An unsupervised indoor positioning method based on crowd-sourced trajectory data

Abstract

The application discloses a kind of crowd-sourced trajectory data-based unsupervised indoor positioning method, utilize the crowd-sourced trajectory of pedestrian, reduce the dependence of positioning method on a large number of prior knowledge, so as to reduce the deployment and use cost of positioning method.Compared with the traditional unsupervised indoor positioning algorithm: the preprocessing of offline stage and the angle matching of online stage, reduce the real-time positioning time of positioning method and improve the positioning accuracy;BSF used in online stage, without considering the linearity and nonlinearity of system and the Gaussian and non-Gaussian of noise, so that the method of the application is closer to the real posterior distribution.Due to the improvement of these two aspects, the positioning accuracy of the indoor positioning method disclosed in the application is higher than that of the traditional unsupervised indoor positioning system and method.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, and in particular to an unsupervised indoor positioning method based on crowdsourced trajectory data. Background Technology

[0002] With the rapid development of location-based services (LBS), fire emergency rescue, and autonomous robot navigation applications in indoor environments, the demand for high-precision, low-cost, and easy-to-maintain indoor positioning services has become increasingly strong. Over the past decade, numerous indoor positioning technologies based on WiFi, Radio Frequency Identification (RFID), Ultra-Wideband (UWB), magnetic fields, and Bluetooth have developed rapidly. Compared to geometric-based wireless signal positioning technologies, location fingerprint-based wireless signal positioning technologies can achieve positioning without requiring specific access points (APs) or beacons, device transmission power, or other information, offering a simpler implementation and greater flexibility. Therefore, location fingerprint-based wireless signal fingerprint positioning technology is gaining popularity. However, the most important and challenging aspect of wireless signal fingerprint positioning is the training and maintenance of the fingerprint map. Currently, building signal fingerprint maps requires professional surveying and calibration, and continuous updates are needed to maintain the positioning performance, resulting in high time and labor costs and reducing the universality of signal fingerprint positioning technology.

[0003] Inertial sensor-based positioning technologies offer advantages such as the widespread availability of sensor integration and the simplicity of algorithm implementation. Among these, the mainstream algorithm is Pedestrian Track Deduction (PDR), which calculates position at a relatively low cost by analyzing changes in physical motion. However, due to noise in inertial sensor data, position deduction often suffers from accumulated errors, leading to track drift. Therefore, many studies combine PDR algorithms with additional positioning information to achieve more accurate positioning. However, the acquisition and processing costs of this additional positioning information are often high, reducing the applicability of such methods in real-world environments.

[0004] The rapid proliferation of smart portable devices, the integration of sensors, and the popularity of crowdsourced data collection methods have made WiFi and inertial measurement unit (IMU) data acquisition convenient, thus enabling the reduction of costs associated with acquiring additional location information and building and maintaining signal fingerprint maps. Furthermore, by integrating additional WiFi sensor data, crowdsourced trajectory data based on the PDR algorithm can reduce trajectory deviations during the online positioning phase. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by fully integrating WiFi sensing information and IMU sensor data, and to propose an unsupervised indoor positioning method based on crowdsourced trajectory data.

[0006] The technical solution adopted by this invention to solve its technical problem includes: offline stage for collecting and processing crowdsourced trajectory data and online stage for real-time tracking of the target.

[0007] I. Offline Phase: Specific steps include:

[0008] 1-1. In an environment with an access point (AP), a pedestrian holds a mobile phone equipped with WiFi and IMU sensors and collects WiFi received signal strength (RSS) data in the area of ​​interest. As the pedestrian moves, the phone collects a crowdsourced trajectory data line containing both RSS and IMU data. Ultimately, crowdsourced trajectory data covering all areas is obtained.

[0009] 1-2. The obtained RSS+IMU crowdsourced trajectory data needs to be segmented according to the pedestrian's stride length. Under the processing of the PDR algorithm, all the collected RSS+IMU crowdsourced trajectory data were segmented into the following triplets. in and It is the RSS sequence measured at the offline time interval t. It is a motion vector calculated from IMU data.

[0010] 1-3. Because the motion vectors in the segmented crowdsourced trajectory data contain a lot of noise, the data is preprocessed to reduce the impact of noise on the overall offline dataset.

[0011] II. The online phase includes the following specific steps:

[0012] 2-1. During the pedestrian's walking process, the RSS+IMU crowdsourced trajectory data acquired and collected by the mobile phone is processed by the PDR algorithm to obtain the {z} of the online phase. t-1 ,z t u t-1} data, where z t-1 and z t It is the RSS sequence measured at the online phase time interval t, u t-1 It is a motion vector calculated from IMU data.

[0013] 2-2. z obtained from the mobile phone sensor t-1 ,z t The sequence is fed into the offline data obtained in steps 1-3, and the KNN algorithm is used for feature matching and angle matching to obtain the observed motion vector. Then combine the u obtained during the online phase t-1The data is fed into a bootstrap filter (BSF) to estimate the current user's position.

[0014] The beneficial effects of this invention are as follows:

[0015] This invention provides an unsupervised indoor positioning method based on crowdsourced trajectory data. Essentially, it's an unsupervised indoor positioning technique based on PDR (Pedestrian Direct Recognition) and filtering algorithms. It utilizes crowdsourced pedestrian trajectories to reduce the reliance on extensive prior knowledge, thereby lowering deployment and usage costs. Compared to traditional unsupervised indoor positioning algorithms: offline preprocessing and online angle matching reduce real-time positioning time and improve positioning accuracy; the BSF (Blog-Based Flow Analysis) used in the online phase eliminates the need to consider system linearity / nonlinearity or noise Gaussian / non-Gaussianity, allowing our method to more closely approximate the true posterior distribution. These improvements result in a higher positioning accuracy than traditional unsupervised indoor positioning systems and methods. Attached Figure Description

[0016] Figure 1 This is a diagram of the positioning framework proposed in this invention.

[0017] Figure 2 This invention presents a simulated environment for acquiring data using mobile phone sensors.

[0018] Figure 3 To calculate u in this invention t-1 A schematic diagram. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention. Figure 1 This is a flowchart illustrating the steps of an unsupervised indoor positioning method based on bootstrap filtering proposed in this invention.

[0020] like Figure 1 As shown, the specific implementation steps of the unsupervised indoor localization method based on bootstrap filtering are as follows:

[0021] Step 1: In an environment with an access point (AP), a pedestrian holds a mobile phone equipped with WiFi and IMU sensors and collects WiFi received signal strength (RSS) data in the area of ​​interest. As the pedestrian moves, the phone collects a crowdsourced trajectory data line based on RSS and IMU signals. Ultimately, crowdsourced trajectory data covering all areas is obtained.

[0022] exist Figure 2In the environment shown, the pedestrian's IMU sensor consists of three parts: an accelerometer, a gyroscope, and a magnetometer. As the pedestrian moves, these three sensors record data about the physical effects at a fixed frequency.

[0023] Step 2: The RSS+IMU crowdsourced trajectory data obtained in Step 1 needs to be segmented according to the pedestrian's stride length. Under the processing of the PDR algorithm, all collected RSS+IMU crowdsourced trajectory data were segmented into the following triplets.

[0024] Based on step 1, data from the accelerometer, gyroscope, and magnetometer were obtained. Using the accelerometer data, the PDR algorithm calculates the step size of the movement; using the gyroscope and magnetometer data, the PDR algorithm calculates the pedestrian's turning angle. Figure 3 As shown, the offline phase It is derived from the following formula:

[0025]

[0026] Where, position t and position t-1 It represents the pedestrian's position before and after the time interval t;

[0027] Step 3: In the offline stage, the IMU sensor in the mobile phone is easily affected by the environment, and the collected data contains a lot of noise. It is necessary to preprocess the motion vectors after step 2 to reduce the impact of noise on the overall offline dataset.

[0028] Step 3.1, dimensionality reduction.

[0029] High-dimensional RSS data can interfere with real-time smartphone operation. Furthermore, the RSS data in the segmented crowdsourced trajectory data from step 2 contains many meaningless values ​​(set to -100dBm if no RSS value is captured), and each access point (AP) cannot cover the entire indoor area. Therefore, feature extraction from RSS data is crucial for localization. An autoencoder (AE) is used to process the RSS sequence. Feature dimensionality reduction is performed to provide effective feature data for subsequent clustering and smoothing processes. The dimensionality-reduced RSS sequence Z is obtained through an autoencoder. h .

[0030]

[0031] Step 3.2, Clustering Processing. Clearly, WiFi signal strength is similar within a certain physical area. Traditional fingerprint positioning widely employs clustering methods to reduce computation and improve positioning accuracy. In the clustering process, the K-means clustering algorithm is used to cluster the RSS sequences. After the RSS sequences are clustered using Kmeans, the clustered RSS sequences are obtained. RSS sequences are clustered based on their similarity. The triplet is divided into several clusters, which significantly reduces the search space of the K-Nearest Neighbors (KNN) algorithm in the online stage. Simultaneously, the reduced data dimensionality further improves the computational efficiency of K-means. Furthermore, with specific class sizes and fewer feature dimensions, the KNN-based smoothing method can greatly improve computational efficiency.

[0032]

[0033] Step 3.3, smoothing process.

[0034] The motion vectors segmented from IMU data by the PDR algorithm contain significant noise. The smoothing process... Using parameters, KNN is used to extract the top-K similar triples, and the average value is used for smoothing to reduce noise in the motion vectors. Then, an offline dataset is constructed. Since a vector is a vector, it is inappropriate to simply add them together without considering their direction and angle. Therefore, an angle threshold of α is set to achieve angle matching, filtering the angles of the top-K motion vectors based on the previous result. As a matching parameter These are the top-K motion vectors, calculated according to the following formula. and Angle between:

[0035]

[0036] According to formula (4), retain the included angle α. i <α Build an offline dataset.

[0037] Here, α needs to be found to a suitable value, because values ​​that are too large or too small will affect the final positioning accuracy of the model. The setting of α depends on the degree of noise angle offset in the crowdsourced trajectory data. The setting method is as follows: when motion vectors covering all areas are collected in the offline phase, the offline motion vector can be obtained. The deviation from the true motion vector. When a large number of motion deviations are obtained, an angle deviation dataset is obtained. The average deviation of the angle deviation dataset is used as the angle threshold α.

[0038] Step 4: During the pedestrian's walking process, obtain RSS+IMU crowdsourced trajectory data collected by the mobile phone, process it through the PDR algorithm, and obtain the {z} for the online stage. t-1 ,z t u t-1}data.

[0039] Step 5, take the z obtained in step 4 t-1 ,z t Using the offline data obtained in step 3 as a parameter, the observed motion vector is obtained through angle matching. Then combine u t-1 The data is fed into the bootstrap filter (BSF) to estimate the current user's position.

[0040] Specifically, the state transition formula is as follows:

[0041]

[0042] in, For the state position, P t-1 It is the location of the pedestrian at the previous moment.

[0043] The measurement estimation formula is as follows:

[0044]

[0045] in, For measuring position; To observe the motion vector, According to the z measured by the mobile phone t-1 ,z t Using the offline dataset obtained in step 3 as parameters, feature matching and angle matching filtering are performed through the KNN algorithm. The subsequent bootstrap filtering is as follows:

[0046] 5-1. Initial state: Initialize a fixed number of particles so that the particles are evenly distributed in space;

[0047] 5-2. In the prediction stage, according to the state transition formula (5), each particle obtains a predicted particle;

[0048] 5-3. Correction stage: Use the measurement estimation formula (6) to obtain the measurement position and evaluate the predicted particles. The closer the particle is to the true state, the greater its weight. The closer the predicted particle is to the measurement position, the greater its weight, and vice versa. Here, the right side of the Gaussian distribution bell curve is used as the particle weight.

[0049] 5-4. Resampling: Filter particles according to their weights, retaining a large number of high-weight particles and a portion of low-weight particles. Estimate the current pedestrian's position based on the sampled predicted particles.

[0050] 5-5. Filtering: Substitute the resampled particles into the state transition equation to obtain new predicted particles. Then, perform correction and resampling repeatedly in a loop to achieve online real-time positioning.

[0051] The above description, in conjunction with specific / preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. Those skilled in the art can make various substitutions or modifications to these described embodiments without departing from the inventive concept, and all such substitutions or modifications should be considered within the scope of protection of the present invention.

[0052] The parts of this invention not described in detail are well-known to those skilled in the art.

Claims

1. An unsupervised indoor positioning method based on crowdsourced trajectory data, characterized in that, include: Offline phase involves collecting and processing crowdsourced trajectory data, while online phase involves real-time tracking of the target. I. Offline Phase: Specific steps include: 1-1. In an environment with an access point (AP), a pedestrian holds a mobile phone equipped with WiFi and IMU sensors and collects WiFi signal strength in the area of ​​interest. As the pedestrian moves, the phone can collect a crowdsourced trajectory data of RSS+IMU, which includes RSS data and IMU data; finally, crowdsourced trajectory data covering all areas is obtained. 1-2. The obtained RSS+IMU crowdsourced trajectory data needs to be segmented according to the pedestrian's stride length; 1-3. Because the motion vectors in the segmented crowdsourced trajectory data contain a lot of noise, the data is preprocessed to reduce the impact of noise on the overall offline dataset; II. The online phase includes the following specific steps: 2-1. During the pedestrian's walking process, the RSS+IMU crowdsourced trajectory data acquired and collected by the mobile phone is processed by the PDR algorithm to obtain the online phase { , , } data, in which and It is the RSS sequence measured at the online phase time interval t. It is a motion vector calculated from IMU data; 2-2. Data obtained from mobile phone sensors , The sequence is fed into the offline data obtained in steps 1-3, and the KNN algorithm is used for feature matching and angle matching to obtain the observed motion vector. Then, combined with the results obtained in the online phase The position of the current user is then fed into the bootstrap filter to estimate the current user's location. The specific method for step 2-2 is as follows: The result obtained in step 2-1 , Using the offline data obtained in steps 1-3 as parameters, the observed motion vector is obtained through angle matching. Then combine The position of the current user is then estimated by feeding it into the bootstrap filter (BSF). Specifically, the state transition formula is as follows: （5） in, For the state position, It is the pedestrian's position at the previous moment; The measurement estimation formula is as follows: （6） in, For measuring position; To observe the motion vector, According to measurements taken by the mobile phone , Using the offline dataset obtained in steps 1-3 as parameters, feature matching and angle matching filtering are performed through the KNN algorithm; the subsequent bootstrap filtering is as follows: (1) Initial state: Initialize a fixed number of particles so that the particles are evenly distributed in space; (2) In the prediction stage, according to the state transition formula (5), each particle obtains a predicted particle; (3) Correction stage: Use the measurement estimation formula (6) to obtain the measurement position and evaluate the predicted particles. The closer the particle is to the true state, the greater its weight. The closer the predicted particle is to the measurement position, the greater its weight, and vice versa. The right side of the Gaussian distribution bell curve is used as the particle weight. (4) Resampling: Filter particles according to particle weights and estimate the current pedestrian's position based on the sampled predicted particles; (5) Filtering: Substitute the resampled particles into the state transition equation to obtain new predicted particles. Then perform correction and resampling in sequence, and continuously iterate to achieve online real-time positioning.

2. The unsupervised indoor positioning method based on crowdsourced trajectory data according to claim 1, characterized in that, The IMU sensor consists of three parts: an accelerometer, a gyroscope, and a magnetometer. As the pedestrian moves, these three sensors record data information under physical influence at a fixed frequency.

3. The unsupervised indoor positioning method based on crowdsourced trajectory data according to claim 2, characterized in that, The specific methods for steps 1-2 are as follows: The obtained RSS+IMU crowdsourced trajectory data needs to be segmented according to the pedestrian's stride length. Under the PDR algorithm, all collected RSS+IMU crowdsourced trajectory data is segmented into the following triples: { , , }; Based on 1-1, data from the accelerometer, gyroscope, and magnetometer were obtained; Using accelerometer data, the PDR algorithm calculates the step length of the movement; using gyroscope and magnetometer data, the PDR algorithm calculates the pedestrian's turning angle; offline phase It is derived from the following formula: = （1） in, and It represents the pedestrian's position before and after the time interval t.

4. The unsupervised indoor positioning method based on crowdsourced trajectory data according to claim 3, characterized in that, The specific methods for steps 1-3 are as follows: During the offline phase, the IMU sensor in the mobile phone is susceptible to environmental influences, and the collected data contains a lot of noise. Therefore, it is necessary to preprocess the motion vectors after the segmentation in steps 1-2 to reduce the impact of noise on the overall offline dataset. Step 3.1, dimensionality reduction; High-dimensional RSS data can interfere with the real-time operation of smartphones; in addition, there are many meaningless values ​​in the RSS data of the crowdsourced trajectory data segmented in steps 1-2. If no RSS value is captured, it is set to −100 dBm, and each access point AP cannot cover the entire indoor area. Therefore, feature extraction from RSS data is crucial for localization; an autoencoder (AE) is used to process the RSS sequence. , Feature dimensionality reduction is performed to provide effective feature data for subsequent clustering and smoothing processes; the dimensionality-reduced RSS sequence is obtained through an autoencoder. ; , （2） Step 3.2, clustering process; The RSS sequences are clustered using the K-means clustering algorithm; after the RSS sequences are entered into the K-means clustering, the clustered RSS sequences are obtained. ; RSS sequences are clustered based on their similarity so that { , , The triples are divided into several clusters; （3） Step 3.3, smoothing process; The motion vectors segmented from IMU data by the PDR algorithm have significant noise; the smoothing process uses { , } is the parameter. The top-K similar triples are extracted using KNN, and the average value is used for smoothing to reduce noise in the motion vectors. Then, an offline dataset is constructed. Since it's a vector, it's inappropriate to simply add them together without considering their direction and angle; therefore, we set the angle threshold to... This serves as an angle matching tool, filtering the angles of the top-K motion vectors based on the previous settings. As a matching parameter These are the top-K motion vectors, calculated according to the following formula. and Angle between: （4） According to formula (4), retain the included angle. of Build an offline dataset; in, The settings depend on the noise and angle offset of the crowdsourced trajectory data; the setting method is as follows: when motion vectors covering all areas are collected during the offline phase, offline motion vectors can be obtained. The deviation from the true motion vector; when a large number of motion deviations are obtained, an angle deviation dataset is generated; the average deviation of the angle deviation dataset is used as the angle threshold. .

Images