A passive multi-target positioning method for training data automatic labeling

Abstract

This invention discloses a passive multi-target localization method with automatic training data annotation. It employs a visible light localization system comprising: an LED control module, an LED emitter, a visible light sensor, a data acquisition module, a target recognition and localization module, and a data storage server. The LED emitter is positioned at the top of the localization space, and the visible light sensor is positioned on the ground and / or walls. The data acquisition module stores the received signals from the visible light sensor into the data storage server. The target recognition and localization module includes first and second data preprocessing modules, a neural network model, and a target position estimation module. The first data preprocessing module processes the acquired data into training samples for the neural network model; the second data preprocessing module processes the acquired data into input data for the neural network model; the neural network model identifies real and false targets, and the target position estimation module estimates the position of real targets. This invention can identify and locate multiple targets, both real and false.

Description

Technical Field

[0001] This invention relates to the field of positioning technology, and in particular to a passive multi-target positioning method with automatic annotation of training data. Background Technology

[0002] Currently, with the emergence of concepts such as smart cities and smart buildings, related Internet of Things (IoT) technologies have attracted significant attention from research institutions and hardware manufacturers. Visible light sensing technology has experienced rapid development in recent years, offering advantages such as being environmentally friendly, having a wide spectrum, and being able to reuse lighting systems. Furthermore, this technology avoids privacy issues arising from the acquisition of facial features. Therefore, passive sensing technologies based on visible light sensing have garnered considerable attention from researchers in recent years. However, multi-target localization technology based on fingerprint matching incurs substantial manual costs. In addition, the problem of spurious target localization during multi-target localization remains a challenging issue. Therefore, addressing these two problems has significant application value for passive multi-target localization based on visible light sensing, and can promote the widespread adoption of smart building systems. Summary of the Invention

[0003] This invention provides a passive multi-target localization method for automatic annotation of training data to solve the technical problems existing in the prior art.

[0004] The technical solution adopted by this invention to solve the technical problems existing in the prior art is: a passive multi-target localization method for automatic annotation of training data, which sets up a visible light localization system, including: an LED control module, multiple LED emitters, multiple visible light sensors, a data acquisition module, a target recognition and localization module, and a data storage server; multiple LED emitters are arranged at a set interval on the top of the localization space for locating moving targets, and multiple visible light sensors are arranged at a set interval on the ground and / or walls of the localization space; the LED control module controls each LED emitter to emit LED light signals; the visible light sensors receive the LED light signals emitted by the LED emitters and send the received signals to the data acquisition module; the data acquisition module collects the received signals from multiple visible light sensors and then... The data is stored in a data storage server. The target recognition and localization module includes a first data preprocessing module, a second data preprocessing module, a BP neural network model, and a target position estimation module. The first data preprocessing module processes the collected data into training samples for the BP neural network model and trains the BP neural network model using the training samples. The second data preprocessing module processes the collected data into input data for the trained BP neural network model. It preprocesses the visible light sensor signal received at the current time point and inputs the preprocessed data into the BP neural network model. The trained BP neural network model identifies and outputs the true and false targets in the localization space. The target position estimation module calls the data in the data storage server to estimate the position coordinates of the true targets identified by the BP neural network model.

[0005] Furthermore, the LED control module controls different LED transmitters to communicate with the visible light sensor using a time-division multiplexing communication method, and each LED transmitter emits an LED light signal with its unique identification code information.

[0006] Furthermore, the data acquisition module sends the light intensity data collected by each visible light sensor to the data storage server for storage; the first data preprocessing module and the second data preprocessing module call the visible light sensor received signals stored in the data storage server, obtain the shadow link according to the light intensity data received by each visible light sensor, and store the shadow link data in the data storage server. The shadow link refers to the visible light link between a certain LED emitter and its corresponding visible light sensor that is blocked.

[0007] Furthermore, the first data preprocessing module, which calls the visible light sensor received signal stored on the data storage server and performs preprocessing, includes the following steps:

[0008] Obtain the shadow links for each sampling time node and calculate the intersection points of these shadow links;

[0009] The obtained shadow link intersections are divided into multiple intersection set sets, and the center coordinates of each intersection set are obtained;

[0010] The center coordinates of the intersection set, the number of shadow links, and the number of intersections are converted into the input matrix of the neural network model.

[0011] Furthermore, the method for collecting and creating training samples for a BP neural network model includes the following steps:

[0012] Step A1: Set the sampling period and sampling time nodes, and select multiple experimental personnel as targets to be identified. Let the experimental personnel move randomly in the positioning space. At each sampling time node, obtain the actual position coordinates of each experimental personnel through the UWB positioning system. The data acquisition module simultaneously collects the received signals of multiple visible light sensors. Store the actual position coordinates of the experimental personnel and the corresponding shadow links into the data storage server.

[0013] Step A2, using the shadow links obtained at each sampling time point Calculate the link intersections of these shadow links; where:

[0014] This represents the set of shadow links corresponding to the f-th LED emitter;

[0015] This represents the i-th shadow link corresponding to the f-th LED emitter;

[0016] i is the sequence number of the shadow link;

[0017] N f The number of shadow links corresponding to the f-th LED emitter;

[0018] Step A3: Divide the obtained shadow link intersections into multiple intersection set sets and obtain the center coordinates of each intersection set;

[0019] Step A4: At each sampling time node, using the actual location coordinates of the experimenters obtained in Step A1 and the center coordinates of the intersection set obtained in Step A3, construct the following cost matrix: Each element of this cost matrix represents the distance between the actual location coordinates of the experimenter and the center coordinates of the intersection set; where:

[0020] i=1,2,…,m; j=1,2,…,n;

[0021] x i,j This represents the distance between the actual position coordinates of the i-th experimenter and the center coordinates of the intersection point of the j-th shadow link;

[0022] xi , i = 1, 2, ..., m, represent the actual location coordinates of the i-th experimenter; m is the number of experimenters;

[0023] j = 1, 2, ..., n, representing the center coordinates of the j-th shadow link intersection point; n is the number of shadow link intersection points;

[0024] Step A5, using the cost matrix, yields the following set of minimum cost equations:

[0025]

[0026]

[0027]

[0028]

[0029] Where: z ij For the corresponding x i,j The elements in the optimal matching matrix; the optimal matching matrix is ​​the matrix that achieves the best match between the true target and the candidate targets in the intersection set;

[0030] Step A6: Process the minimum cost equation system and automatically divide the intersection set into the intersection set corresponding to the true target and the intersection set corresponding to the pseudo target;

[0031] Step A7: Construct the training sample set for training the BP neural network model by taking the center coordinates of the intersection set, the number of shadow links, the number of link intersections, and the true / false label values ​​corresponding to each intersection set.

[0032] Further, in step A1, the method for collecting the actual location coordinates of each experimenter includes: deploying a UWB positioning system in the positioning space, the UWB positioning system including a positioning tag, a positioning base station, and a coordinator; the positioning tag serving as the positioning target; the positioning base station used to measure its distance from the positioning tag; the coordinator used to collect the positioning data measured by the UWB positioning system; the UWB positioning system collects the actual positioning coordinate information of the target in the following manner: the experimenter wears the positioning tag, the positioning base station completes the distance measurement with the positioning tag and sends the measurement data to the coordinator, the coordinator receives the measurement data sent by the positioning base station and forwards it to the data storage server; the data storage server saves the location coordinates of each experimenter obtained by the UWB positioning system at each sampling time point.

[0033] Furthermore, in step A3, the DBSCAN algorithm is used to divide the obtained shadow link intersections into multiple intersection set sets.

[0034] Furthermore, in step A6, the Hungarian algorithm is used to process the minimum cost equation system.

[0035] Furthermore, the method for real-time multi-target localization includes the following specific steps:

[0036] Step B1: The second data preprocessing module obtains the shadow link based on the light intensity data received by each visible light sensor and calculates the intersection of the shadow links;

[0037] Step B2: Use the DBSCAN algorithm to divide the intersection points of the obtained shadow links into multiple intersection point sets, and treat each intersection point set as a candidate target;

[0038] Step B3: Calculate the feature information of each set of shadow link intersections. The feature information of the intersection set includes: the center coordinates of the intersection set, the number of shadow links, and the number of intersections.

[0039] Step B4: Input the feature information of the intersection point set into the trained BP neural network model, and the BP neural network model will perform target recognition and output the true and false values ​​of each candidate target.

[0040] Step B5: Eliminate outliers from the set of intersection points corresponding to each true target;

[0041] Step B6: Use the average of the intersection coordinates in each set of shadow link intersections as the position estimate of the true target.

[0042] Furthermore, in step B5, the isolated forest algorithm is used to exclude outliers from the set of intersections corresponding to each true target.

[0043] The advantages and positive effects of this invention are as follows: This invention utilizes the advantages of visible light sensing, such as being environmentally friendly, having a wide spectrum, and allowing for the reuse of lighting, while simultaneously enabling the identification and location of multiple targets, achieving low cost and high accuracy, thus promoting the widespread adoption of smart building systems. Furthermore, this invention avoids the privacy violations associated with the acquisition of facial or fingerprint features. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the operation of a visible light positioning system during the offline data acquisition and training phase of the present invention.

[0045] Figure 2 This is a schematic diagram of the operation of a visible light positioning system in the real-time positioning stage according to the present invention.

[0046] Figure 3 This is a flowchart illustrating the workflow of a passive multi-target localization method for automatic annotation of training data according to the present invention.

[0047] In the figure: x(m), y(m), z(m) represent the three axes of the three-dimensional coordinate system. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0049] The Chinese definitions of the following English words, abbreviations, and phrases used in this application are as follows:

[0050] ZigBee: A low-speed, short-range, two-way wireless network technology.

[0051] WiFi: Wireless Local Area Network technology.

[0052] UWB: Ultra-wideband technology.

[0053] DBSCAN: Density-based clustering algorithm.

[0054] In the description of this invention, the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," and "bottom," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. The terms "connected" and "linked" used in this invention should be interpreted broadly. For example, they can refer to a fixed connection or a detachable connection; they can refer to a direct connection or an indirect connection through intermediate components. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0055] Please see Figures 1 to 3A passive multi-target localization method with automatic training data annotation is disclosed. A visible light localization system is set up, comprising: an LED control module, multiple LED emitters, multiple visible light sensors, a data acquisition module, a target recognition and localization module, and a data storage server. Multiple LED emitters are arranged at predetermined intervals on the top of a localization space for locating moving targets, and multiple visible light sensors are arranged at predetermined intervals on the ground and / or walls of the localization space. The LED control module controls each LED emitter to emit LED light signals. The visible light sensors receive the LED light signals emitted by the LED emitters and send the received signals to the data acquisition module. The data acquisition module collects the received signals from the multiple visible light sensors and stores them in the data storage server. The target recognition and localization module includes a first data... The system comprises a preprocessing module, a second data preprocessing module, a BP neural network model, and a target position estimation module. The first data preprocessing module processes the collected data into training samples for the BP neural network model and trains the model using these samples. The second data preprocessing module processes the collected data into input data for the trained BP neural network model. It preprocesses the visible light sensor signal received at the current time point and inputs the preprocessed data into the BP neural network model. The trained BP neural network model inputs the preprocessed data from the second data preprocessing module and further identifies and outputs the true and false targets in the positioning space. The target position estimation module calls data from the data storage server to estimate the position coordinates of the true targets identified by the BP neural network model.

[0056] The visible light sensor receives the LED light signal emitted by the LED transmitter and sends the received signal to the data acquisition module. The visible light sensor and the data acquisition module can communicate wirelessly, such as via ZigBee, WiFi, or Bluetooth.

[0057] Preferably, the LED control module can control different LED transmitters to communicate with the visible light sensor using a time-division multiplexing communication method, and each LED transmitter emits an LED light signal with its unique identification code information.

[0058] Preferably, the data acquisition module can send the light intensity data collected by each visible light sensor to the data storage server for storage; the first data preprocessing module and the second data preprocessing module can call the visible light sensor received signals stored in the data storage server, obtain the shadow link according to the light intensity data received by each visible light sensor, and store the shadow link data in the data storage server. The shadow link refers to the visible light link that is blocked between a certain LED emitter and its corresponding visible light sensor.

[0059] Preferably, the first data preprocessing module, in its method of calling and preprocessing the visible light sensor received signals stored on the data storage server, may include the following steps:

[0060] Obtain the shadow links for each sampling time node and calculate the intersection points of these shadow links;

[0061] The obtained shadow link intersections are divided into multiple intersection set sets, and the center coordinates of each intersection set are obtained;

[0062] The center coordinates of the intersection set, the number of shadow links, and the number of intersections are converted into the input matrix of the neural network model.

[0063] Preferably, the method for collecting and creating training samples for a BP neural network model may include the following steps:

[0064] Step A1 allows setting the sampling period and sampling time nodes, using multiple experimental personnel as targets to be identified, and having the personnel move randomly in the positioning space; at each sampling time node, the actual position coordinates of each experimental personnel can be obtained through the UWB positioning system; the data acquisition module simultaneously acquires the received signals from multiple visible light sensors; the actual position coordinates of the experimental personnel and the corresponding shadow links can be stored in the data storage server;

[0065] Step A2 can utilize the shadow links obtained at each sampling time point. Calculate the link intersections of these shadow links; where:

[0066] This represents the set of shadow links corresponding to the f-th LED emitter;

[0067] This represents the i-th shadow link corresponding to the f-th LED emitter;

[0068] i is the sequence number of the shadow link;

[0069] N f The number of shadow links corresponding to the f-th LED emitter;

[0070] Step A3: Divide the obtained shadow link intersections into multiple intersection set sets and obtain the center coordinates of each intersection set;

[0071] Step A4: At each sampling time node, using the actual location coordinates of the experimenters obtained in Step A1 and the center coordinates of the intersection set obtained in Step A3, the following cost matrix can be constructed: Each element of this cost matrix represents the distance between the actual location coordinates of the experimenter and the center coordinates of the intersection set; where:

[0072]

[0073] x i,j This represents the distance between the actual position coordinates of the i-th experimenter and the center coordinates of the intersection point of the j-th shadow link;

[0074] x i , i = 1, 2, ..., m, represent the actual location coordinates of the i-th experimenter; m is the number of experimenters;

[0075] j = 1, 2, ..., n, representing the center coordinates of the j-th shadow link intersection point; n is the number of shadow link intersection points;

[0076] Step A5 yields the following set of minimum cost equations from the cost matrix:

[0077]

[0078]

[0079]

[0080]

[0081] Where: z ij For the corresponding x i,j The elements in the optimal matching matrix; the optimal matching matrix is ​​the matrix that achieves the best match between the true target and the candidate targets in the intersection set;

[0082] Step A6: Process the minimum cost equation system and automatically divide the intersection set into the intersection set corresponding to the true target and the intersection set corresponding to the pseudo target;

[0083] Step A7: Construct the training sample set for training the BP neural network model by taking the center coordinates of the intersection set, the number of shadow links, the number of link intersections, and the true / false label values ​​corresponding to each intersection set.

[0084] Preferably, in step A1, the method for collecting the actual location coordinates of each experimenter may include: deploying a UWB positioning system in the positioning space, the UWB positioning system including a positioning tag, a positioning base station, and a coordinator; the positioning tag serving as the positioning target; the positioning base station used to measure its distance from the positioning tag; the coordinator used to collect the positioning data measured by the UWB positioning system; the UWB positioning system collects the actual positioning coordinate information of the target in the following manner: the experimenter wears the positioning tag, the positioning base station completes the distance measurement with the positioning tag and sends the measurement data to the coordinator, the coordinator receives the measurement data sent by the positioning base station and forwards it to the data storage server; the data storage server saves the location coordinates of each experimenter obtained by the UWB positioning system at each sampling time point.

[0085] Preferably, in step A3, the DBSCAN algorithm can be used to divide the obtained shadow link intersections into multiple intersection set sets.

[0086] Preferably, in step A6, the Hungarian algorithm can be used to process the minimum cost equations.

[0087] Preferably, the method for real-time multi-target localization may include the following specific steps:

[0088] Step B1: The second data preprocessing module obtains the shadow link based on the light intensity data received by each visible light sensor and calculates the intersection of the shadow links;

[0089] Step B2: The DBSCAN algorithm can be used to divide the intersections of the obtained shadow links into multiple intersection set sets, and each intersection set can be regarded as a candidate target.

[0090] Step B3: Calculate the feature information of each set of shadow link intersections. The feature information of the intersection set includes: the center coordinates of the intersection set, the number of shadow links, and the number of intersections.

[0091] Step B4: Input the feature information of the intersection point set into the trained BP neural network model, and the BP neural network model will perform target recognition and output the true and false values ​​of each candidate target.

[0092] Step B5: Eliminate outliers from the set of intersection points corresponding to each true target;

[0093] Step B6: Use the average of the intersection coordinates in each set of shadow link intersections as the position estimate of the true target.

[0094] Preferably, in step B5, the isolated forest algorithm can be used to exclude outliers from the set of intersections corresponding to each true target.

[0095] The working principle of the present invention will be further explained below with reference to a preferred embodiment:

[0096] A passive multi-target localization method with automatic training data annotation is characterized by setting up a visible light localization system, which includes: an LED control module, multiple LED emitters, multiple visible light sensors, a data acquisition module, a target recognition and localization module, and a data storage server; multiple LED emitters are arranged at a set interval on the top of the localization space for locating moving targets, and multiple visible light sensors are arranged at a set interval on the ground and / or walls of the localization space; the LED control module controls each LED emitter to emit LED light signals; the visible light sensors receive the LED light signals emitted by the LED emitters and send the received signals to the data acquisition module; the data acquisition module collects the received signals from the multiple visible light sensors and stores them in the data storage server; The target identification and positioning module includes a first data preprocessing module, a second data preprocessing module, a BP neural network model, and a target position estimation module. The first data preprocessing module processes the collected data into training samples for the BP neural network model and trains the BP neural network model using the training samples. The second data preprocessing module processes the collected data into input data for the trained BP neural network model. It preprocesses the visible light sensor signal received at the current time point and inputs the preprocessed data into the BP neural network model. The trained BP neural network model identifies and outputs the true and false targets in the positioning space. The target position estimation module calls data from the data storage server to estimate the position coordinates of the true targets identified by the BP neural network model.

[0097] Multiple LED emitters are positioned at predetermined intervals on the ceiling of the positioning space used to locate moving targets, while multiple visible light sensors are positioned at predetermined intervals on the floor of the positioning space. An LED control module controls each LED emitter to emit LED light signals. The visible light sensors receive the LED light signals emitted by the LED emitters and transmit the received signals to a data acquisition module. The LED control module controls the different LED emitters to operate in a time-division multiplexing manner, and the signals emitted by different LED emitters have their own unique identification codes. The LED emitters can be mounted on the ceiling, and visible light sensors for detecting light intensity information can be deployed at equal intervals on the floor of the positioning space. Furthermore, a UWB positioning system can be configured during the offline data acquisition and training phases to obtain the actual location information of each experimenter during light intensity information acquisition.

[0098] A passive multi-target localization method with automatic training data annotation consists of two stages: offline data acquisition and training, and real-time localization. The method steps for these two stages are described in detail below:

[0099] 1. Offline data collection and training phase:

[0100] During the data acquisition and training phase, a UWB positioning system was deployed in the positioning space. UWB antennas were fixed in the positioning space, and multiple personnel wore UWB positioning tags. The coordinator of the UWB positioning system was connected to a data storage server, and was used to send the personnel's location information to the data storage server. Considering that the accuracy of UWB positioning is at the centimeter level, the personnel location information acquired by the UWB positioning system was approximated as the target's true location.

[0101] The method for collecting and creating training samples for a BP neural network model includes the following specific steps:

[0102] Step 1: Have multiple experimenters move randomly in the positioning space, and record the position coordinates {x1, x2, ..., x...} of each experimenter obtained by the UWB positioning system at each positioning time point. m Store it in the data storage server.

[0103] Step 2: The visible light sensor nodes deployed on the ground will obtain the shadow link for the current positioning period based on the collected light intensity data, and send the link information to the data storage server for storage.

[0104] Step 3: Utilize the shadow links for each cycle saved in Step 2 Calculate the link intersections of these shadow links.

[0105] Step 4: Use the DBSCAN algorithm to divide the shadow link intersections obtained in Step 3 into multiple intersection set sets, and obtain the center coordinates of each intersection set. The corresponding coordinates represent the center coordinates of the 1st, 2nd, and nth intersection point sets.

[0106] Step 5: In each positioning cycle, using the position coordinates of the experimenter measured by UWB in Step 1 and the center coordinates of the intersection set in Step 4, construct the following cost matrix: Each element of this cost matrix represents the distance between the actual location coordinates of the experimenter and the center coordinates of the intersection set; where:

[0107] i=1,2,…,m; j=1,2,…,n;

[0108] x i,j This represents the distance between the actual position coordinates of the i-th experimenter and the center coordinates of the intersection point of the j-th shadow link;

[0109] x i , i = 1, 2, ..., m, represent the actual location coordinates of the i-th experimenter; m is the number of experimenters;

[0110] j = 1, 2, ..., n, representing the center coordinates of the j-th shadow link intersection point; n is the number of shadow link intersection points;

[0111] Step 6: The following set of minimum cost equations is obtained through the cost matrix:

[0112]

[0113]

[0114]

[0115]

[0116] Where: z ij For the corresponding x i,j The elements in the optimal matching matrix; the optimal matching matrix is ​​the matrix that achieves the best match between the true target and the candidate targets in the intersection set;

[0117] Step 7: Use the Hungarian algorithm to process the minimum cost equations in Step 6, and automatically divide the intersection set into the intersection set corresponding to the true target and the intersection set corresponding to the pseudo target.

[0118] Step 8: Extract feature information based on the set of intersection points of real and false targets and their corresponding shadow link information. The feature information includes the number of intersection points and the number of shadow links in the set of intersection points.

[0119] Step 9: Use a BP neural network to train the feature information of real and false targets to obtain a BP neural network model for recognizing real and false targets in the localization scene.

[0120] 2. Real-time positioning stage:

[0121] Step 10: Infer each shadow link based on the collected light intensity data, and calculate the intersection of the shadow links.

[0122] Step 11: Use the DBSCAN algorithm to divide the intersection points of the shadow links obtained in Step 10 into individual intersection point sets, and treat each intersection point set as a candidate target.

[0123] Step 12: For each set of shadow link intersections obtained in Step 11, calculate the feature information of the intersection set, including: the center coordinates of the intersection set, the number of shadow links, and the number of intersections.

[0124] Step 13: Use the BP neural network model trained in Step 9 to process the feature information corresponding to each candidate target in Step 12 to determine whether the candidate target is real or fake.

[0125] Step 14: Use the Isolation Forest algorithm to eliminate outliers in the intersection set corresponding to each true target.

[0126] Step 15: Use the average value of the intersection coordinates in each set of shadow link intersections optimized in Step 14 as the position estimate of the true target.

[0127] The aforementioned LED control module, LED transmitter, visible light sensor, data acquisition module, target recognition and positioning module, data storage server, first data preprocessing module, second data preprocessing module, BP neural network model, target position estimation module, UWB positioning system, positioning tag, positioning base station and coordinator and other functional modules and components can adopt applicable functional modules and components in the prior art; or adopt functional modules and components in the prior art and construct them using conventional technical means.

[0128] The embodiments described above are only used to illustrate the technical ideas and features of the present invention. Their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The patent scope of the present invention should not be limited by these embodiments. That is, all equivalent changes or modifications made in accordance with the spirit disclosed in the present invention still fall within the patent scope of the present invention.

Claims

1. A passive multi-target localization method for automatic labeling of training data, characterized in that, The visible light positioning system comprises an LED control module, a plurality of LED emitters, a plurality of visible light sensors, a data acquisition module, a target recognition and positioning module, and a data storage server. The plurality of LED emitters are arranged at the top of a positioning space for positioning a moving target at a set interval. The plurality of visible light sensors are arranged on the ground and / or wall of the positioning space at a set interval. The LED control module controls the LED emitters to emit LED light signals. The visible light sensors receive the LED light signals emitted by the LED emitters and send the received signals to the data acquisition module. The data acquisition module collects the received signals from the plurality of visible light sensors and stores them in the data storage server. The target recognition and positioning module comprises a first data preprocessing module, a second data preprocessing module, a BP neural network model, and a target position estimation module. The first data preprocessing module is used to process the collected data into training samples of the BP neural network model. The training samples are used to train the BP neural network model. The second data preprocessing module is used to process the collected data into input data of the trained BP neural network model. The received signals of the visible light sensors at the current time node are preprocessed, and the preprocessed data is input into the BP neural network model. The trained BP neural network model identifies and outputs the true and false targets in the positioning space. The target position estimation module calls the data in the data storage server to estimate the position coordinates of the true target identified by the BP neural network model. The data acquisition module sends the light intensity data collected by each visible light sensor to the data storage server for storage. The first data preprocessing module and the second data preprocessing module call the received signals of the visible light sensors stored in the data storage server, obtain the shadow link according to the light intensity data received by each visible light sensor, and store the shadow link data in the data storage server. The shadow link refers to the blocked visible light link between a certain LED emitter and its corresponding visible light sensor.

2. The passive multi-target localization method of automatic labeling of training data according to claim 1, characterized in that, The LED control module controls different LED emitters to communicate with the visible light sensors using time division multiplexing, and each LED emitter emits an LED light signal with a unique identity code information.

3. The passive multi-target localization method of automatic labeling of training data according to claim 1, characterized in that, The method for calling and preprocessing the received signals of the visible light sensors at the current time node stored in the data storage server by the first data preprocessing module comprises the following steps: Obtain the shadow link at each sampling time node, and calculate the link intersection points of the shadow links. Divide the obtained shadow link intersection points into a plurality of intersection point sets, and obtain the center coordinates of each intersection point set. Convert the center coordinates of the intersection point sets, the number of shadow links, and the number of intersection points into an input matrix of the neural network model.

4. The passive multi-target localization method of automatic labeling of training data according to claim 3, characterized in that, The method for collecting and making the training samples of the BP neural network model comprises the following steps: Step A1, set the sampling period and sampling time node, take multiple experimenters as the to-be-identified targets, and let the experimenters move randomly in the positioning space; at each sampling time node, the actual position coordinates of each experimenter are obtained through a UWB positioning device; a data acquisition module synchronously acquires the received signals of multiple visible light sensors; The actual position coordinates of the experimenters and the corresponding shadow links are stored in a data storage server; Step A2, using the shadow links obtained at each sampling time node , calculating the intersection points of the links of these shadow links; wherein: a set of shadow links representing the fth LED emitter; represents the ith shadow link corresponding to the fth LED emitter; sequence number for the shadow link; the number of shadow links corresponding to the fth LED emitter pair; Step A3, the intersection points of the acquired shadow links are divided into multiple intersection point sets, and the center coordinates of each intersection point set are obtained; Step A4, at each sampling time node, using the actual position coordinates of the experimental personnel obtained in step A1 and the center coordinates of the intersection set obtained in step A3, a cost matrix is constructed as follows: Each element of the cost matrix is the distance between the actual position coordinates of the experimental personnel and the center coordinates of the intersection set; wherein: ; i = 1, 2,..., m; j = 1, 2,..., n; represents the distance between the actual position coordinates of the ith tester and the intersection center coordinates of the jth shadow link; , , represents the actual position coordinates of the i-th experimenter; m is the number of experimenters; , , represents the jth shadow link intersection center coordinate; n is the number of shadow link intersections; Step A5, the following minimum cost equation set is obtained through a cost matrix: ； ； ； wherein: is an element in the optimal matching matrix corresponding to the optimal matching matrix; the optimal matching matrix is such that the true target and the candidate targets in the intersection set achieve optimal matching; Step A6, the minimum cost equation set is processed, and the intersection point sets are automatically divided into intersection point sets corresponding to true targets and intersection point sets corresponding to false targets; Step A7, the center coordinates of the intersection point sets, the number of shadow links, the number of link intersection points, and the true-false label values corresponding to each intersection point set are constructed into a training sample set for training a BP neural network model.

5. The passive multi-target localization method of automatic labeling of training data according to claim 4, characterized in that, In step A1, the method for acquiring the actual position coordinates of each experimenter includes: deploying a UWB positioning system in the positioning space, the UWB positioning system including positioning tags, positioning base stations, and a coordinator; the positioning tags serve as positioning targets; the positioning base stations are used to measure distances from the positioning tags; the coordinator is used to collect positioning data measured by the UWB positioning system; the actual positioning coordinate information of the targets is acquired by the UWB positioning system in the following manner: the experimenters wear the positioning tags; after the positioning base stations complete distance measurement with the positioning tags, the positioning base stations send the measurement data to the coordinator; after the coordinator receives the measurement data sent by the positioning base stations, the coordinator forwards the measurement data to a data storage server; the data storage server saves the position coordinates of each experimenter acquired by the UWB positioning system at each sampling time node.

6. The passive multi-target location method of automatic labeling of training data according to claim 4, characterized in that, In step A3, the DBSCAN algorithm is used to divide the acquired intersection points of the shadow links into multiple intersection point sets.

7. The passive multi-target location method of automatic labeling of training data according to claim 4, characterized in that, In step A6, the Hungarian algorithm is used to process the minimum cost equation set.

8. The passive multi-target location method of automatic labeling of training data according to claim 1, characterized in that, The method for performing real-time multi-target positioning includes the following specific steps: Step B1, a second data preprocessing module acquires shadow links according to light intensity data received by each visible light sensor, and calculates intersection points of the shadow links; Step B2, the DBSCAN algorithm is used to divide the acquired intersection points of the shadow links into multiple intersection point sets, and each intersection point set is regarded as a candidate target; Step B3, feature information of each acquired shadow link intersection point set is calculated, the feature information of the intersection point set including: center coordinates of the intersection point set, number of shadow links, and number of intersection points; Step B4, the feature information of the intersection point set is input into a BP neural network model that has been trained, target identification is performed by the BP neural network model, and true and false values of each candidate target are output; Step B5, abnormal points in each intersection point set corresponding to a true target are excluded; Step B6, the average value of the intersection point coordinates in each shadow link intersection point set is used as a position estimate of the true target.

9. The passive multi-target location method of automatic labeling of training data according to claim 8, characterized in that, In step B5, the isolation forest algorithm is used to exclude abnormal points in each intersection point set corresponding to a true target.

Images