CN113727273B

CN113727273B - Personnel indoor semantic track reconstruction method based on wireless crowdsourcing data

Info

Publication number: CN113727273B
Application number: CN202110952934.5A
Authority: CN
Inventors: 庄园; 曹晓祥; 王轩; 杨先圣
Original assignee: Wuhan University WHU
Current assignee: Wuhan University WHU
Priority date: 2021-08-19
Filing date: 2021-08-19
Publication date: 2022-04-26
Anticipated expiration: 2041-08-19
Also published as: CN113727273A

Abstract

The invention provides a personnel indoor semantic track reconstruction method based on wireless crowdsourcing data, which is realized without additional equipment installation, additional field measurement work and a map. Firstly, establishing a relational database of POI and wireless signal information and a wireless fingerprint database through crowdsourcing data, and mainly comprising three works: valuable wireless signal information is extracted from original observation data through fuzzy search, observed wireless signals and POI are paired one by one through a DBSCAN clustering algorithm, and construction of a relative semantic graph is completed. Secondly, selecting the most similar fingerprint points by utilizing a neural network in the semantic positioning process; and a plurality of new relative features and a two-classification method are adopted to replace the traditional features and methods, so that the migration capability of the classification model is greatly improved, and the output probability of the model can be used for describing the confidence of each matching result. And finally, correcting some mismatching results by using a semantic map, and finally outputting the semantic track of the pedestrian in the scene.

Description

Personnel indoor semantic track reconstruction method based on wireless crowdsourcing data

Technical Field

The embodiment of the invention relates to the technical field of personnel positioning and semantic track reconstruction, in particular to a personnel indoor semantic track reconstruction method based on wireless crowdsourcing data.

Background

The investigation of the activity track of the confirmed personnel during the epidemic situation, the accurate marketing and the like all have strong demands on the technical research of the reconstruction of the personnel track. Due to the fact that the outdoor GNSS technology is mature day by day, the outdoor personnel track reconstruction technology is not difficult to implement, and a complete absolute motion track can be recovered. Compared with the outdoor method, the indoor absolute track recovery method needs more hardware assistance and faces more complex environments, and the combination of the approximate position and the semantic information in practical application can completely support various application requirements.

Most of the current indoor people track reconstruction methods determine the absolute positions (WiFi, Bluetooth, UWB, inertial navigation and the like) of people by means of various indoor positioning technologies, and associate the corresponding absolute positions with some adjacent POIs on the basis of determining the absolute positions of pedestrians, so that the semantic track reconstruction of people is realized. The various positioning technologies used in these technologies often require complex hardware support, which also involves complex field installation and measurement work, and also requires maps and other aspects. There are significant difficulties in landing from either cost or practical implementation. In practical applications, the general location provided by the GNSS or the mobile base station and some semantic information are enough to describe the semantic track of the pedestrian, and also enough to support applications such as activity track reconstruction in epidemic investigation, character image based on the semantic track in precise marketing, and the like, so that a weakly dependent and low-cost semantic track reconstruction method is necessary.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a personnel indoor semantic track reconstruction method based on wireless crowdsourcing data, so as to realize personnel indoor semantic track reconstruction under the conditions of no map, no additional equipment installation and layout and no additional measurement.

The technical scheme of the invention is as follows: a personnel indoor semantic track reconstruction method based on wireless crowdsourcing data comprises the following steps: firstly, defining the following symbols; AP represents an access point, RSS represents signal receiving strength, POI represents a point of interest, GNSS represents a global satellite navigation positioning system, and UWB represents an ultra wide band; the name of the AP, MAC (physical address) and RSS (signal strength) are typically included in the wireless scan list. The following operations are then carried out:

step1, crowd-sourced data cleaning is carried out to screen out a wireless scanning list containing POI information and payment information, wireless connection information and inertial sensor information which are subjected to auxiliary judgment;

step2, associating the wireless signals with POI, and marking position labels on the wireless signal data so as to generate a POI fingerprint library, wherein the POI fingerprint library comprises a scanning wireless scanning list and corresponding POI names; step3, reversely deducing the relative position relationship among POI according to the RSS relationship in the associated fingerprint data established in the step2 to obtain a relative semantic map;

step4, constructing a fingerprint matching model, training the fingerprint matching model through the established POI fingerprint library, and outputting a matching result and the confidence coefficient of the result by the model;

and 5, determining POI (point of interest) passed by the pedestrian through the fingerprint matching model in the step4, and correcting mismatching through a relative semantic map.

Further, the specific implementation manner of the crowdsourcing data cleaning in the step1 is as follows;

step11, establishing a POI list under a positioning scene, wherein the POI list needs to contain fields as many as possible;

step12, matching the list in Step11 by a fuzzy search means and a formulated matching rule, and screening out data containing related common POI name information in the list from the crowdsourced wireless data, wherein the matching rule comprises Chinese and English, capital and small cases, abbreviation and partial fields;

step13, information for assisting judgment in crowdsourcing data is retained.

Further, the specific implementation manner of step2 is as follows;

step21, associating part of the wireless scanning list with a certain POI according to the auxiliary judgment information screened out in the Step1, and distinguishing POIs with the same name through the MAC of the associated AP; if the two POIs are both KFCs and the names of the APs in the two POIs are both the same and are both 'AP-KFCs', the two APs can be distinguished through the physical addresses (MACs) of the two APs, namely the two POIs are distinguished;

step22, for some wireless scanning lists without associated auxiliary information, the RSS of some AP in the data is always ranked in the scanning list, and the name of the AP contains some POI field in the POI name list in Step1, and the association operation is also carried out;

step23, when new scan data appears on some POI that can be exactly associated with wireless scan data, DBSCAN, a density-based clustering method with noise is used to determine whether to associate the new scan data with it, and the screened data and the data determined under the POI are divided into two categories: when a plurality of data are determined under the POI, each data is merged and clustered with the screened data, and finally whether the data can be associated with the POI or not is determined according to result voting, and fuzzy results are not associated;

step24, after the above operations are finished, establishing a wireless POI fingerprint database, including a scanning wireless scanning list and a corresponding POI name; defining a fingerprint point, namely a POI, finding out some wireless data scanned on the POI through crowdsourcing data, namely fingerprint information on the fingerprint point, namely a wireless scanning list, wherein the list comprises scanned AP names, MAC and corresponding received signal strength RSS, and then normalizing the RSS in each piece of fingerprint data;

wherein

rss_stdRespectively representing the mean and variance of the RSS list,

indicating the received signal strength of the ith AP in the scan list.

Further, the specific implementation manner of training the fingerprint matching model in the step4 is as follows;

step41, pairing the fingerprint information in the wireless fingerprint library in the Step2 pairwise, and if the two pieces of information come from the same POI, using the paired information for subsequent calculation to generate a positive sample; if the two pieces of information are from different POIs, such a pair of information is used for subsequent calculation to generate a negative sample;

step42, calculating new relative features by using the pairing formed by Step41, wherein the new relative features comprise similarity features, sorting features, shift features and coincident number features, so as to obtain a new feature vector, and the feature vector is a positive sample and a negative sample of subsequent model training;

step43, training a fingerprint matching model by using a neural network through the sorted positive and negative samples, wherein the last layer of the network is designed to be a softmax layer in the structural design of the neural network, so that the probability of the positive samples, namely the probability of the POI fingerprint data being matched, is output by the fingerprint matching model, and a threshold value pro _ threshold is obtained through a verification set.

Further, the specific implementation manner of step5 is as follows;

step51, the data to be subjected to trajectory reconstruction has a scanning list at some time points according to the scanning period, if the APs ranked in the scanning list at the front do not appear in the POI name list in Step1, the POI of the scanning list at the time point is positioned as None, otherwise, the process goes to Step 52;

step52, normalizing the RSS of the scan list scan _ list at the time point, and screening all fingerprint points of the AP in the scan _ list from the wireless POI fingerprint database;

step53, making the screened fingerprint point data and scan _ list as relative features, and calculating to obtain a plurality of feature vectors;

step54, inputting the feature vectors generated in Step53 into the fingerprint matching model trained in Step4, outputting the matching probability of each screened fingerprint point, and extracting the fingerprint point with the highest probability;

step55, comparing the probability of the fingerprint point with the highest probability with a threshold value pro _ threshold, if the probability is greater than pro _ threshold, outputting the POI name and the approximate GNSS position corresponding to the POI name as a result, otherwise, outputting None;

step56, each scanning list corresponds to a time stamp, and the result of mismatching of time and relative relation in the reconstructed track is adjusted and corrected by means of the relative semantic map generated in the Step 3;

step57, a semantic track is finally restored, circles are used for representing the matched POIs, the sizes of the circles correspond to the matching confidence degrees, arrows represent the precedence relationship, and the length of a line connecting the POIs represents the time length from the last POI to the next POI.

Compared with the prior art, the invention has the following innovation points and advantages:

1. besides wireless crowdsourcing data collected in a scene (information related to POI is required in wireless data), any other prior data including base station positions, maps, road networks and the like are not required.

2. The general position provided by the GNSS or the mobile base station is added with some semantic information to describe the indoor track of the person, but not the absolute precise position, so that the requirement of laying any additional equipment for the scene is avoided, and only the existing AP in the scene is relied on.

3. The association of the wireless signals in the crowd-sourced data with the POI is established through the crowd-sourced data partially containing the user payment information, the connection information, the RSS strength information and the hidden POI information in the wireless, so that the association work of the POI with a deterministic position (in practice, site survey and the like) is avoided.

4. After the wireless signal and the POI are associated, the relative semantic map established according to the association relationship can reflect the relative position relationship between the POI in the scene, and can correct the reconstructed semantic track.

5. The method has the advantages that the original absolute features used in fingerprint positioning are converted into various relative features, the original multi-classification problem is converted into a two-classification problem, the small sample learning problem and the problem that the model cannot migrate are avoided, model training is completed through a neural network, and the output result can express the probability of matching adjacent points and can also be used for expressing the confidence coefficient of the matching result.

Drawings

FIG. 1 is a block diagram of the system architecture of the present invention;

FIG. 2 is a list of relative features of the present invention;

FIG. 3 is a diagram of the relative semantic map construction of the present invention;

FIG. 4 is a schematic diagram of a reconstructed trajectory according to the present invention.

Detailed Description

The present invention will be described in further detail with reference to the following description and embodiments in conjunction with the accompanying drawings.

Some terms used in the present invention are explained first:

AP (access point), Access Point

RSS (received signal strength), signal reception strength

POI (Point of interest), point of interest

GNSS (Global Navigation Satellite System), Global Navigation Satellite System and positioning System

UWB (ultra Wide Bandwidth), ultra Wide band

DBSCAN (sensitivity-Based Spatial Clustering of Applications with Noise), Noise-Based Density Clustering method

The method comprises two modules of basic data preparation and semantic track reconstruction, the overall structure is shown in figure 1, and the basic data preparation and semantic track reconstruction module is subdivided into four main contents, namely POI and wireless signal association, wireless POI fingerprint database and relative semantic map construction, positioning matching model training and track reconstruction, wherein the contents of each block are explained in the following one by one:

the work in the basic data preparation includes: crowd-sourced data cleaning, association of a wireless signal scanning list and a POI, wireless POI fingerprint database and relative semantic map construction:

the crowd-sourced data cleaning is mainly to screen out a wireless scanning list containing POI information, payment information, wireless connection information, inertial sensor information (which can be used for judging whether the person stays for a long time or not moves) and the like for auxiliary judgment.

Some auxiliary judgment information in data cleaning can associate a certain wireless scanning list with a certain POI, for example, a certain piece of crowdsourcing data is a wireless signal scanning list generated in the process of paying in the 'KFC', and a certain AP in the scanning list contains a field of the 'KFC', so that the wireless scanning list can be associated with the POI of the 'KFC'; for another example, if a certain crowdsourcing data is connected to an AP for a long time, RSS ordering of the AP is always in front and the AP includes a "KFC" field, the crowdsourcing data may also be associated with the POI of "KFC". A part of POI can be associated with crowdsourced scanned wireless data by auxiliary information, wherein other data without auxiliary judgment information can be ranked at the front through RSS of the AP, POI information is contained in the AP as bottom-entering logic to expand the association relationship between the POI and the wireless data, and when repeated scanning data occurs to a certain POI, DBSCAN clustering can be used for determining whether the POI is included in a new piece of associated data. Thus, the corresponding relationship between the POI and the wireless scanning list is established, and the wireless POI fingerprint database is obtained. A field of another POI name may be included in a certain AP in a scan list associated with a certain POI, a relative positional relationship between the two POIs may be approximately established according to RSS, for example, AP1 including a "NIKE" field and AP2 including an "ANTA" field are respectively scanned in "KFC", a first near position from "KFC" is "NIKE", a second near position is "ANTA" may be obtained according to the corresponding strength, a relative positional relationship between each POI on the two POIs of "NIKE" and "ANTA" may also be obtained according to the scan list, relationships established between the POIs may be mutually calibrated, the relative positional relationship is completely established, that is, the relative semantic map. And (3) training a fingerprint matching model by using a conventional NN method through the established wireless POI fingerprint library design, and outputting a matching result and the confidence coefficient (probability) of the result by using the model.

The semantic track reconstruction comprises two main contents of POI matching and reconstruction track optimization and adjustment:

POI matching, namely fingerprint matching positioning, wherein a single scanning list with a POI name appearing in front in a fingerprint library is screened out before matching, the matching result of an undisposed scanning list is defaulted to be None, the screened list is input into a trained fingerprint matching model, the matching result with the confidence coefficient larger than a threshold value is used as the positioning result of the scanning list, and the matching result smaller than the threshold value is also the None. And after all matching is finished, generating a semantic track of POI-1- > POI-2- > POI-3- >, wherein the time duration of each POI can be approximately obtained, and then correcting partial mismatching results through a relative semantic map.

Part A Crowdsourcing data cleansing

1. The purpose is as follows: crowdsourcing data is of poor quality, and this step is to screen out useful data information.

2. The method comprises the following steps:

step1: an approximate POI list (which may be surveyed in advance or some common POI names may be selected) in a positioning scene is established, such as { "KFC", "huabei", "NIKE", "ANTA", "361", "benevolence hall", "non-printed good product", "snow", etc., which contains as many fields as possible.

And Step2, screening the crowdsourced wireless data containing the related common POI name information in the list by a fuzzy search means (the established rule comprises a series of matching rules of Chinese and English, capital and small cases, short abbreviations, partial fields and the like, and the matching rules are matched with the list in Step 1). Such as:

scan list 1: { "KFC _ 01": 57 dBm; -66dBm for "ap _ 01"; "ap _ 11": 87 dBm; ... };

scan list 2: { "ap _ 01": -57 dBm; "NIKE _ OFFICE": 68 dBm; "ap _ 21": 89 dBm; ... };

scan list 3: { "ap _ 05": -66 dBm; -86dBm for "ap _ 04"; "ap _ 111": 97 dBm; ... };

scan list 4: { "ap _ 05": -44 dBm; "anta _ store": 46 dBm; "huawei _ office": 67 dBm; ... };

......

scan lists 1,2,4 are retained after screening, with scan list3 discarded.

Step3, some information for assisting judgment in crowdsourcing data is reserved, such as:

eg.1:11:11:30 payment occurs in "KFC", at which point there is a wireless scan list scan _ list1 [ { "KFC _ 01": -57 dBm; -66dBm for "ap _ 01"; "ap _ 11": 87 dBm; ... } information is retained;

eg.2:11:11:30 code scanning behavior occurs in "NIKE", at which time there is a wireless scan list scan _ list2 [ a { "ap _ 01": -57 dBm; "NIKE _ OFFICE": 68 dBm; "ap _ 21": 89 dBm; ... } information is retained;

the terminal is always connected with the AP of KFC-01 in the period of eg.3:11:11: 30-12: 00:21, the RSS is always sorted in the scanning list in the front, and the information is reserved;

......

part B, associating the wireless signal with the POI, and generating a POI fingerprint database

1. The purpose is as follows: the wireless signal data is tagged with a location, not an absolute location tag, but rather a POI name.

2. The method comprises the following steps:

step1: the auxiliary judgment information screened out in Step3 of Part a may relate the partial scan list to a certain POI certainty, such as eg.1 and eg.2 exemplified in Step3 of Part a, wherein scan _ list1 and scan _ list2 may relate to two POI certainty, i.e., "KFC" and "Nike". POIs with the same name may be distinguished by the MAC of the associated AP; if the two POIs are both KFCs and the names of the APs in the two POIs are both the same and are both 'AP-KFCs', the two APs can be distinguished through the physical addresses (MACs) of the two APs, namely the two POIs are distinguished;

step2, some scanning list has no associated auxiliary information, and the RSS of some AP in the data is always ranked in the scanning list at the front, and the name of the AP contains a POI field of the POI name list in Part A, and the association operation can be carried out.

Step3 when new scan data appears on some POI that can be exactly associated with wireless data, the DBSCAN clustering method is used to determine whether to associate new data with it: and dividing the screened data and the data determined under the POI into two categories (when a plurality of data are determined under the POI, combining and clustering each data with the screened data, and finally determining whether each data can be associated with the POI according to result voting, wherein fuzzy results are not associated).

Step4: after the operation is finished, a wireless POI fingerprint library is established, which comprises a scanning wireless scanning list and a corresponding POI name, a fingerprint point, namely a POI, in the invention, some wireless data scanned on the POI, namely fingerprint information on the fingerprint point, namely a wireless scanning list can be found out through crowdsourcing data, the list comprises the scanned AP name, MAC and corresponding signal strength (RSS), and the RSS in each piece of fingerprint data is normalized.

Wherein

rss_stdRespectively representing the mean and variance of the RSS list,

indicating the signal strength of the ith AP in the scan list.

Part C relative semantic map

1. The purpose is as follows: and (4) reversely deducing an approximate relative relationship between POIs through the relationship established in Part B and the RSS relationship in the fingerprint data.

2. The method comprises the following steps:

step1, according to the fingerprint database established in B, the AP list of any one fingerprint data in each POI may include many other POI names, and in combination with the corresponding RSS size, the relative position relationship between the POI and the POIs may be established. Such as:

“KFC”：

scan_list1—{“KFC_01”:-57dBm；“NIKE_OFFICE”:-66dBm；“ANTA_STORE”:-74dBm；......}

scan_list2—{“KFC_01”:-53dBm；“ANTA_STORE”:-71dBm；“NIKE_OFFICE”:-73dBm；......}

scan_list3—{“KFC_01”:-55dBm；“NIKE_OFFICE”:-68dBm；“ANTA_STORE”:-77dBm；......}

......

“Nike”:

scan_list1—{“NIKE_OFFICE”:-66dBm；“KFC_01”:-69dBm；“ANTA_STORE”:-78dBm；......}

scan_list2—{“NIKE_OFFICE”:-66dBm；“KFC_01”:-70dBm；“ANTA_STORE”:-76dBm；......}

scan_list3—{“KFC_01”:-67dBm；“NIKE_OFFICE”:-69dBm；“ANTA_STORE”:-79dBm；......}

......

“ANTA”:

scan_list1—{“ANTA_STORE”:-48dBm；“ap_01”:-66dBm；“ap_01”:-76dBm；“ap_01”:-86dBm；“ap_01”:-88dBm；“KFC_01”:-99dBm；......}

......

“HUAWEI”：

scan_list1—{“HUAWEI”:-48dBm；“ap_01”:-66dBm；“ap_01”:-76dBm；“ap_01”:-86dBm；“ap_01”:-88dBm；.....}

step2. Each piece of fingerprint data in Step1 can establish a set of relative relations, such as each subgraph in the upper column in FIG. 3, and in combination with each subgraph result, such as the result of scan _ list3 in "KFC" in Step1 is inconsistent with the results of scan _ list1 and scan _ list2, the result of scan _ list3 can be corrected by scan _ list1 and scan _ list2, and the combined result is used as the final relative semantic map for output.

Step3 similarly, data between different POIs can be calibrated, such as: the scan _ list1 and scan _ list2 in the "KFC" in Step1 can be used to check and correct the result of the scan _ list1 in the "ANTA".

Step4: and finally obtaining a relative position relationship map, namely a relative semantic map, of each POI in the scene.

Part D fingerprint matching model training

1. The purpose is as follows: and (4) expressing the nonlinear relation by the neural network, and training a fingerprint matching model by using the neural network.

2. The method comprises the following steps:

and Step1, pairing the fingerprint information in the wireless fingerprint library in the Part B in pairs, and if the two pieces of information are from the same POI, using the paired information for subsequent calculation to generate a positive sample. Such a pair of information is used for subsequent calculations to generate a negative example if the two pieces of information are from different POIs. Such as: the scan _ list1 and scan _ list2 pair of "KFC" in Part C is used to calculate the positive sample features and the scan _ list1 and scan _ list1 pair of "HUAWEI" in Part C is used to calculate the negative sample features.

Step2, calculating new relative features by using the pairing formed by Step1, wherein the new relative features comprise similarity features, ordering features, shifting features and coincidence features, namely, comparing and calculating two RSS scanning lists in the pairing to obtain a new feature vector, and the feature vector is a positive sample and a negative sample of subsequent model training (specifically, the used features are shown in FIG. 2, the features are mainly divided into four categories, and each category comprises specific features). Such as: the number of coincidences of APs in two lists, namely scan _ list1 and scan _ list2 of "KFC" in Part C, the RSS similarity of the two lists (specifically using the similarity as shown in fig. 2), the sorting of the AP with the strongest RSS in scan _ list1 in scan _ list2 (specifically using the sorting feature as shown in fig. 2), the number of bits that a certain AP in scan _ list1 needs to be moved so that the position of the AP is the same as that of scan _ list2 (specifically using the shifting feature as shown in fig. 2), and the like.

Step3, training a fingerprint matching model by utilizing a neural network through the sorted positive and negative samples, wherein the last layer of the network is designed to be a softmax layer in the structural design of the neural network, so that the model output can be the probability of the positive sample, namely the probability of matching the POI fingerprint data. And searches for a probability threshold pro threshold on a correct match through the validation set data.

Part E semantic track reconstruction and correction

1. The purpose is as follows: POI (point of interest) passed by the pedestrian is determined in a fingerprint matching mode, and some mismatching is corrected through a relative semantic map.

2. The method comprises the following steps:

step1: data to be subjected to trajectory reconstruction has scanning lists at certain time points according to scanning cycles, if the APs ranked at the top in some scanning lists do not appear in the POI name list in Part a (matching is performed by using fuzzy search as well), the scanning list time point locates the matched POI as None, otherwise, Step2 is entered.

Step2, the RSS of the scan list scan _ list at the time point is normalized, which is the same as Step4 in Part B. And screening all fingerprint points of the AP in the scan _ list from the wireless POI fingerprint library.

And Step3, performing relative feature calculation on the screened fingerprint point data and the scan _ list, and calculating to obtain a plurality of feature vectors (the screened fingerprint points all correspond to one feature vector) in the same calculation method as Step2 in Part D.

And Step4, inputting the feature vector generated in Step3 into a fingerprint matching model trained in Step3 in Part D, outputting the matching probability of each screened fingerprint point, and extracting the fingerprint point with the highest probability.

Step 5: comparing the probability of the fingerprint point with the highest probability with a threshold value pro _ threshold obtained in Part D, if the probability is greater than pro _ threshold, outputting the POI name and the approximate GNSS position corresponding to the POI name as a result, otherwise, outputting None.

Step 6: each scanning list corresponds to a time stamp, and the result that the time in the reconstructed track is not matched with the relative relation can be adjusted and corrected by means of the relative semantic map generated in Part C. Such as: the semantic map generated in Part C shows that the 'KFC' and the 'Nike' are in adjacent positions, but the time from the 'KFC' to the 'Nike' in the generated track is very long, which indicates that a mismatching exists, and the fingerprint point with the second probability in the matching is verified, and so on.

Step 7: finally, a semantic track can be recovered, such as: { "KFC" (xmin) "ANTA" -) (× min) "NIKE" -) (× min) "ADIDAS" -) - (× min) "FLA > (× min)" HUAWEI "-) (× min)" MEIZU "-) -) - (× min)" MEIZU "-) - (× min)" and manor "- (× min) -) - }, arrows indicate a precedence relationship, and" × min "indicates the duration of time from one POI to the next POI (fig. 4 makes a simple illustration of the generated semantic track, circles indicate the matched POIs, the size of the circles corresponds to the confidence of the matches, and the lengths of lines connecting between POIs indicate the duration of the POI from the previous POI to the next POI).

The above description is only a preferred embodiment of the invention and should not be taken as limiting the invention, and any modification, equivalent replacement, or improvement made within the spirit and principle of the invention should be included in the protection scope of the invention.

Claims

1. A personnel indoor semantic track reconstruction method based on wireless crowdsourcing data is characterized by comprising the following steps:

firstly, defining the following symbols; AP represents an access point, RSS represents signal receiving strength, POI represents a point of interest, GNSS represents a global satellite navigation positioning system, and UWB represents an ultra wide band; the following operations are then carried out:

step1, crowd-sourced data cleaning is carried out to screen out a wireless scanning list containing POI information and payment information, wireless connection information and inertial sensor information which are subjected to auxiliary judgment; the wireless scanning list comprises the name of the AP, a physical address MAC and signal strength RSS;

step2, associating the wireless signals with POI, and marking position labels on the wireless signal data so as to generate a POI fingerprint library, wherein the POI fingerprint library comprises a scanning wireless scanning list and corresponding POI names;

the specific implementation manner of the step2 is as follows;

step21, associating part of the wireless scanning list with a certain POI according to the auxiliary judgment information screened out in the Step1, and distinguishing POIs with the same name through the MAC of the associated AP;

step24, after the above operations are finished, establishing a POI fingerprint database, including a scanning wireless scanning list and a corresponding POI name; defining a fingerprint point, namely a POI, finding out some wireless data scanned on the POI through crowdsourcing data, namely fingerprint information on the fingerprint point, namely a wireless scanning list, wherein the list comprises scanned AP names, MAC and corresponding received signal strength RSS, and then normalizing the RSS in each piece of fingerprint data;

wherein

rss_stdRespectively representing the mean and variance of the RSS list,

indicating the received signal strength of the ith AP in the scanning list;

step3, reversely deducing the relative position relationship among POI according to the RSS relationship in the associated fingerprint data established in the step2 to obtain a relative semantic map;

the specific implementation manner of training the fingerprint matching model in the step4 is as follows;

step43, training a fingerprint matching model by utilizing a neural network through the sorted positive and negative samples, wherein the last layer of the network is designed as a softmax layer in the structural design of the neural network, so that the probability of the positive samples, namely the probability of the POI fingerprint data, output by the fingerprint matching model is obtained, and a threshold value pro _ threshold is obtained through a verification set;

step5, determining POI (point of interest) through which the pedestrian passes by the fingerprint matching model in the step4, and correcting mismatching through a relative semantic map;

the specific implementation manner of the step5 is as follows;

2. The method for reconstructing the indoor semantic track of the person based on the wireless crowdsourcing data as claimed in claim 1, wherein: the specific implementation manner of the crowdsourcing data cleaning in the step1 is as follows;

step13, information for assisting judgment in crowdsourcing data is retained.