CN117560761A

CN117560761A - Dynamic floor positioning method based on multiple sensors

Info

Publication number: CN117560761A
Application number: CN202311414263.2A
Authority: CN
Inventors: 毛永毅; 王晓甜
Original assignee: Xian University of Posts and Telecommunications
Current assignee: Xian University of Posts and Telecommunications
Priority date: 2023-10-30
Filing date: 2023-10-30
Publication date: 2024-02-13

Abstract

The invention relates to a dynamic floor positioning method based on multiple sensors, which belongs to the field of floor positioning, and the multi-floor indoor positioning comprises 2 stages: the off-line stage is mainly responsible for constructing a fingerprint database, and comprises a floor identification fingerprint database and a training LDA multi-classification model; the online stage is mainly responsible for distinguishing the initial floor of the online fingerprint, and acquiring the corner information of pedestrians in the stairwell of the building and the priori knowledge of the stairwell of the building through the MEMS sensor to acquire the dynamic real-time floor information, and experiments show that the static floor recognition accuracy is 96.7%, the corner recognition accuracy is 100%, and the dynamic floor recognition accuracy is 100%. The method can obtain the positioning result after the movement of pedestrians is finished in the process of going upstairs and downstairs once, and compared with the existing fingerprint floor positioning algorithm, the method only needs one fingerprint positioning, and reduces the time cost required by positioning.

Description

Dynamic floor positioning method based on multiple sensors

Technical Field

The invention belongs to the field of floor positioning, and particularly relates to a dynamic floor positioning method based on multiple sensors.

Background

In recent years, with the development of the internet of things, the demand for location-based services is increasing, and the development of positioning technology is promoted. The global positioning system GPS (global positioning system) can meet the positioning requirement of an outdoor scene, but in an indoor environment, the GPS receiver and the satellite cannot normally communicate, and the indoor position cannot be accurately identified. Common technologies for indoor positioning include bluetooth, radio Frequency Identification Devices (RFID), loRa, wi-Fi, ultra Wideband (UWB), and the like.

The coverage range of WiFi positioning is wide, accurate positioning can be carried out indoors, and the limitation that GPS cannot receive signals is reduced; meanwhile, the Wi-Fi network is covered in most places, so that the WiFi positioning technology is easy to popularize and use. The accuracy of Wi-Fi positioning may be affected by a number of factors such as building materials, wi-Fi signal penetration capabilities, etc., and therefore in some cases, wiFi positioning may be subject to large errors.

Floor positioning is an important aspect of indoor positioning technology, is closely related to location-based services, and many problems to be solved and improved still remain in the current WiFi-based floor positioning technology, so that it is very important to maintain and improve the accuracy of combined positioning through research on the WiFi-based floor positioning technology.

Disclosure of Invention

Aiming at the defect of low positioning precision in the current floor positioning technology based on WiFi, the invention provides a dynamic floor positioning method based on multiple sensors. The positioning method disclosed by the invention is based on an RSS fingerprint positioning algorithm based on WiFi, integrates acceleration information and floor air pressure information of each moving direction of pedestrian activity, realizes accurate positioning of indoor floors and positions, overcomes the defects of low positioning precision and the like in indoor positioning, and effectively improves the accuracy and stability of indoor positioning.

In order to achieve the above purpose, the present invention is realized by the following technical scheme: a dynamic floor positioning method based on multiple sensors comprises the following steps:

step 1) construction of fingerprint database

Collecting RSS values of a plurality of APs at different floors for constructing a fingerprint database;

step 2) training an LDA multi-classification model according to the fingerprint database established in the step 1);

step 3) initial static floor discrimination

Performing multi-classification on the floor fingerprint data by using an LDA multi-classification model to obtain an initial static floor;

step 4) dynamic floor discrimination: the method comprises the following steps of obtaining dynamic real-time floor information by obtaining the angular speed, acceleration and barometer data of pedestrians during walking in a stairwell through an MEMS sensor, wherein the specific process is as follows:

the method comprises the steps of detecting acceleration and barometer data of a pedestrian, judging the motion state of the pedestrian according to a step frequency fusion air pressure difference, recording the number of corners by taking the motion state as a precondition to judge the current floor of the pedestrian, calculating according to a calculation formula of the relation between the total number of the corners and the floor, if the calculation result is an integer, indicating that the pedestrian just passes through the corner of the stair, judging whether the pedestrian goes upstairs/downstairs or is stationary by identifying the state, if the state is upstairs/downstairs, keeping the floor result, and if the state is stationary, subtracting 0.5 layer from the calculation floor, namely, the pedestrian is on a landing between the stairs.

Further, the LDA multi-classification model in the step 2) is:

according to the multi-category LDA definition, the inter-category dispersion matrix of F categories is expressed as

μ _i Is the mean vector of all the reference points of an AP in the i layer, n _i Is the number of all reference points in the ith layer of the AP, mu is the average vector of the reference points of the AP in all layers, and the in-class dispersion matrix of F classes is expressed as

S _wi Is represented by the covariance matrix of the RSS of the AP at all reference points of the ith layer

Wherein k represents the number of reference points, f represents the floor, i represents the number of APs,the method is characterized in that the RSS value of the AP at the j-th reference point in the f floor is obtained according to a two-class LDA optimization target solving algorithm, a multi-class optimization target is obtained based on the principle that the intra-class variance is minimum after projection and the inter-class variance is maximum, the LDA is applied to the multi-class condition, namely multi-class data is projected to a low-dimensional space, the low-dimensional space at the moment is not a straight line but a hyperplane, the dimension of the low-dimensional space is assumed to be d, and the corresponding base vector is set to be (omega) ₁ ,ω ₂ ,…,ω _d ) The matrix of basis vectors is W (AP _k X d), generalizing the two classifications to multiple classifications, optimizing the objective function to be

Wherein,is generalized Rayleigh Li Shang, whose maximum is matrix +.>The product of the maximum d values expressed by the formula (5) is the matrix +.>The product of the maximum d eigenvalues, the corresponding matrix W is the projection matrix of the eigenvector corresponding to the maximum d eigenvalues; the probability density function, the mean vector and the variance matrix of the Gaussian distribution of the fingerprint data of the AP set on the ith floor can be obtained after the projection matrix W is utilized for projection;

when the tasks of the data of multiple types in multiple classifications meet Gaussian distribution and the variances are the same, the LDA obtains an optimal Bayesian classifier, and the optimal Bayesian classifier:

h*(x)＝arg max p(c|x) (6)

for each sample x, one can choose to maximize the posterior probability p (c|x), x being the sample sum, c being the class label; if the class mark is required to make the value p (c|x) large, namely the value of dividing the center distance of two training samples belonging to different classes by the center distance value of the training sample of the same class is large, namely the maximum value of formula (6), in the LDA classification, input data are RSS data training samples of all APs at reference points, output a predicted result of the floor where pedestrians are located, abstract the floor positioning as a multi-classification problem, sample data are RSS data of all APs at all reference points of different floors, and a specific floor classification result is obtained through the multi-classification prediction model.

Further, in the step 4), when the pedestrian walks in the stairwell, the heading angle is calculated by adopting a formula (5)

head (k) is the calculated heading angle, head (0) is the initial heading angle, k is the number of acquisitions, w _i Is angular velocity data, Δt is a sampling period, and the course angular transformation detection algorithm adopts formula (7):

head(k+τ)-head(k)＞Δh _thr (8)

first, a detection window tau and a threshold deltah are set _thr Detecting whether the variation of the course angle in the window is larger than a threshold value, if so, adding one to count, and obtaining the dynamic floor by the number of times of the corners and the structure of the stairwell when the pedestrian passes through the corners once.

Further, when the corner of the stairwell is 1, the calculation formula of the relation between the total number of the corners and the floors is as follows:

num represents the total number of corners the pedestrian passes by, and N represents the number of floors.

Further, the specific rule for distinguishing the pedestrian motion state in the step 4) is as follows: the comparison of the barometer variation is only carried out when the pedestrian is detected to stride, so that the pedestrian upstairs and downstairs states are determined, if the barometer variation is positive, the pedestrian states are downstairs, and if the barometer variation is negative, the pedestrian states are upstairs.

Furthermore, the zero-crossing comparison algorithm is adopted to detect the step frequency, the acceleration of the pedestrian crossing one step will have zero moment in the process of rising, wave crest, wave trough and acceleration of the pedestrian crossing one step, and the pedestrian crossing one step is considered to be crossing one step when the zero moment is detected once.

Compared with the prior art, the invention has the beneficial effects that: the invention provides a dynamic floor positioning scheme based on multiple sensors, and provides a multi-floor indoor positioning model in a WiFi scene. The floor discrimination in floor positioning is divided into discrimination of an initial static floor and discrimination of a dynamic floor, and the supervised dimension reduction technology (LDA) is utilized to carry out multi-classification on the floor fingerprint data to obtain the initial static floor. By introducing MEMS sensors, acceleration of each moving direction of pedestrian activity is obtained, barometer is used for obtaining air pressure information, priori knowledge of a stairwell is combined, pedestrians are identified in the process of going upstairs and downstairs in the stairwell, RSS fingerprints of WiFi are combined for realizing multi-floor indoor positioning, and dynamic floor information and user positions are obtained.

The method can reflect real-time floor change of the traveler in the multi-story building. Experiments show that the static floor recognition accuracy is 96.7%, the corner recognition accuracy is 100%, and the dynamic floor recognition accuracy is 100%. The method can obtain the positioning result after the movement of pedestrians is finished in the process of going upstairs and downstairs once, and compared with the existing fingerprint floor positioning algorithm, the method only needs one fingerprint positioning, and reduces the time cost required by positioning.

Drawings

FIG. 1 is a schematic diagram of the multi-sensor based dynamic floor positioning of the present invention.

Fig. 2 is an RSS profile.

FIG. 3 is a graph showing the change of the data of the air pressure of the upstairs and downstairs.

Fig. 4 is a block diagram of a dynamic floor recognition algorithm.

Fig. 5 is an experimental environment in which (a) is a laboratory distribution and (b) is a schematic and reference point for AP deployment per layer.

Fig. 6 is a static floor identification schematic.

Fig. 7 is a heading angle and step frequency detection in which (a) indicates that a pedestrian passing from five floors to three floors passes through 3 stair corners and (b) indicates a point in time at which the pedestrian strides during this period.

Fig. 8 shows the actual and detected corners for each floor.

FIG. 9 is a graph of barometric pressure as a function of number of steps.

Fig. 10 is a floor positioning by different initial floors.

Figure 11 floor identification under multiple activities.

Detailed Description

The invention is described in further detail below with reference to the drawings and examples of implementation.

As shown in fig. 1, the invention provides a WLAN multi-floor indoor positioning method based on RSS fingerprints, aiming at the problem of multi-floor indoor positioning in WLAN scene, and the multi-floor indoor positioning is shown in fig. 1. Multi-floor indoor positioning includes 2 phases: an offline phase and an online phase. The off-line stage is mainly responsible for constructing a fingerprint database, including floor recognition fingerprint database and training an LDA multi-classification model. The online stage is mainly responsible for distinguishing the initial floors of online fingerprints, and acquiring the corner information of pedestrians in stairs and priori knowledge of the stairs through the MEMS sensor to obtain dynamic and real-time floor information.

Multi-floor classification algorithm based on LDA

Experimental analysis shows that the average RSS difference between adjacent layers of the same AP is between-13.37 dBm and-20.89 dBm. Therefore, the number of floors can be distinguished through a classification model according to the difference of RSS values of different floors, and the common classification model comprises SVM, KNN, LDA, a neural network and the like, and the distribution condition of the RSS values of different reference points of the same AP signal in the same floor basically meets Gaussian distribution. As shown in fig. 2, the location x=0 of the AP, the grid or location far from the AP has a lower RSS value, and the grid or location near the AP has a higher RSS value. The data LDA obeying the Gaussian distribution can be classified well, the LDA algorithm is simple, the training time is short, and the implementation is easy, so that the LDA is selected for floor discrimination.

LDA is a supervised dimension reduction technology, and based on the LDA classification idea, high-dimension data samples are projected to the optimal discrimination vector space, so that the projected new sample data can meet the condition of minimum intra-class distance and maximum inter-class distance, and the original data is classified in the vector space. The algorithm can ensure that the sample has the best separability in the vector space after projection while reducing the dimension of the sample data. The classification concept of LDA can be generalized to the multi-classification problem by deforming the intra-class divergence and inter-class divergence formulas.

Is represented by the covariance matrix of the RSS of the AP at all reference points of the ith layer

Wherein k represents the number of reference points, f represents the floor, i represents the number of APs,is the RSS value of the AP at the j-th reference point within the f floor. And obtaining a multi-classification optimization target based on the principle that the intra-class variance is minimum and the inter-class variance is maximum after projection according to a two-classification LDA optimization target solving algorithm. LDA is applied in multi-class situations, i.e. projection of multi-class data into a low-dimensional space, where the low-dimensional space will not be a straight line but a hyperplane. Assuming that the dimension of the low-dimensional space is d, the corresponding basis vector is set to (ω ₁ ,ω ₂ ,…,ω _d ) The matrix of basis vectors is W (AP _k X d). The two classifications are generalized to multiple classifications, and the objective function is optimized as

Wherein,is generalized Rayleigh Li Shang, whose maximum is matrix +.>Is the maximum eigenvalue of (c). The product of the maximum d values expressed by equation (5) is the matrix +.>The product of the maximum d eigenvalues, where the corresponding matrix W is the projection matrix of the eigenvector corresponding to the maximum d eigenvalues. And (3) obtaining a probability density function, a mean vector and a variance matrix of the Gaussian distribution of the fingerprint data of the AP set on the ith floor after projection by using the projection matrix W. Better classification results can be obtained for LDA with the same or similar fingerprint data.

When the tasks meet Gaussian distribution and the variances are the same in multiple types of data in multiple classifications, the LDA obtains an optimal Bayesian classifier.

Optimal bayesian classifier:

h*(x)＝arg max p(c|x) (6)

for each sample x, one can choose to maximize the posterior probability p (c|x) (x is the sample sum, c is the class label) if c (class label) is required such that the value p (c|x) is large, i.e., the center distance of two training samples belonging to different classes divided by the center distance value of the training samples of the same class is large, i.e., the maximum value of equation (5). Furthermore, (4) and (5) are equal. Thus, the LDA may build a bayesian optimal classifier in this case. The input data in the LDA classification is RSS data (training samples) of each AP at a reference point, and is output as a floor where a pedestrian is located (prediction result). The floor positioning is abstracted into a multi-classification problem, sample data are RSS data of each AP at each reference point of different floors, and a specific floor classification result is obtained through the multi-classification prediction model.

Corner detection and pedestrian upstairs and downstairs movement identification

When a pedestrian walks in the stairwell, the data of the inertial sensor of the mobile phone is acquired through the phyphox software, so that the data of the angular speed, the acceleration and the barometer of the pedestrian during walking are obtained, and the heading data and the state data (upstairs and downstairs and the static state) of the pedestrian in the stairwell are obtained. Course angle calculation adopts the formula (5):

head (k) is the calculated heading angle, head (0) is the initial heading angle, k is the number of acquisitions, w _i Is angular velocity data and Δt is the sampling period. The course angle transformation detection algorithm adopts the following formula (6):

head(k+τ)-head(k)＞Δh _thr (8)

first, a detection window tau and a threshold deltah are set _thr Detecting whether the variation of course angle in window is greater than threshold value, if it is greater than threshold value, then adding one, making the pedestrian pass through one corner, utilizing corner number and stairwell structure to obtain dynamic floor, because most of stairwell corner numbers are 1 and 2, said section mainly can be used for researching relationship of corner number and floor of stairwell, and the relationship of the corner number of stairwell passed by pedestrian and floorIs shown in Table 1 below:

TABLE 1 relationship between number of stair corners and floors

Number of stairwell corners	Layer 1	2 layers	Layer 3
				1	0	1	3
2	0	2	5

The following formula represents the relationship of the number of floors, the total number of corners and the number of corners between each floor, and the formula (9) represents the relationship of the total number of corners and floors that a pedestrian passes when the corner of the adjacent floor is 1:

FIG. 3 shows that the barometer is easy to generate abnormal fluctuation in the moving process of going upstairs and downstairs, and the air pressure of the downstairs curve in the drawing is abnormal fluctuation between 300 and 400 on the time axis, so that misjudgment can be generated on the moving state of pedestrians. And introducing barometer fusion acceleration step frequency detection to obtain the upstairs and downstairs states of the pedestrians, namely comparing the variation of the barometer only when the pedestrians are detected to stride, and determining the upstairs and downstairs states of the pedestrians. The zero-crossing comparison algorithm is adopted to detect the step frequency, the acceleration of the pedestrian crossing a step can be zero in the process of rising the acceleration of the pedestrian crossing a wave crest, falling the acceleration of the pedestrian crossing a wave trough, and the pedestrian crossing a step is considered to be crossing the pedestrian crossing once the zero moment is detected.

Dynamic floor recognition algorithm

The online stage is divided into two steps of initial floor discrimination and online floor discrimination, firstly, RSS signal values are put into an LDA multi-floor classification model trained in the offline stage to obtain the initial floor.

When an initial floor is obtained, pedestrians can pass through a plurality of 180-degree corners when walking through stairs, corresponding relations exist between the number of the corners and the floors, the current floor of the pedestrians can be obtained by judging the number of times of direction change of the pedestrians, the up-and-down floor state of the pedestrians is obtained by integrating step frequency detection through barometers, and a dynamic floor recognition algorithm is shown in figure 4. The method comprises the steps of detecting acceleration and barometer data of a pedestrian, judging the motion state of the pedestrian according to a step frequency fusion air pressure difference, recording the number of corners by taking the motion state as a precondition to judge the current floor of the pedestrian, calculating according to a calculation formula of the relation between the total number of the corners and the floor, if the calculation result is an integer, indicating that the pedestrian just passes through the corner of the stair, judging whether the pedestrian goes upstairs/downstairs or is stationary by identifying the state, if the state is upstairs/downstairs, keeping the floor result, and if the state is stationary, subtracting 0.5 layer from the calculation floor, namely, the pedestrian is on a landing between the stairs.

The method of the invention is verified by specific experimental simulation.

Experimental environment

This particular embodiment collects and analyzes RSS values of 9 APs for initial floor discrimination by AP selection in a three-tier laboratory. The un-circled WiFi in fig. 5 (b) represents the reference point for data acquisition, and the circled WiFi represents the distribution of recorded APs. Wherein apx_y is the number of AP, x represents the floor, y represents what number of AP, table 1 reflects the difference value of average signals of each AP in different floors, data are collected and divided according to experimental environment, and the experimental environment is shown in fig. 2.

And acquiring the average value of multiple RSS signal data of each AP of each reference point, reducing random errors, assigning a value to a signal which cannot be received when acquiring the data, and setting the value to be-110 (dBm). Table 1 reflects the mean (dBm) of RSS data for a portion of APs at adjacent floors.

AP	3_floor(avg)	4_floor(avg)	5_floor(avg)
				AP1_1	-73.72	-92.98	-107.21
AP2_1	-87.50	-74.13	-92.39
				AP3_1	-58.31	-73.11	-94.10

Table 1 average value of partial AP at each floor

The difference value of the RSS data between adjacent floors is-13.37 (dBm) to-20.89 (dBm), and the difference of the RSS data mean values between different floors is obvious, so that the floor judgment is facilitated.

The invention collects three-layer WiFi signals of the experiment No. two building for simulation verification, each layer is about 60 square meters (20 multiplied by 3 meters), and the total area of the three layers is 180 square meters. The data acquisition environment is shown in fig. 2, 155 reference points are arranged on each layer, the square size is 0.6x0.6 meter, RSS data of the selected AP on each reference point are acquired, the average value is acquired for a plurality of times, and the MEMS sensor signals during pedestrian movement are acquired to analyze the dynamic floors.

Static floor recognition algorithm simulation

The LDA multi-classification model is trained by the data acquired above, and the training accuracy is 97.6%. To verify the accuracy of the model, twenty points were randomly tested at each floor for a total of sixty reference points, and the results were shown in fig. 6.

The fact that the points at the end of the fifth building and the beginning of the fourth building have a certain degree of misjudgment can be seen through the figure 6, the fact that the Wi-Fi signals are similar in the characteristics of the RSS signal values of the last point of the fifth building and the beginning of the fourth building results in misjudgment is caused, and the fact that the signal characteristics of all the reference points are obviously only one point in other positions is wrong in point identification. In general, the data LDA obeying Gaussian distribution can be well classified, the LDA algorithm is simple, the training time is short and easy to realize, the result shows that the classification accuracy is high, the static floor recognition accuracy is 96.4%, and the initial floor data can be provided for dynamic floor recognition.

Dynamic floor recognition algorithm simulation

Pedestrians can be in a static state when going up and down stairs, such as thinking, smoking, making video calls and the like. Therefore, it is necessary to consider the influence of the stationary state on the floor determination. When temperature and humidity change rapidly, the floor height estimation method based on only air pressure is likely to generate a large error. Therefore, the invention adopts the step frequency fusion air pressure difference to judge the motion state of the pedestrian, and records the number of corners by taking the motion state as a precondition to judge the current floor of the pedestrian. Heading angle and acceleration-based step frequency detection is shown in fig. 7 below.

Fig. 7 (a) shows that the pedestrians passing through from the fifth floor to the third floor pass through 3 stair corners, so that there are three course changes of about 180 degrees, the accuracy of the corner detection shown in fig. 8 is 100%, and (b) shows the time point of the pedestrian stepping in the process, and the total 65 steps are obtained in the process of going down the stairs through the step frequency detection algorithm, so that the accuracy of 66 steps is 98.5%, and the accuracy can provide a basis for judging whether the pedestrian steps.

Fig. 8 is a comparison of the number of corners and the actual number of corners using the sliding window detection algorithm of equation (7), in which the detection window τ=100 and the threshold Δh are set, since the data frequency is 50Hz, and experiments show that the pedestrian passes a corner time of about two seconds _thr =180, the sliding distance of the window is 100, and whether the maximum variation of the heading angle in the window is greater than a threshold is detected to determine whether the corner is passed.

Fig. 9 is a graph showing the variation of the air pressure with the number of steps when going upstairs and downstairs, wherein the average value of the variation of the air pressure in each five steps is 0.1523hPa, the maximum and minimum values are 0.179 and 0.104, the minimum value is taken as a threshold value, the variation of the air pressure when going upstairs is negative, the variation of the air pressure when going downstairs is positive, and the current upstairs and downstairs states of the exercise can be judged by the variation of the air pressure in five steps.

Analysis of floor positioning results under different initial floors the experiment respectively uses third floor, fourth floor and fifth floor as initial floors to respectively make up-building and down-building experiments, and the experiment is used for verifying accuracy of floor calculation based on corner detection under ideal condition, the accuracy of the experiment result is 100%, and the experiment result is shown in fig. 10.

The x-axis of fig. 10 is the travel time, the y-axis is the floor, fig. 10 (a) is the change of the floor of the pedestrian moving from the third floor to the fifth floor, (b) is the process from the fifth floor to the third floor, and (c) and (d) are the processes of going up and down to the third floor and the fifth floor respectively with the fourth floor as the initial floor.

Fig. 10 is a diagram showing a floor determination in an ideal state, and does not consider the state of a pedestrian in a stairwell (going up and down a stairs, stopping), so that a floor positioning analysis with a plurality of activities is performed. The experiment takes a third building as a starting floor, and goes upstairs, downstairs and stairs corners to stop during the period of reaching a fifth building. The experimental results are shown in fig. 11 below:

in the above experiments the floor position of the pedestrian was calculated with the proposed method and compared with the actual floor position. Fig. 10 shows that the algorithm of the invention can obtain the specific position of the pedestrian when the stairwell is stationary, and the recognition rate of the floor is high. The triangle line in the figure shows the real position, the circle line shows the traditional fingerprint matching algorithm, the detection of the floor change in the floor identification is not linear enough, and the target floor is reached in advance. The floor result can be obtained after the corner is detected, and the pedestrian does not reach the target floor in the last section of stairs which reach the target floor, so that the problem can be solved by prolonging the detection time of the air pressure, the whole detection result is not influenced, and the specific floor of the pedestrian in the process of going up and down the floor can be accurately obtained.

Assuming in the experiment that the initial floor is already determined, the invention provides a dynamic floor identification method based on corner detection. The 3 aspects of positioning accuracy, fingerprint positioning times, and whether to detect the platform position were compared with other floor methods as shown in table 2.

Table 2 comparison of system performance

MA-LDA reaches higher floor recognition accuracy, because the algorithm itself limits to building different classifiers for different floors, the calculation amount is large, and two fingerprint positioning is needed in the course of changing the pedestrian floor and the stair platform cannot be recognized. The EAP-LDA method achieves higher accuracy, but the online positioning time is longer because the AP needs to be selected in the online stage, which is unfavorable for deployment on the terminal equipment. The BPFL method is optimized in fingerprint positioning times and online positioning time, and can also be used for positioning the platform position, but the floor and platform recognition accuracy is lower due to the fact that the floor estimation is carried out by adopting the barometric method to calculate the height. Compared with the methods, the method has the advantages that the identification rate of the target floor is greatly improved, one-time online fingerprint positioning is reduced, and the position of the stair platform can be accurately positioned.

The invention provides a dynamic floor identification method based on corner detection. The floor and the landing where the pedestrian is can be accurately identified in the floor identification under various activity conditions. Experimental results show that the LDA multi-classification algorithm is used for initial floor recognition, the training precision is 97.8%, and the testing precision is 96.7%. Under the condition of determining the initial floor, the current floor is judged by detecting corners in the walking process of pedestrians, and the recognition accuracy is 100%. Compared with the traditional fingerprint floor positioning, the method can accurately identify the position of the platform where the pedestrian is located under the condition that the fingerprint database of the landing area of the stairwell is not increased, provides a new measuring method and a new estimating means for floor identification in indoor positioning, and improves the positioning problem of the landing area of the stairwell.

The foregoing is merely illustrative of the present invention, and the present invention is not limited thereto, and any changes or substitutions easily contemplated by those skilled in the art within the scope of the present invention should be included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A dynamic floor positioning method based on multiple sensors is characterized in that: the method comprises the following steps:

step 1) construction of fingerprint database

step 3) initial static floor discrimination

2. A multi-sensor based dynamic floor positioning method according to claim 1, characterized in that: the LDA multi-classification model in the step 2) is as follows:

Wherein,is generalized Rayleigh Li Shang, whose maximum is matrix +.>The product of the maximum d values expressed by the formula (5) is the matrix +.>The product of the maximum d eigenvalues, the corresponding matrix W is the projection matrix of the eigenvector corresponding to the maximum d eigenvalues; using projection matricesAfter W projection, a probability density function, a mean vector and a variance matrix of Gaussian distribution of fingerprint data of the AP set on the ith floor can be obtained;

h*(x)＝argmaxp(c|x) (6)

3. A multi-sensor based dynamic floor positioning method according to claim 2, characterized in that: the course angle calculation in the step 4) adopts the formula (5) when the pedestrian walks in the stairwell

head(k+τ)-head(k)＞Δh _thr (8)

first, a detection window tau and a threshold deltah are set _thr Detecting whether the variation of course angle in the window is larger than a threshold value, if so, counting and adding one, and obtaining the pedestrian by using the number of times of the corner and the structure of the stairwellDynamic floors.

4. A multi-sensor based dynamic floor positioning method according to claim 3, characterized in that: when the corner of the stairwell is 1, calculating the relation between the total number of the corners and the floors:

5. A multi-sensor based dynamic floor positioning method according to claim 1, characterized in that: the specific rule for judging the pedestrian motion state in the step 4) is as follows: the comparison of the barometer variation is only carried out when the pedestrian is detected to stride, so that the pedestrian upstairs and downstairs states are determined, if the barometer variation is positive, the pedestrian states are downstairs, and if the barometer variation is negative, the pedestrian states are upstairs.

6. The multi-sensor based dynamic floor positioning method according to claim 5, wherein: the zero-crossing comparison algorithm is adopted to detect the step frequency, the acceleration of pedestrians crossing one step can be zero in the process of rising, wave crest, wave bottom and wave trough, and the moment of zero is detected once to consider that the pedestrians cross one step.