includes acceleration information and angular velocity information. Determine the location of each turning termination point based on the behavior state of the moving target; Data was collected at each steering endpoint. The RSS vector of each AP; Based on the position of the turning termination point, heading, and collected data The RSS vectors of each AP together constitute the sample set. The heading is calculated based on the angular velocity information in the behavior data; Gaussian mixture clustering algorithm is used to obtain Clusters of turning termination points ; Location of the cluster center Its corresponding heading information Together they form the cluster center point set of turning termination points ,in , Representing the cluster center points respectively Axis coordinates; If the turning point terminates at the cluster center point In the neighborhood radius There is another point within the range. and Connected, then for The set of connected points Inner element; at the same time, if This indicates Inner elements and Indirect connection, all with The set of points that are directly or indirectly connected This represents a locally connected space, and all locally connected spaces constitute a set of landmark branch feature points with sequence labels. .

2. The method according to claim 1, characterized in that, The step of identifying the behavioral state of the moving target at each moment based on the behavioral data includes: Define behavior labels ,for consecutive integers within the interval, where The number of all possible behavior state categories, including going straight and turning; Regarding the current moment Sampled by inertial sensor Before the moment Continuous data at each moment constitutes the observation data. ; Using the Long Short-Term Memory (LSTM) model to give Time-based behavior label vector and through The function extracts the index of the maximum value in a vector to obtain the label of the current row, as shown in the following formula: (1) in, Indicates going straight. Indicates a change of direction.

3. The method according to claim 2, characterized in that, Step 2, obtaining the center location information of the landmark, includes: Targeting landmarks Let its landmark branch feature point set be set. The center point of the landmark Landmark branch road feature points Intersection with its reverse extension line The calculation formula is as follows: (2) when At that time, the landmark only contains two branch feature points, whose backward extensions intersect at one point. There is a unique solution; when When a landmark contains multiple branch feature points, and the intersection of their backward extensions is not unique, the real physical point is represented by the least squares solution of the system of equations. The approximate solution is, The solution.

4. The method according to claim 3, characterized in that, Step 2, obtaining the radius of influence and Wi-Fi feature fingerprint, includes: For different landmarks, the radius of influence of the landmark is defined using different thresholds. : Targeting landmarks ,definition The set of all clusters to which the cluster centers of all turning termination points belong is ,but It is expressed as follows: (3) In the formula, ; Correspondingly, for: (4) In the formula, for Number of elements in the set for Collect data at each point within the set. RSS vectors of APs.

5. The method according to claim 4, characterized in that, Step 3 includes: Define adaptive threshold for: (6) in, as a landmark The mean of the similarity measure, This indicates the turning termination point determined in real time during the matching phase. The calculation formula is as follows: (5) when At that time, it was detected that the pedestrian's current turning termination point was located at the landmark. Location, and then based on landmarks Correct its position and course.

6. The method according to claim 5, characterized in that, Step 3 also includes extracting the steering feature trajectory: Assumption and as a landmark Any two landmark branch feature points within the area; The formula for determining road segments is defined as follows: (7) In the formula, and They are respectively and The course, if If the two landmark branch feature points are located in different road segments, then the landmarks can be extracted. Set of turning feature trajectories within the range Including the turning characteristic trajectory of the same branch road and different branch turning characteristic trajectories ,in For turning feature points, express Unique identifier within the set; Turning feature points are defined as: given landmarks If it exists Then in Inside On the bisector of the line, there must exist a point. and the circle centered at that point and the line segment and The shorter one is tangent to , At the same time with Hand over , That is, the landmark A turning feature point; like , , for Vector angle, satisfying ,but and The position can be calculated using equation (8): (8) when hour, It conforms to the actual turning characteristic trajectory trend.

7. The method according to claim 6, characterized in that, Step 3 also includes: Incorporating heading weights into the trajectory similarity metric, based on the pedestrian's turning termination point time. Guided by this, and then assess pedestrians. Turning trajectory at all times and detected landmarks Turning characteristic trajectory within range Similarity results between The calculation method is as follows: (9) In the formula, The Frescher distance between two trajectories is the minimum value of the elements in the set of the maximum Euclidean distance between the sequence pairs of points in the two discrete trajectories. Representing two trajectories and Heading difference weights: (10) Therefore, by judging the similarity between the pedestrian's turning trajectory and the current landmark turning feature trajectory, the location of the pedestrian's turning termination point and heading information are updated.

8. The method according to claim 7, characterized in that, Step 3 also includes: A dual-threshold binarization peak-valley detection algorithm is used to obtain the number of steps and an LSTM-based behavior recognition algorithm is used to dynamically estimate the step size. like The pedestrian's position is constantly detected as the turning termination point, with coordinates as follows: The heading is Matching the location of landmark branch feature points on the turning feature trajectory and heading The formulas for updating pedestrian position and heading are then expressed as: (11) In the formula, for Pedestrian stride length at all times and These are the updated x and y coordinates of the pedestrian's location.

Images