CN107884744B - Passive indoor positioning method and device - Google Patents
Passive indoor positioning method and device Download PDFInfo
- Publication number
- CN107884744B CN107884744B CN201710951127.5A CN201710951127A CN107884744B CN 107884744 B CN107884744 B CN 107884744B CN 201710951127 A CN201710951127 A CN 201710951127A CN 107884744 B CN107884744 B CN 107884744B
- Authority
- CN
- China
- Prior art keywords
- sample
- coefficient
- equal
- determining
- frequency wavelet
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S5/00—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations
- G01S5/02—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations using radio waves
- G01S5/0252—Radio frequency fingerprinting
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Position Fixing By Use Of Radio Waves (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
A passive, indoor positioning method, comprising: taking the RSS value of a radio frequency network link collected by a target at each coordinate in a non-open room as a training sample, and taking a coordinate number as a sample label; performing two-dimensional double-correlation distributed wavelet filtering processing on the training sample to determine a filtered training sample; establishing an Adaboost.M2 ensemble learning model based on a Kini decision tree, training by adopting a filtered training sample and a sample label, and determining the trained model; and acquiring an RSS value of the target moving in the monitoring area at will, taking the RSS value as a test sample, performing two-dimensional dual-correlation distributed wavelet filtering on the test sample, inputting the processed test sample into the trained model, and determining a positioning result. The method can distinguish and filter RSS sample noise and random interference, retain normal jump data, has strong generalization capability in the positioning process, and can improve positioning accuracy and stability on the whole. The invention also provides a corresponding device.
Description
Technical Field
The invention relates to the technical field of indoor personnel target positioning of a wireless sensing network, in particular to the technical field of passive indoor target positioning based on radio frequency received signal strength in a non-spacious complex indoor environment.
Background
The indoor positioning method of the personnel target based on the wireless radio frequency sensing network does not need to actively cooperate and carry an electronic tag by the positioned target, and the principle is as follows: when a human body target is in a monitored area, a related radio frequency link in a network can be shielded or interfered, so that shielding is generated, namely, a shadow effect is generated, and therefore a Received Signal Strength (RSS) value of a network sensor node is influenced, and the position of the target can be judged by collecting and analyzing the RSSI sample condition. Due to the characteristics of the radio frequency signal, the method has the advantages of insensitivity to environmental temperature, humidity, light and non-metallic barriers, convenience in networking, low cost and the like, so that the method has wide application prospects in the fields of intelligent building monitoring, military security monitoring, old monitoring, patient monitoring and the like.
The positioning method mainly comprises two technical implementation schemes at present: the first category is machine learning based solutions. For example, the method is realized by using algorithms such as an artificial neural network, fingerprint matching, a support vector machine and the like, coordinates need to be manually calibrated in a monitoring area, RSS data of a sensor network is collected to serve as a training sample to train a machine learning model, and positioning is carried out through test data in an online stage. The second type is a positioning scheme based on a model of human body influence on the radio frequency signal link. The scheme does not need off-line learning, and mainly finds an ideal model of influence of a human body on the strength of a received signal of a corresponding link, such as an ellipse model, a saddle model and the like. In the positioning process, corresponding parameters are deduced through an influence model and an RSS value, and then the target is positioned by matching with the technologies of particle filtering, Bayesian probability, radio frequency tomography and the like. However, in a non-open environment, such as a common indoor environment, the structure of the area varies, the material of the furniture is varied, and the shapes of the furniture are different, so that there is a great difference in the influences of various constructive or non-constructive fades, reflections, scattering, etc. caused by radio frequency signals of different links, and the current positioning method has the following problems in this case:
(1) the RSS samples do not filter well. When the monitoring area is relatively complex, the RSS data is affected by various noises and random interferences, and the intensity varies, and some samples even have strong distortion. The moving speed of the detected target is not constant, so that some signals and noise time-frequency characteristics are mixed and difficult to distinguish. The traditional filtering method has unstable effect, the strict filtering condition can influence the characteristics of data and even pollute effective data, and the loose condition can leave noise and distortion points, which can influence the positioning accuracy.
(2) The accuracy and generalization capability of the machine learning positioning mode need to be improved. When the environment is complex, the fixed link influence model is difficult to correctly express the relation between the RSS sample change and the target position change, and a machine learning mode is a good choice. However, in a complex environment, the relationship between link change and a target position is complex, the importance of each link is difficult to determine, the node weight distribution is not fixed, the quantity of collected training samples is limited, and the types of the samples are not comprehensive. However, since the generalization capability of the trained single machine learning model is not strong and the adaptability is poor, when the target is located at a position between certain training sample coordinates, namely the target is located at a non-calibration coordinate, the target is likely to be located on a coordinate far away from the real position relative to the surrounding coordinates, thereby affecting the locating effect. For example, when the target is located between coordinates 3 and 4, the model with poor generalization ability and adaptability is likely to locate the target to the distant coordinate 10 or 13, and the actual expectation is that the model can output the location result as the coordinate 3 or 4.
Disclosure of Invention
Technical problem to be solved
The invention aims to provide a passive indoor positioning method and a passive indoor positioning device, which are used for solving at least one of the problems of poor filtering effect on RSS samples, low positioning accuracy and low generalization capability of a machine learning method in a complex environment, namely a non-open environment, such as a common indoor environment, of the conventional RSS-based passive indoor positioning technology.
(II) technical scheme
The invention provides a passive indoor positioning method, which comprises the following steps:
step A, taking a non-open indoor environment as a monitoring area, taking all acquired RSS values of radio frequency network links as training samples when a target is at each coordinate in the monitoring area, and taking coordinate numbers as sample labels;
b, performing two-dimensional double-correlation distributed wavelet filtering processing on the training sample to determine the filtered training sample;
step C, establishing an Adaboost.M2 ensemble learning model based on a Kini decision tree, training by using the filtered training samples and sample labels, and determining the trained ensemble learning model; and
and step D, collecting all radio frequency network link RSS values when the target moves in the monitoring area at will, using the RSS values as test samples, carrying out two-dimensional double-correlation distributed wavelet filtering on the test samples, determining the filtered test samples, inputting the filtered test samples into the trained integrated learning model, and determining a positioning result.
In some embodiments of the invention, step B comprises the sub-steps of:
substep B1, forming training samples into a sample matrix, wherein each column represents a training sample, the number of columns is the number of samples, each row represents the RSS value of the same dimension in different samples, the sample matrix is subjected to wavelet decomposition of each dimension data according to the dimension, a wavelet function selects db1 wavelet, and a low-frequency wavelet coefficient ca and a h-th layer high-frequency wavelet coefficient cd are obtainedhWherein, one dimension refers to a radio frequency link, h is more than or equal to 1 and less than or equal to s, s is the number of layers of wavelet decomposition, and s is more than or equal to 2;
substep B2, retaining low frequency wavelet coefficient ca, and utilizing high frequency wavelet coefficients of h layer and h +1 layer to obtain two-dimensional double correlation longitudinal correlation coefficient corrhDetermining the high-frequency wavelet coefficient energy PcdhAnd longitudinal correlation coefficient energy PcorrhDetermining a normalized longitudinal correlation coefficient corrnh；
Substep B3 of comparing the high frequency wavelet coefficients cdhSetting the high-frequency wavelet coefficient greater than the normalized longitudinal correlation coefficient to zero, retaining the high-frequency wavelet coefficient less than or equal to the normalized longitudinal correlation coefficient, retaining all the high-frequency wavelet coefficients of the last layer, and finally determining the retained high-frequency wavelet coefficient cdi；
Sub-step B4, dividing the preserved high frequency wavelet coefficient into U segments, U is more than or equal to 2, each segment selects d data, d is positive integer, and the preserved high frequency wavelet coefficients of the j segment and the j +1 segment are respectively utilized according to layers to calculate the transverse non-time-shift correlation coefficient R1 in the two-dimensional double correlationijJ is more than or equal to 1 and less than or equal to U-1, the high-frequency wavelet coefficient reserved in the j +1 segment is subjected to time shift, wherein the time shift is a, a is a positive integer less than or equal to d/2, and the transverse time shift correlation coefficient R2 is calculated according to layers respectivelyijDetermining the difference value Rm between the time-shift correlation coefficient and the non-time-shift correlation coefficientijAnd selecting each layer R1ijR1 with the smallest absolute valueijCorresponding data segment cdriAnd each layer RmijRm with the smallest absolute valueijCorresponding data segment cdrmiCombined to filter threshold parameter estimation data cdmi；
Substep B5, according to cdmiDetermining a filter threshold parameter sigma, then determining a filter threshold thr, performing distributed filtering on the reserved high-frequency wavelet coefficient, and obtaining a filtered high-frequency wavelet coefficient cdfi(ii) a And
substep B6 of using the low frequency wavelet coefficients ca and the filtered high frequency wavelet coefficients cdfiAnd performing wavelet reconstruction to obtain a filtered training sample.
In some embodiments of the present invention, the normalized longitudinal correlation coefficient is determined in sub-step B2 according to the following formula:
corrh＝cdh·cdhh+1
wherein n is a high-frequency wavelet coefficient cdhThe length of the sequence is that g is more than or equal to 1 and less than or equal to n, and g is a high-frequency wavelet coefficient cdhSequence number of data in the sequence.
In some embodiments of the present invention, in sub-step B4, R1 is determined according to the following formulaij、R2ij、 RmijAnd selecting cdri、cdrmi、cdmi：
Rmij＝|R1ij-R2ij|
cdrmi＝cdij，
cdri＝cdijIf, if
cdmi＝{cdrmi，cdri}
Wherein cdij+1Is cdijNext adjacent non-time-shifted fragment, cd'ij+1Is cdij+1Right shift by a time shift amount to form a time shift small segment, and Cov is cdij+1And cdijCovariance between, Var is cdij+1And cdijThe variance between.
In some embodiments of the present invention, the sub-step B5 determines the filtering threshold parameter σ and the threshold thr according to the following formulas, and performs distributed wavelet filtering to obtain the filtered high frequency wavelet coefficient cdfi：
Wherein q is more than or equal to 2 and less than or equal to s, and L reserves high-frequency wavelet coefficient cd for each layeriL is U × d, mean is an intermediate value,y is the serial number of each data in each layer reserved high-frequency wavelet coefficient, and y is more than or equal to 1 and less than or equal to L.
In some embodiments of the invention, step C comprises the sub-steps of: and a substep C1 of determining the Gini coefficient of each attribute in the training sample, and selecting the attribute A corresponding to the maximum value of the Gini coefficient as the optimal attribute to perform the growth of the decision tree, wherein the formula of the Gini coefficient Gini (A) is as follows:
wherein V is the branch number of the decision tree, N is the number of coordinates, S'bcSample subset S 'sorted with A as splitting attribute'cThe number of samples belonging to class b in the group, E is the number of total samples at the splitting point, and b is more than or equal to 1 and less than or equal to N; and a substep C2 of determining the trained ensemble learning model based on the decision tree in the substep C1:
iteratively updating sample weights:
determining the error rate of the decision tree:
determining weights of the decision tree:
wherein, W0(f) Is the initial weight of the f-th sample, N is the number of coordinates, M is the number of training samples contained in each coordinate, Wk(f, z) is the sum of the probabilities that in the k-th iteration sample f is classified into all error classes z, k is the number of iterations, hk(xf，zf) Probability of classifying the f-th sample for the k-th decision tree into its correct class zf, hk(xfZ) the kth decision tree classifies the f sampleTo remove its correct class zfExcept for the possibility of all other incorrect categories z,kis the error rate of the kth decision tree, akIs the weight of the kth decision tree and Sum is the Sum.
In some embodiments of the present invention, sub-step B1 further includes the steps of: and taking all the radio frequency network link RSS values acquired when no target exists in the monitoring area as reference samples, carrying out difference processing on the training samples and the reference samples, and determining difference signals to be used as the training samples.
Based on the same inventive concept, the invention also provides a passive indoor positioning device, comprising:
a memory to store instructions; and
and the processor is used for executing the passive indoor positioning method according to the instruction.
(III) advantageous effects
Compared with the prior art, the passive indoor positioning method and the passive indoor positioning device provided by the invention at least have the following advantages:
1. the RSS data samples are subjected to longitudinal and transverse two-dimensional double-correlation distributed wavelet filtering processing, noise of the RSS samples is distinguished and filtered in a targeted mode, normal jump data mixed in high-frequency noise and generated due to target movement are reserved as far as possible, meanwhile, the RSS data filtering method has a good filtering effect on abnormal jump points of the data and the high-frequency noise, and has a two-dimensional double-correlation self-adaptive filtering threshold value.
2. The Adaboost.M2 integrated learning model based on the kini decision tree solves the problem of poor classification effect of targets at non-calibrated coordinates due to poor generalization ability and poor adaptability of a single machine learning model, and can reasonably distribute and select weights of sample characteristics and attributes under the conditions that the relation between link changes and target positions is complex and the importance of each link is difficult to determine, so that multiple groups of learning models are trained to improve the generalization ability and the positioning accuracy.
Drawings
Fig. 1 is a schematic step diagram of a passive indoor positioning method according to an embodiment of the present invention.
Fig. 2A is a schematic view of the monitoring environment after coordinate division according to embodiment 1 of the present invention.
Fig. 2B is a schematic view of the monitoring environment after coordinate division according to embodiment 2 of the present invention.
Fig. 3A is a plot of the localization effect of the non-filtered and kuney decision tree-based adaboost. m2 ensemble learning model of example 1.
Fig. 3B is a graph of the localization effect of the unfiltered and deep neural network DNN-based model of example 1.
FIG. 3C is a diagram of the positioning effect of embodiment 1 without filtering and based on the fingerprinting model.
Fig. 3D is a diagram of the positioning effect of the non-filtered SVM-based positioning method of embodiment 1.
Fig. 3E is a plot of the localization effect of the non-filtered and kuney decision tree-based adaboost. m2 ensemble learning model of example 2.
Fig. 3F is a graph of the localization effect of the unfiltered and deep neural network DNN-based model of example 2.
FIG. 3G is a diagram of the positioning effect of embodiment 2 without filtering and based on the fingerprinting model.
Fig. 3H is a diagram of the positioning effect of the unfiltered SVM-based positioning method of embodiment 2.
Fig. 4A is a positioning effect diagram of the two-dimensional double correlation distributed wavelet filtering and the macaroni decision tree-based adaboost. m2 ensemble learning model of embodiment 1.
Fig. 4B is a plot of the localization effect of the unassociated threshold wavelet filtered and kuney decision tree-based adaboost. m2 ensemble learning model of example 1.
Fig. 4C is a diagram of localization effects of the correlation wavelet entropy filtering and the kuney decision tree-based adaboost. m2 ensemble learning model of example 1.
Fig. 4D is a plot of the localization effect of the high-frequency coefficient all-zero wavelet filter and the gibbost. m2 ensemble learning model based on the kini decision tree of embodiment 1.
Fig. 4E is a positioning effect diagram of the two-dimensional double-correlation distributed wavelet filtering and the macaroni decision tree-based adaboost. m2 ensemble learning model of embodiment 2.
Fig. 4F is a plot of the localization effect of the unassociated threshold wavelet filter of example 2 and based on the adaboost. m2 ensemble learning model of the kini decision tree.
Fig. 4G is a positioning effect diagram of the correlation wavelet entropy filtering and the macaroni decision tree-based adaboost. m2 ensemble learning model of example 2.
Fig. 4H is a plot of the localization effect of the high-frequency coefficient all-zero wavelet filter and the gibbost. m2 ensemble learning model based on the kini decision tree of example 2.
Fig. 5A is a positioning effect diagram of the two-dimensional double correlation distributed wavelet filtering and the macaroni decision tree-based adaboost. m2 ensemble learning model of embodiment 1.
Fig. 5B is a graph of localization effect of the two-dimensional double correlation distributed wavelet filtering and based on the deep neural network DNN model of embodiment 1.
Fig. 5C is a diagram of the two-dimensional dual correlation distributed wavelet filtering and positioning effect based on the fingerprint identification model of embodiment 1.
Fig. 5D is a diagram of the two-dimensional double correlation distributed wavelet filtering and the positioning effect based on the SVM model of embodiment 1.
Fig. 5E is a positioning effect diagram of the two-dimensional double-correlation distributed wavelet filtering and the macaroni decision tree-based adaboost. m2 ensemble learning model of embodiment 2.
Fig. 5F is a map of the localization effect of the two-dimensional double correlation distributed wavelet filtering of example 2 based on the deep neural network DNN model.
Fig. 5G is a diagram of the two-dimensional double correlation distributed wavelet filtering and the positioning effect based on the fingerprint identification model of embodiment 2.
Fig. 5H is a diagram of the localization effect of the two-dimensional double correlation distributed wavelet filtering of the embodiment 2 based on the SVM model.
Fig. 6 is a schematic structural diagram of a passive indoor positioning apparatus according to an embodiment of the present invention.
Detailed Description
In order to overcome the defects that the existing positioning method has poor filtering effect of RSS samples and the accuracy and generalization capability of a machine learning positioning mode need to be improved, the invention provides a passive indoor positioning method which is mainly suitable for more complicated indoor environments, namely non-open environments, for example, common indoor environments, filtering the acquired RSS sample data by using a two-dimensional dual-correlation wavelet filtering method, and training an Adaboost.M2 integrated learning model based on a kini decision tree by using the filtered data to obtain the trained model. And finally, collecting the RSS value of the target moving in the monitoring area as a test sample, carrying out two-dimensional double-correlation distributed wavelet filtering processing on the test sample, inputting the processed test sample into the trained model, and determining a positioning result. Generally, in the passive positioning method, the target can be directly positioned without carrying an electronic tag and a signal transceiver.
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to specific embodiments and the accompanying drawings.
The embodiment of the invention provides a passive indoor positioning method, as shown in figure 1, the method comprises the following steps:
and step A, taking a non-open common indoor environment as a monitoring area, taking all acquired RSS values of the radio frequency network links as training samples when a target is at each coordinate in the monitoring area, and taking the coordinate number as a sample label. Wherein the determination of the coordinates comprises the steps of: selecting any position as an initial coordinate on the ground of a barrier-free area in the non-open indoor environment, and setting a coordinate at a preset distance from the initial coordinate along the longitudinal direction and the transverse direction of the initial coordinate, wherein the coordinate comprises the initial coordinate. Generally, the training samples can be acquired by arranging a plurality of radio frequency sensors around the monitoring area, the heights of the radio frequency sensors are generally 1.5m to 2m (meters), and the heights of the radio frequency sensors are preferably the same.
Let the number of coordinates be N, where N is 18 in example 1 and 12 in example 2. M training samples are collected under each coordinate, M is 480 in two embodiments, and R is the training sample collected at the I coordinateIJ＝{rssIJ1，rssIJ2，......，rssIJH}(I＝1, 2.. N, J ═ 1, 2.. M), H is the number of links, in this embodiment, H ═ 48, and then the I point training sample set is RI＝{RI1，RI2，......RIM}，RICorresponding label set as YI＝{I1，I2，......IMAll collected at the same coordinate I point, so I1＝I2＝...＝IMI. The training sample population may be represented as R ═ R1，R2，......RNY, its corresponding label Y ═ Y1，Y2，......YN}。
Next, the coordinate determination is exemplified by combining embodiment 1 and embodiment 2, fig. 2A is a schematic view of the entire monitoring environment after the coordinate division of embodiment 1 of the present invention, fig. 2B is a schematic view of the entire monitoring environment after the coordinate division of embodiment 2 of the present invention, as shown in fig. 2A to fig. 2B, the monitoring areas of the two embodiments are general indoor environments with obstacles, and the furniture size, placement, material and the like of the two embodiments are different.
The number of coordinates in example 1 was 18, and the numbers were 1 to 18, respectively. It should be noted that, in order to reduce the influence caused by interference, all the RSS values of the radio frequency network link acquired when there is no target in the monitored area may be used as a reference sample, and the training sample and the reference sample are subjected to a difference processing to determine a difference signal, which is used as the training sample. Therefore, in this embodiment, a total of 19 training data samples are added when the target is not located in the monitoring area. Similarly, the number of the coordinates in example 2 is 12, and the numbers are 1-12 respectively, and in addition, the situation when the monitoring area has no target, 13 training samples are provided. When training samples are taken, the target stands at each coordinate point, respectively, and 480 samples are taken at each coordinate point, each sample consisting of 48 dimensions (number of links between sensors communicating with each other).
And B, performing two-dimensional double-correlation distributed wavelet filtering processing on the training sample to determine the filtered training sample.
Wherein, step B mainly comprises the following substeps:
substep B1, forming training samples into a sample matrix, wherein each column represents a training sample, the number of columns is the number of samples, each row represents the RSS value of the same dimension in different samples, the sample matrix is subjected to wavelet decomposition of each dimension data according to the dimension, a wavelet function selects db1 wavelet, and a low-frequency wavelet coefficient ca and a h-th layer high-frequency wavelet coefficient cd are obtainedhWherein, one dimension refers to a radio frequency link, h is more than or equal to 1 and less than or equal to s, s is the number of layers of wavelet decomposition, and s is more than or equal to 2; the number of layers of the wavelet decomposition can be selected preferentially according to the data length of each dimension of the training sample matrix.
Substep B2, retaining low frequency wavelet coefficient ca, and utilizing high frequency wavelet coefficients of h layer and h +1 layer to obtain two-dimensional double correlation longitudinal correlation coefficient corrhDetermining the high-frequency wavelet coefficient energy PcdhAnd longitudinal correlation coefficient energy PcorrhDetermining a normalized longitudinal correlation coefficient corrnh；
corrh＝cdh·cdh+1
Wherein n is a high-frequency wavelet coefficient cdhThe length of the sequence is that g is more than or equal to 1 and less than or equal to n, and g is a high-frequency wavelet coefficient cdhSequence number of data in the sequence.
Substep B3 of comparing the high frequency wavelet coefficients cdhSetting the high-frequency wavelet coefficient larger than the normalized longitudinal correlation coefficient to zero, retaining the high-frequency wavelet coefficient smaller than or equal to the normalized longitudinal correlation coefficient, retaining all the high-frequency wavelet coefficients of the last layer, and finally determining to retain the high-frequency waveletCoefficient cdi；
And a sub-step B4 of dividing the preserved high-frequency wavelet coefficient into U sections, wherein U is more than or equal to 2, each section selects d data, d is a positive integer, and d is more than or equal to 10 and less than or equal to 20. Then, the reserved high-frequency wavelet coefficients of the j-th segment and the j + 1-th segment are respectively utilized according to layers to calculate a transverse non-time-shift correlation coefficient R1 in the two-dimensional double correlationijJ is more than or equal to 1 and less than or equal to U-1, the j +1 th segment of the reserved high-frequency wavelet coefficient is subjected to time shift, wherein the time shift is a, a is a positive integer less than or equal to d/2, and the transverse time shift correlation coefficient R2 is calculated according to layers respectivelyijDetermining the difference value Rm between the time-shift correlation coefficient and the non-time-shift correlation coefficientijAnd selecting each layer R1ijR1 with the smallest absolute valueijCorresponding data segment cdriAnd each layer RmijRm with the smallest absolute valueijCorresponding data segment cdrmiCombined to filter threshold parameter estimation data cdmi. R1 is determined according to the following formulaij、R2ij、 RmijAnd selecting cdri、cdrmi、cdmi：
Rmij＝|R1ij-R2ij|
cdrmi＝cdijIf, if
cdri＝cdijIf, if
cdmi＝{cdrmi，cdri}
Wherein cdij+1Is cdijNext adjacent non-time-shifted fragment, cd'ij+1Is cdij+1Right shift by a time shift amount to form a time shift small segment, and Cov is cdij+1And cdijCovariance between, Var is cdij+1And cdijThe variance between.
Substep B5, according to cdmiDetermining a filter threshold parameter sigma, then determining a filter threshold thr, performing distributed filtering on the reserved high-frequency wavelet coefficient, and obtaining a filtered high-frequency wavelet coefficient cdfi。
In sub-step B5, the filtering threshold parameter sigma and threshold thr are determined according to the following formula, and distributed wavelet filtering is performed to obtain the filtered high-frequency wavelet coefficient cdfi：
Wherein q is more than or equal to 2 and less than or equal to s, and L reserves high-frequency wavelet coefficient cd for each layeriThe length of the high-frequency wavelet coefficient is equal to or greater than 1 and equal to or less than L, mean is an intermediate value, y is the serial number of each data in each layer reserved high-frequency wavelet coefficient, and y is equal to or greater than U multiplied by d.
Substep B6 of using the low frequency wavelet coefficients ca and the filtered high frequency wavelet coefficients cdfiAnd performing wavelet reconstruction to obtain a filtered training sample.
And step C, establishing an Adaboost. M2 ensemble learning model based on the Kini decision tree, training by using the filtered training samples and sample labels, and determining the trained ensemble learning model.
The method comprises the following substeps:
and a substep C1 of determining the Gini coefficient of each attribute in the training sample, and selecting the attribute A corresponding to the maximum value of the Gini coefficient as the optimal attribute to perform the growth of the decision tree, wherein the formula of the Gini coefficient Gini (A) is as follows:
wherein V is the branch number of the decision tree, N is the number of coordinates, S'bcSample subset S 'sorted with A as splitting attribute'cThe number of samples belonging to class b in the group, E is the number of total samples at the splitting point, and b is more than or equal to 1 and less than or equal to N; and
and a substep C2 of determining the trained ensemble learning model based on the decision tree in the substep C1:
iteratively updating sample weights:
determining the error rate of the decision tree:
determining weights of the decision tree:
wherein, W0(f) Is the initial weight of the f-th sample, N is the number of coordinates, M is the number of training samples contained in each coordinate, Wk(f, z) is the sum of the probabilities that in the k-th iteration sample f is classified into all error classes z, k is the number of iterations, hk(xf，zf) Probability of classifying the f-th sample for the k-th decision tree into its correct class zf, hk(xfZ) the probability that the kth decision tree classifies the result of the f-th sample as all other incorrect classes z except its correct class zf,kis the error rate of the kth decision tree, akIs the weight of the kth decision tree and Sum is the Sum. In addition, generally speaking, the number of decision trees also represents the number of decision treesAnd (5) performing secondary iteration.
And D, collecting all radio frequency network link RSS values when the target moves randomly in the monitoring area, using the RSS values as test samples, carrying out two-dimensional dual-correlation distributed wavelet filtering on the test samples, determining the filtered test samples, inputting the filtered test samples into the trained integrated learning model, and determining a positioning result.
Next, the method provided by the present invention is tested and verified:
during the verification process, the test samples are collected when the human target continuously moves from any point of the monitoring area to any point, and each test sample can be composed of 48 dimensions (in other embodiments, different dimensions can be selected). Because the structure of the common indoor environment is varied, the material composition of furniture is varied and has different sizes, so that the phenomena of various construction or non-construction fading, reflection, scattering and the like caused by radio frequency signals are different, various noises and random interference exist in the environment, and the acquired data samples contain various noises and data mutation.
In the moving process of the target, the moving speed of the target can be changed at any time, so that the amount of samples collected at each position is different, the target can pass through the areas among certain coordinate points except the position where the target passes through the calibrated coordinates, and the samples collected in the areas where the target is located among the calibrated coordinates are the key points for verifying the generalization capability of the positioning model and the accuracy of the positioning method. For example, when the target is located between coordinates 3 and 4, and is affected by various factors, other positioning methods may position the target at a distant coordinate (e.g., coordinates 10 or 13) so that the positioning result has a large deviation, and it is actually expected that the model can output the positioning result as coordinates 3 or 4.
In the embodiments 1 and 2, in a common indoor environment, different positioning tests are performed by using the positioning method of the adaboost.m2 ensemble learning model based on the kini decision tree, and compared with 3 different positioning methods using a fingerprint model, an SVM model, and a deep neural network DNN model in the related positioning model in the field. The positioning effect with or without filtering processing is compared respectively; positioning effects under different wavelet filtering modes; and the positioning effect of different positioning models when the two-dimensional double correlation wavelet filtering method is adopted. The deep neural network model is debugged, a 5-layer full-connection DNN structure is selected, and regularization and dropout layers are used, so that the problems of local minimum and overfitting can be avoided as much as possible.
In the two embodiments, compared with embodiment 2, in embodiment 1, the environmental noise and data abnormal jump are less, and the data quality is higher, so that embodiment 2 can more clearly embody the filtering and generalization capability advantages of the passive indoor positioning method in the non-spacious ordinary indoor environment. The coordinate accuracy for both embodiments is 0.5 meters, so the experimental results will verify the positioning accuracy and stability with a resolution of 0.5 meters.
Fig. 3A to 3H are graphs of positioning results of the positioning method of the adaboost.m2 ensemble learning model based on the kini decision tree and the positioning methods of other models in a test sample when there is no filtering process, and table 2 is a comparison table of the positioning results of fig. 3A to 3H. The abscissa represents the test specimen numbers arranged in time series, and the ordinate represents the coordinate result of positioning. Wherein the test samples of example 1 and example 2 were taken when the target passed the calibration coordinates of 16-13-7-4-3-2-9-11 and 1-6-7-8-9-10-11, respectively. It can be seen from the figure that the Adaboost. M2 ensemble learning model based on the Kini decision tree has better stability than other models, and the accuracy (see Table 1) at the resolution of 0.5 m is obviously higher than that of the positioning method adopting other models.
TABLE 1
Fig. 4A to 4H are positioning effect graphs obtained by using the ensemble learning model of the present invention in different wavelet filtering modes, and table 2 is a positioning effect comparison table of fig. 4A to 4H. The abscissa represents the test specimen numbers arranged in time series, and the ordinate represents the coordinate result of positioning. The different wavelet filtering modes are two-dimensional double-correlation distributed wavelet filtering, single-correlation threshold wavelet filtering, correlation wavelet entropy filtering and high-frequency coefficient all-zero wavelet filtering of the invention, and the number of layers is set to be 3. It can be seen from the figure that the positioning method using the two-dimensional double correlation distributed wavelet filtering method of the present invention for filtering has higher positioning accuracy (see table 2) than the positioning method using other wavelet filtering methods at a resolution of 0.5 m, and has better stability.
TABLE 2
Fig. 5A to 5H are positioning result diagrams of different positioning models when the two-dimensional dual correlation distributed wavelet filtering method of the present invention is adopted, and table 3 is a positioning result comparison table of fig. 5A to 5H. The abscissa represents the test specimen numbers arranged in time series, and the ordinate represents the coordinate result of positioning. It can be seen from the figure that the positioning method of the Adaboost. M2 ensemble learning model based on the Kiney decision tree has higher positioning accuracy (see Table 3) and better stability under the resolution of 0.5 meter compared with the positioning method adopting other positioning models. Meanwhile, by comparing table 1, it can be found that the positioning accuracy of the positioning method of the same model is improved to different degrees after the data preprocessing is performed by the two-dimensional double-correlation distributed wavelet filtering of the invention.
TABLE 3
In another aspect of the embodiments of the present invention, there is also provided a passive indoor positioning device, fig. 6 is a schematic structural diagram of the passive indoor positioning device according to the embodiments of the present invention, and as shown in fig. 6, the passive indoor positioning device includes: a memory 61 for storing instructions; and a processor 62 for executing the aforementioned passive indoor positioning method according to instructions in the memory 61. The method can distinguish and filter RSS sample noise and random interference, retain normal jump data, has strong generalization capability in the positioning process, and can improve the positioning accuracy and stability on the whole.
The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are only exemplary embodiments of the present invention, and are not intended to limit the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (7)
1. A passive, indoor positioning method, comprising:
step A, taking a non-open indoor environment as a monitoring area, taking all acquired RSS values of radio frequency network links as training samples when a target is at each coordinate in the monitoring area, and taking coordinate numbers as sample labels;
b, performing two-dimensional double-correlation distributed wavelet filtering processing on the training sample to determine the filtered training sample; wherein, step B includes the substeps:
substep B1, forming training samples into a sample matrix, wherein each column represents a training sample, the number of columns is the number of samples, each row represents the RSS value of the same dimension in different samples, the sample matrix is subjected to wavelet decomposition of each dimension data according to the dimension, a wavelet function selects db1 wavelet, and a low-frequency wavelet coefficient ca and a h-th layer high-frequency wavelet coefficient cd are obtainedhWherein, one dimension refers to a radio frequency link, h is more than or equal to 1 and less than or equal to s, s is the number of layers of wavelet decomposition, and s is more than or equal to 2;
substep B2, retaining low frequency wavelet coefficient ca, and utilizing high frequency wavelet coefficients of h layer and h +1 layer to obtain two-dimensional double correlation longitudinal correlation coefficient corrhDetermining the high-frequency wavelet coefficient energy PcdhAnd longitudinal correlation coefficient energy PcorrhDetermining a normalized longitudinal correlation coefficient corrnh；
Substep B3 of comparing the high frequency wavelet coefficients cdhNormalization to corresponding layerSetting the high-frequency wavelet coefficient greater than the normalized longitudinal correlation coefficient to zero, retaining the high-frequency wavelet coefficient less than or equal to the normalized longitudinal correlation coefficient, retaining all the high-frequency wavelet coefficients of the last layer, and finally determining the retained high-frequency wavelet coefficient cdi；
Sub-step B4, dividing the preserved high frequency wavelet coefficient into U segments, U is more than or equal to 2, each segment selects d data, d is positive integer, and the preserved high frequency wavelet coefficients of the j segment and the j +1 segment are respectively utilized according to layers to calculate the transverse non-time-shift correlation coefficient R1 in the two-dimensional double correlationijJ is more than or equal to 1 and less than or equal to U-1, the j +1 th segment of the reserved high-frequency wavelet coefficient is subjected to time shift, wherein the time shift is a, a is a positive integer less than or equal to d/2, and the transverse time shift correlation coefficient R2 is calculated according to layers respectivelyijDetermining the difference value Rm between the time-shift correlation coefficient and the non-time-shift correlation coefficientijAnd selecting each layer R1ijR1 with the smallest absolute valueijCorresponding data segment cdriAnd each layer RmijRm with the smallest absolute valueijCorresponding data segment cdrmiCombined to filter threshold parameter estimation data cdmi；
Substep B5, according to cdmiDetermining a filter threshold parameter sigma, then determining a filter threshold thr, performing distributed filtering on the reserved high-frequency wavelet coefficient, and obtaining a filtered high-frequency wavelet coefficient cdfi(ii) a And
substep B6 of using the low frequency wavelet coefficients ca and the filtered high frequency wavelet coefficients cdfiPerforming wavelet reconstruction to obtain a training sample after filtering;
step C, establishing an Adaboost.M2 ensemble learning model based on a Kini decision tree, training by using the filtered training samples and sample labels, and determining the trained ensemble learning model; and
and step D, collecting all radio frequency network link RSS values when the target moves in the monitoring area at will, using the RSS values as test samples, carrying out two-dimensional double-correlation distributed wavelet filtering on the test samples, determining the filtered test samples, inputting the filtered test samples into the trained integrated learning model, and determining a positioning result.
2. A method according to claim 1, wherein in sub-step B2 the normalized longitudinal correlation coefficient is determined according to the formula:
corrh＝cdh·cdh+1
wherein n is a high-frequency wavelet coefficient cdhThe length of the sequence is that g is more than or equal to 1 and less than or equal to n, and g is a high-frequency wavelet coefficient cdhSequence number of data in the sequence.
3. The method of claim 1 wherein in sub-step B4, R1 is determined according to the formulaij、R2ij、RmijAnd selecting cdri、cdrmi、cdmi：
Rmij＝|R1ij-R2ij|
cdrmi＝cdijIf, if
cdri＝cdijIf, if
cdmi＝{cdrmi，cdri}
Wherein cdij+1Is cdijNext adjacent non-time-shifted fragment, cd'ij+1Is cdij+1Right shift by a time shift amount to form a time shift small segment, and Cov is cdij+1And cdijCovariance between, Var is cdij+1And cdijThe variance between.
4. The method according to claim 1, wherein in sub-step B5, a filtering threshold parameter σ and a threshold thr are determined according to the following formula, and distributed wavelet filtering is performed to obtain a filtered high frequency wavelet coefficient cdfi：
Wherein q is more than or equal to 2 and less than or equal to s, L is the length of each layer reserved high-frequency wavelet coefficient cdi, L is U × d, mean is an intermediate value, y is the serial number of each data in each layer reserved high-frequency wavelet coefficient, and y is more than or equal to 1 and less than or equal to L.
5. The method according to claim 1, wherein step C comprises the sub-steps of:
and a substep C1 of determining the Gini coefficient of each attribute in the training sample, and selecting the attribute A corresponding to the maximum value of the Gini coefficient as the optimal attribute to perform the growth of the decision tree, wherein the formula of the Gini coefficient Gini (A) is as follows:
wherein V is the branch number of the decision tree, N is the number of coordinates, S'bcSample subset S 'sorted with A as splitting attribute'cThe number of samples belonging to class b in the group, E is the number of total samples at the splitting point, and b is more than or equal to 1 and less than or equal to N; and
and a substep C2 of determining the trained ensemble learning model based on the decision tree in the substep C1:
iteratively updating sample weights:
determining the error rate of the decision tree:
determining weights of the decision tree:
wherein, W0(f) Is the initial weight of the f-th sample, N is the number of coordinates, M is the number of training samples contained in each coordinate, Wk(f, z) is the sum of the probabilities that in the k-th iteration sample f is classified into all error classes z, k is the number of iterations, hk(xf，zf) Classifying the f sample into the correct class z for the k decision treefPossibility of (a), hk(xfZ) the kth decision tree classifies the f-th sample as the correct class zfExcept for the possibility of all other incorrect categories z,kis the error rate of the kth decision tree, akIs the weight of the kth decision tree, Sum。
6. The method of claim 1, wherein sub-step B1 further comprises the steps of: and taking all the radio frequency network link RSS values acquired when no target exists in the monitoring area as reference samples, carrying out difference processing on the training samples and the reference samples, and determining difference signals to be used as the training samples.
7. A passive indoor positioning device, comprising:
a memory to store instructions; and
a processor for performing the method of any of claims 1 to 6 in accordance with the instructions.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710951127.5A CN107884744B (en) | 2017-10-12 | 2017-10-12 | Passive indoor positioning method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710951127.5A CN107884744B (en) | 2017-10-12 | 2017-10-12 | Passive indoor positioning method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107884744A CN107884744A (en) | 2018-04-06 |
| CN107884744B true CN107884744B (en) | 2021-01-01 |
Family
ID=61781631
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710951127.5A Active CN107884744B (en) | 2017-10-12 | 2017-10-12 | Passive indoor positioning method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107884744B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111325261B (en) * | 2020-02-14 | 2023-12-05 | 中国科学院半导体研究所 | Passive indoor positioning algorithm and device for dynamic environment |
| CN117295157B (en) * | 2023-11-24 | 2024-02-13 | 北京国旺盛源智能终端科技有限公司 | Positioning method and system for wearable back splint terminal |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7038617B1 (en) * | 2005-05-26 | 2006-05-02 | Lockheed Martin Corporation | Sensor doppler/wavelet processing |
| CN102186194A (en) * | 2011-05-09 | 2011-09-14 | 松日数码发展(深圳)有限公司 | Method for establishing passive target measurement model based on wireless sensor network |
| CN103023589A (en) * | 2012-12-06 | 2013-04-03 | 中山大学 | Indoor passive motion detection method and device |
| CN104883732A (en) * | 2015-04-14 | 2015-09-02 | 哈尔滨工程大学 | Enhanced indoor passive human body location method |
| CN105116375A (en) * | 2015-07-16 | 2015-12-02 | 北京理工大学 | Robust passive passiveness target positioning method based on multiple frequency points |
-
2017
- 2017-10-12 CN CN201710951127.5A patent/CN107884744B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7038617B1 (en) * | 2005-05-26 | 2006-05-02 | Lockheed Martin Corporation | Sensor doppler/wavelet processing |
| CN102186194A (en) * | 2011-05-09 | 2011-09-14 | 松日数码发展(深圳)有限公司 | Method for establishing passive target measurement model based on wireless sensor network |
| CN103023589A (en) * | 2012-12-06 | 2013-04-03 | 中山大学 | Indoor passive motion detection method and device |
| CN104883732A (en) * | 2015-04-14 | 2015-09-02 | 哈尔滨工程大学 | Enhanced indoor passive human body location method |
| CN105116375A (en) * | 2015-07-16 | 2015-12-02 | 北京理工大学 | Robust passive passiveness target positioning method based on multiple frequency points |
Non-Patent Citations (2)
| Title |
|---|
| 基于RSS的室内无线定位***的设计与实现;李健伟;《中国优秀硕士学位论文全文数据库信息科技辑》;20170215;第5-30页 * |
| 室内定位特征选择算法研究;张乐玫;《software》;20150131;第38-46页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107884744A (en) | 2018-04-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107563999A (en) | A kind of chip defect recognition methods based on convolutional neural networks | |
| CN111126471A (en) | Microseism event detection method and system | |
| CN105243388B (en) | Waveform classification based on dynamic time warping and partitioning algorithm | |
| Lu et al. | Classification of Camellia (Theaceae) species using leaf architecture variations and pattern recognition techniques | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN107274451A (en) | Isolator detecting method and device based on shared convolutional neural networks | |
| CN104732244B (en) | The Classifying Method in Remote Sensing Image integrated based on wavelet transformation, how tactful PSO and SVM | |
| CN106604229A (en) | Indoor positioning method based on manifold learning and improved support vector machine | |
| CN106056097B (en) | Millimeter wave detection method of small target | |
| CN103544499B (en) | The textural characteristics dimension reduction method that a kind of surface blemish based on machine vision is detected | |
| CN104374738A (en) | Qualitative analysis method for improving identification result on basis of near-infrared mode | |
| CN110020712B (en) | Optimized particle swarm BP network prediction method and system based on clustering | |
| CN106600595A (en) | Human body characteristic dimension automatic measuring method based on artificial intelligence algorithm | |
| CN109919241A (en) | Hyperspectral unknown class target detection method based on probability model and deep learning | |
| CN106204543A (en) | Fabric defect detection method based on single category support vector machines | |
| CN104850867A (en) | Object identification method based on intuitive fuzzy c-means clustering | |
| CN108877947A (en) | Depth sample learning method based on iteration mean cluster | |
| CN107884744B (en) | Passive indoor positioning method and device | |
| CN109740631A (en) | Object-based OBIA-SVM-CNN Remote Image Classification | |
| CN111983569B (en) | Radar interference suppression method based on neural network | |
| CN106529458A (en) | Deep neural network space spectrum classification method for high-spectral image | |
| CN114818809A (en) | Fault diagnosis method for SCA-SVM motor rolling bearing based on cross wavelets | |
| CN103971362B (en) | SAR image change-detection based on rectangular histogram and elite genetic algorithm for clustering | |
| CN114330486A (en) | Power system bad data identification method based on improved Wasserstein GAN | |
| CN110427019A (en) | A kind of NPDA Fault Classification and control device based on Multivariate Discriminant Analysis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |