Unmanned vehicle indoor and outdoor seamless positioning method based on area identification

Abstract

The invention discloses an unmanned vehicle indoor and outdoor seamless positioning method based on area identification, and belongs to the field of automatic driving. The method comprises the following steps: firstly, designing a joint prediction model of a support vector machine (PSO-SVM) and a hidden Markov model (HMM) based on particle swarm optimization, and using acquired GNSS environment characteristic data as model input to obtain a region identification result; then, three region models (an indoor model, an indoor and outdoor transition region model and an indoor model) are defined to describe the environment position of the vehicle, and corresponding measurement information is selected according to a region recognition result; and finally, the switching probability of the three environment sub-models in the IMM algorithm is updated by using the area switching probability, so that the accuracy and the positioning precision of indoor and outdoor seamless positioning environment area identification of the unmanned vehicle are improved. Through experimental verification, compared with a single sensor and an IMMEKF algorithm, the method provided by the invention can better adapt to environmental changes, so that the accuracy and robustness of unmanned vehicle positioning are improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving positioning, and in particular to a method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition. Background Art

[0002] In recent years, driverless technology has developed rapidly, among which high-precision vehicle positioning is the key to driverless planning. Indoor and outdoor seamless positioning technology uses the unmanned vehicle terminal as a carrier to effectively integrate the various sensor data sources currently received, integrate and adapt various positioning technology methods in different scenarios, and use the seamless positioning framework in different indoor and outdoor regional environments to achieve indoor and outdoor seamless positioning of unmanned vehicles, aiming to meet the requirements of high-precision positioning in various situations. The Inertial Navigation System (INS) can provide relative attitude, speed and position estimation. The GNSS / INS combined navigation system can effectively make up for the short-term failure of the Global Navigation Satellite System (GNSS).

[0003] However, in complex urban scenes, GNSS is easily interfered by buildings, resulting in signal attenuation, significantly reducing positioning performance, and even failing to provide positioning services; in indoor environments, the positioning capability of the GNSS / INS combined navigation system is limited; one of the challenges faced by the traditional indoor and outdoor seamless positioning framework is that the vehicle cannot accurately identify the regional boundaries during movement. In the process of data fusion, the probability selection of the sub-model is not optimal, resulting in low positioning accuracy of unmanned vehicles during indoor and outdoor movement. Therefore, combining UltraWide Band (UWB) with GNSS and Inertial Measurement Unit (IMU) to form a GNSS / IMU / UWB positioning system can effectively achieve seamless positioning of unmanned vehicles in complex urban environments. Although PSO-SVM recognition has certain pattern classification capabilities, it is easy to fall into local optimal solutions when facing complex and changeable regional features, resulting in limited recognition accuracy and difficulty in accurately adapting to subtle differences under different working conditions, thereby affecting the overall recognition accuracy. HMM recognition is overly dependent on the assumption of state transition probability, and has poor adaptability to sudden or abnormal situations. In actual complex scenarios, it is prone to misjudgment and cannot stably and efficiently complete high-precision area recognition tasks. In addition, the traditional IMM algorithm, due to the lack of reasonable measurement selection, will introduce a large amount of redundant or noisy data, which increases the burden on the algorithm during data processing, interferes with the extraction of real and effective information, and ultimately greatly reduces the positioning accuracy. At the same time, it lacks optimization of the sub-model switching probability, cannot dynamically adjust the model weight according to the real-time scenario, and is difficult to flexibly respond to changing environments, resulting in delayed response when switching between different working conditions, and it is difficult to improve positioning accuracy. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition, comprising the following steps:

[0006] S1. Build a PSO-SVM / HMM joint prediction model, collect GNSS environment feature values ​​through unmanned vehicles, and input the feature values ​​into the joint prediction model for training;

[0007] GNSS environmental characteristic values ​​include: total number of visible satellites, mean signal-to-noise ratio, and environmental labels corresponding to known environmental areas;

[0008] The joint prediction models built include: PSO-SVM prediction model and HMM prediction model;

[0009] The steps to build a joint prediction model are as follows:

[0010] S1.1, through the PSO-SVM prediction model, the optimal objective function is obtained, as follows:

[0011] For the kth SVM binary classifier, the samples belonging to the kth SVM binary classifier are defined as positive classes, and the remaining samples are defined as negative classes; the expression is as follows:

[0012]

[0013] The training goal of the model is to find a hyperplane that maximizes the distance between the hyperplane and the two types of sample points;

[0014] The hyperplane expression is as follows:

[0015] (ω (k) ) T x i +b (k) =0;

[0016] In the formula, ω (k) represents the normal vector of the hyperplane in the kth SVM classifier; b (k) represents the intercept; x i Indicates the sample to be tested;

[0017] The calculation expression of the interval is as follows:

[0018] 2 / ||ω (k) ||;

[0019] For any sample point (x i ,y i (k) ) to the hyperplane is expressed as:

[0020]

[0021] Among them, any sample point (x i ,y i (k) )middle, is the sample x to be tested i The corresponding regional status label category, Regional state ∈ S, S represents the regional state set and; the sample to be tested x i It is a second-order vector, which contains two pieces of information: the total number of visible satellites of GNSS and the mean signal-to-noise ratio;

[0022] The objective function to be optimized is expressed as:

[0023]

[0024] Where n is the total number of samples; C represents the penalty parameter; ξ(k) represents the slack variable, ξ i (k) ≥0, used to handle the situation where samples are linearly inseparable;

[0025] Introducing the Lagrange multiplier α into the objective function i (k) and α ii (k) , construct the Lagrangian function, the expression is as follows:

[0026]

[0027] In the formula, α i (k) represents the first Lagrange multiplier; α ii (k) represents the second Lagrange multiplier; y i (k) and ii (k) Respectively represent the sample x to be tested i and the sample x to be tested ii The regional state label category, where i, ii∈n, i≠ii; is the sample x to be tested i Mapping in high-dimensional space; is the sample x to be tested ii Mapping in high-dimensional space; is the kernel function;

[0028] In the PSO algorithm, the kernel parameter γ is introduced and used together with the penalty parameter C as the position vector of the particle to be input into the particle swarm algorithm for optimization. The above dual problem is solved to obtain the optimal α (k)* , and then we get ω (k) and b (k) ;

[0029] S1.2, construct HMM prediction model;

[0030] A hidden Markov model based on indoor and outdoor area recognition is established, and the expression is as follows:

[0031] λ=(S,O,A,B,π)

[0032] Where S represents the regional state set, including outdoor area 1, indoor-outdoor transition area 2 and indoor area 3, S∈(1,2,3); O={o1,o2} represents the observation vector composed of the total number of visible satellites and the mean signal-to-noise ratio; A represents the probability matrix of transferring to the other two areas starting from the current location of the unmanned vehicle; B represents the probability of obtaining observation values ​​from three hidden states (outdoor area 1, indoor-outdoor transition area 2 and indoor area 3) after obtaining a set of observation values ​​O; π represents the probability of the unmanned vehicle being in different areas at the initial moment;

[0033] S1.3, determine the model weight information and build a PSO-SVM / HMM joint prediction model;

[0034] Assume that the validation set has n samples and the true category label is Y = {y1,y2,...,y i ,...,y n}, where y i The regional state ∈S; then the classification result of PSO-SVM on the verification set is: in, The regional state ∈S of HMM is: in, The regional state ∈S;

[0035] Definition N k is the number of real samples of category k in Y, and the region state of k∈S, N s,k N is the number of samples correctly classified by PSO-SVM for the validation samples, h,k is the number of samples correctly classified by HMM for the validation samples, P S,K represents the classification accuracy of PSO-SVM, P H,K Represents the classification accuracy of HMM, and the total accuracy is recorded as P K =P S,K +P H,K ;

[0036] Defined in joint classification prediction and They represent the classification prediction result weight coefficient of PSO-SVM and the classification prediction weight coefficient of HMM in k respectively;

[0037] Classification predictions include: cases where the results are inconsistent in indoor areas and cases where the results are consistent in indoor areas;

[0038] S2, connect GNSS information as an outdoor positioning source to the unmanned vehicle, provide the vehicle posture through the IMU, and arrange 4 UWB base stations in the indoor rectangular area as indoor positioning sources, build three positioning models (outdoor model, indoor-outdoor transition model, indoor model) and use the interactive multi-model algorithm (IMM) to initially solve the vehicle positioning;

[0039] GNSS information includes: the longitude of the unmanned vehicle and the latitude of the unmanned vehicle;

[0040] S3. Input the GNSS environmental feature values ​​collected during the movement of the unmanned vehicle into the joint prediction model, perform regional identification and division on the unmanned vehicle's movement trajectory, and output the regional transfer probability. The steps are as follows:

[0041] S3.1. Divide the results of S2 into a training set and a test set in an 8:2 ratio; train the model using the training set, and use the trained model to test the test set;

[0042] S3.2, output PSO-SVM classification prediction results;

[0043] For the sample x to be tested i , respectively calculate the hyperplane f when the regional state ∈S of k k (x) = (ω (k) ) T x i +b (k) , the sample to be tested is classified as f k (x) is the largest category, and the final classification result is defined as:

[0044] V * = {V1 * ,V2 * ,...V i * ,...,V n *},V i * The regional state ∈S;

[0045] Sample to be tested x i After the hyperplane is input, the result will be calculated. If the result is greater than 0, it is classified as positive, and if it is less than 0, it is classified as negative. In this way, the binary classification of a single model can be achieved, so the hyperplane is a decision function;

[0046] S3.3, output HMM classification prediction results;

[0047] Solve the hidden sequence q with the highest probability in the test set by using the Viterbi algorithm t * =ψ t+1 (qt+1 * ) where q (t+1) represents the hidden sequence at the next moment; ψ t+1 Indicates that the backtracking pointer at time t+1 is used; the resulting hidden sequence is recorded as:

[0048] H * ={S1 * ,S2 * ,...,S i * ,...,S n *}, where S i * The regional state ∈S;

[0049] S3.4, output the PSO-SVM joint classification prediction result;

[0050] Get the prediction result V of PSO-SVM * The prediction H with HMM * After that, the classifier weight information is combined and The final joint classification result is recorded as:

[0051] S3.5, output region transfer probability;

[0052] The regional transfer matrix is ​​C T =[c ij ] 3×3 , where c ij Represents the area q after calculation i * Transfer to other areas after calculation q j * The number of transfers: If T c,t * =q i * And T c,t+1 * =q j * , then c ij =c ij +1, where T c,t * represents the classification label corresponding to time t; let Indicates that from q i * The total number of transfers departing from the region, the regional switching probability matrix is ​​expressed as: Among them, m ij The regional state ∈S;

[0053] S4, optimizing the sub-model switching probability in the interactive multi-model algorithm according to the regional transfer probability;

[0054] In the IMM algorithm, the model transition probability matrix is ​​defined as: Among them, p ij The regional state ∈S, then the updated model transition probability after the update is: P m =ω p P+ω m M=[p m,ij ] 3×3 ; Among them, ω p With ω m They represent the weight coefficient of the regional switching probability and the weight coefficient of the model transition probability respectively;

[0055] S5. Reselect measurement information according to the area identification and division results;

[0056] Specifically, the measurement selection is re-performed according to the regional recognition result. If the recognition result is a digital label 1, GNSS information is selected as the measurement; if the recognition result is 2, GNSS / UWB information is selected as the measurement. When selecting the measurement in the indoor and outdoor transition area, the measurement weight information should be added to better select the measurement and adjust it according to experience; if the recognition result is 3, UWB information is selected as the measurement;

[0057] S6, inputting the updated sub-model switching probability and measurement information into the extended Kalman filter of the interactive multi-model algorithm to perform fusion relocation of the vehicle;

[0058] Specifically, the measurement information updated by regional identification is used as the measurement information of the interactive multi-model, and the IMU measurement value is used as the state input. The outdoor model (ODR), the semi-outdoor model (I-OTR), and the indoor model (IDO) are introduced as regional positioning sub-models to describe the motion state of the vehicle. The sub-model switching probability in the IMM algorithm is optimized according to the regional transfer probability, and the extended Kalman filter is used as the filter of the interactive multi-model, which is beneficial to the filtering of the nonlinear motion model. The present invention enables the vehicle to provide high-precision positioning both indoors and outdoors, and compared with the traditional IMM positioning and the filtering of a single model, the positioning accuracy of the present invention is improved.

[0059] Beneficial effects of the present invention:

[0060] Compared with traditional analysis, the present invention improves the positioning accuracy of unmanned vehicles in complex environments by using a PSO-SVM / HMM joint prediction model and considering three situations: GNSS / IMU fusion (outdoor area), GNSS / UWB / IMU fusion (indoor-outdoor transition area) and UWB / IMU fusion (indoor area).

[0061] The present invention improves the recognition accuracy of the unmanned vehicle movement area by adjusting the weight coefficient through the model, which is better than a single classification prediction model.

[0062] The present invention utilizes extended Kalman filtering as a filter of interactive multi-models, which is beneficial to filtering of nonlinear motion models. The present invention enables the vehicle to provide high-precision positioning during indoor and outdoor movement, and has good environmental adaptability and positioning accuracy stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a step diagram of the present invention;

[0064] Figure 2 It is a technical framework diagram of the present invention;

[0065] Figure 3 A technical roadmap for solving classification accuracy for the present invention;

[0066] Figure 4 This is a schematic diagram of the comparison of the area recognition accuracy of the present invention;

[0067] Figure 5 A schematic diagram showing a comparison of the positioning trajectory of the method of the present invention and that of a single sensor;

[0068] Figure 6 Schematic diagram showing the comparison of the single-point positioning error of the sampling point in the X direction between the method of the present invention and a single sensor;

[0069] Figure 7 It is a schematic diagram comparing the single-point positioning error of the sampling point in the Y direction between the method of the present invention and a single sensor;

[0070] Figure 8 It is a schematic diagram for comparing the positioning trajectories of the method of the present invention, the single positioning model and the IMM_EKF algorithm;

[0071] Fig. 9 It is a schematic diagram comparing the single-point positioning error of the sampling point in the X direction of the method of the present invention, the single positioning model and the IMM_EKF algorithm;

[0072] Fig.10 It is a schematic diagram comparing the single-point positioning error of the sampling point in the Y direction between the method of the present invention, the single positioning model and the IMM_EKF algorithm. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] like Figure 1 As shown, a method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition includes the following steps:

[0075] S1. Build a PSO-SVM / HMM joint prediction model, collect GNSS environment feature values ​​through unmanned vehicles, and input the feature values ​​into the joint prediction model for training;

[0076] GNSS environmental characteristic values ​​include: total number of visible satellites, mean signal-to-noise ratio, and environmental labels corresponding to known environmental areas;

[0077] The joint prediction models built include: PSO-SVM prediction model and HMM prediction model;

[0078] The steps to build a joint prediction model are as follows:

[0079] S1.1, through the PSO-SVM prediction model, the optimal objective function is obtained, as follows:

[0080] For the kth SVM binary classifier, the samples belonging to the kth SVM binary classifier are defined as positive classes, and the remaining samples are defined as negative classes; the expression is as follows:

[0081]

[0082] The training goal of the model is to find a hyperplane that maximizes the distance between the hyperplane and the two types of sample points;

[0083] The hyperplane expression is as follows:

[0084] (ω (k) ) T x i +b (k) =0;

[0085] In the formula, ω (k) represents the normal vector of the hyperplane in the kth SVM classifier; b (k) represents the intercept; x i Indicates the sample to be tested;

[0086] The calculation expression of the interval is as follows:

[0087] 2 / ||ω (k) ||;

[0088] For any sample point (x i ,y i (k) ) to the hyperplane is expressed as:

[0089]

[0090] Among them, any sample point (x i ,y i (k) )middle, is the sample x to be tested i The corresponding regional status label category, Regional state ∈ S, S represents the regional state set and; the sample to be tested x i It is a second-order vector, which contains two pieces of information: the total number of visible satellites of GNSS and the mean signal-to-noise ratio;

[0091] The objective function to be optimized is expressed as:

[0092]

[0093] Wherein, in this embodiment, n=1406; C>0, which represents a penalty parameter for controlling the degree of penalty for misclassification, C∈(1-10), in this embodiment, C=1.8; ξ (k) represents the slack variable, ξ i (k) ≥0, used to handle the situation where samples are linearly inseparable;

[0094] Introducing the Lagrange multiplier α into the objective function i (k) and α ii (k) , construct the Lagrangian function, the expression is as follows:

[0095]

[0096] In the formula, α i (k) represents the first Lagrange multiplier; α ii (k) represents the second Lagrange multiplier; y i (k) and ii (k) Respectively represent the sample x to be tested i and the sample x to be tested ii The regional state label category, where i, ii∈n, i≠ii; is the sample x to be tested i Mapping in high-dimensional space; is the sample x to be testedii Mapping in high-dimensional space; is a kernel function, which is a Gaussian kernel function in this embodiment. and The inner product of

[0097] In the PSO algorithm, the kernel parameter γ is introduced and used together with the penalty parameter C as the position vector of the particle to be input into the particle swarm algorithm for optimization. The above dual problem is solved to obtain the optimal α (k)* , and then we get ω (k) and b (k) ; The kernel parameter γ is directly solved by MATLAB built-in tools;

[0098] S1.2, construct HMM prediction model;

[0099] A hidden Markov model based on indoor and outdoor area recognition is established, and the expression is as follows:

[0100] λ=(S,O,A,B,π)

[0101] Where S represents the regional state set, including three cases: outdoor area 1, indoor-outdoor transition area 2, and indoor area 3. The values ​​are shown in Table 1. S∈(1, 2, 3); O={o1,o2} represents the observation vector composed of the total number of visible satellites and the mean signal-to-noise ratio; A represents the probability matrix of the transfer from the current location of the unmanned vehicle to the other two areas, that is, A=[a ij ], indicating that from the current area s i To the other two j The probability of ij =P(s j,t+1 |s i,t ), t represents the current time t, t+1 represents the next time, for any s i ,s j The regional state ∈S satisfies The probability matrix is ​​a 3×3 square matrix, and its specific values ​​are: B represents the probability of obtaining observations from three hidden states (outdoor area 1, indoor-outdoor transition area 2, and indoor area 3) after obtaining a set of observations O, that is, B = [b i (o t )] indicates that the hidden state is s i When the sequence O is observed t The probability of i (o t )=p(o t |q t =s i ), B is a 3×2 square matrix, and its specific value is: π represents the probability of the unmanned vehicle being in different areas at the initial moment, that is, π={π i} indicates that the initial moment is s i The probability of π is a 3×1 matrix:

[0102] Table 1: Values ​​of S

[0103]

[0104] S1.3, determine the model weight information and build a PSO-SVM / HMM joint prediction model;

[0105] like Figure 3 As shown, there are n samples, and the true category label is Y = {y1, y2, ..., y i ,...,y n}, where y i The regional state ∈S; then the classification result of PSO-SVM on the verification set is: in, The regional state ∈S of HMM is: in, The regional state ∈S;

[0106] Definition N k is the number of real samples of category k in Y, and the region state of k∈S, N s,k N is the number of samples correctly classified by PSO-SVM for the validation samples, h,k is the number of samples correctly classified by HMM for the validation samples, P S,K represents the classification accuracy of PSO-SVM, P H,K Represents the classification accuracy of HMM, and the total accuracy is recorded as P K =P S,K +P H,K ;

[0107] In this embodiment, N1=975 (there are 975 outdoor area points), N2=365 (there are 365 indoor and outdoor transition area points), N3=66 (there are 66 indoor area points); N S,1 =875;N S,2 =365; N S,3 =60 (the reason why the total number of classifications is inconsistent with the actual number of samples is that N S，K is the number of correctly classified points, not the final classification result; for example, in this embodiment, there are 975 real outdoor sample points, but the correct outdoor results obtained by SVM classification are only 875, and the same is true for HMM classification results); N h,1 =955;Nh,2 =343; N h,3 =60; calculated P s,1 =0.90; P s,2 =1;P s,3 =0.90; P h,1 =0.98; P h,2 =0.94; P h,3 =0.90; P k,1 =1.88; P K,2 =1.94; P K,3 =1.8;

[0108] Defined in joint classification prediction and They represent the classification prediction result weight coefficient of PSO-SVM and the classification prediction weight coefficient of HMM in k respectively;

[0109] Classification predictions include: cases where the results are inconsistent in indoor areas and cases where the results are consistent in indoor areas;

[0110] In the case of inconsistency, there are the following specific situations: when the SVM classification result is outdoor and the HMM result is semi-outdoor, the classification accuracy is used to distinguish and determine the final result; in this embodiment, the real outdoor area has 975 data points, but the SVM correctly classifies only 875 points with a classification accuracy of 0.90, and the HMM correctly classifies 955 points with a classification accuracy of 0.98; when the classification is all outdoor areas, there are 80 points with inconsistent classifications. At this time, the weight coefficient is introduced, and the PSO-SVM weight coefficient in the outdoor area is calculated to be 0.47, and the HMM weight coefficient is 0.53; in the case of the indoor-outdoor transition area, there are 22 points with inconsistent classifications. The weight coefficient is introduced, and the PSO-SVM weight coefficient in the transition area is calculated to be 0.52, and the HMM weight coefficient is 0.48;

[0111] The results in the indoor area are consistent, and there is no need to introduce weight coefficients. The weight coefficients of the two classifiers are expressed as: The HMM weight coefficient is expressed as: The region recognition results are as follows: Figure 4 As shown;

[0112] S2, such as Figure 2As shown in the figure, GNSS information is used as an outdoor positioning source to connect to the unmanned vehicle, the vehicle posture is provided by IMU, and four UWB base stations are arranged in the indoor rectangular area as indoor positioning sources. Three positioning models (outdoor model, indoor-outdoor transition model, indoor model) are constructed and the vehicle positioning is initially solved by the interactive multi-model algorithm (IMM);

[0113] GNSS information includes: the longitude of the unmanned vehicle and the latitude of the unmanned vehicle;

[0114] S3. Input the GNSS environmental feature values ​​collected during the movement of the unmanned vehicle into the joint prediction model, perform regional identification and division on the unmanned vehicle's movement trajectory, and output the regional transfer probability. The steps are as follows:

[0115] S3.1. Divide the results of S2 into a training set and a test set in an 8:2 ratio; train the model using the training set, and use the trained model to test the test set;

[0116] S3.2, output PSO-SVM classification prediction results;

[0117] For the sample x to be tested i , respectively calculate the hyperplane f when the regional state ∈S of k k (x) = (ω (k) ) T x i +b (k) , the sample to be tested is classified as f k (x) is the largest category, and the final classification result is defined as:

[0118] V * = {V1 * ,V2 * ,...V i * ,...,V n *},V i * The regional state ∈S;

[0119] Sample to be tested x i After the hyperplane is input, the result will be calculated. If the result is greater than 0, it is classified as positive, and if it is less than 0, it is classified as negative. In this way, the binary classification of a single model can be achieved, so the hyperplane is a decision function;

[0120] S3.3, output HMM classification prediction results;

[0121] Solve the hidden sequence q with the highest probability in the test set by using the Viterbi algorithm t * =ψ t+1 (qt+1 * ) where q (t+1) represents the hidden sequence at the next moment; ψ t+1 Indicates that the backtracking pointer at time t+1 is used; the resulting hidden sequence is recorded as:

[0122] H * ={S1 * ,S2 * ,...,S i * ,...,S n *}, where S i * The regional state ∈S;

[0123] S3.4, output the PSO-SVM joint classification prediction result;

[0124] Get the prediction result V of PSO-SVM * The prediction H with HMM * After that, the classifier weight information is combined and The final joint classification result is recorded as:

[0125] S3.5, output region transfer probability;

[0126] The regional transfer matrix is ​​C T =[c ij ] 3×3 , where c ij Represents the area q after calculation i * Transfer to other areas after calculation q j * The number of transfers: If T c,t * =q i * And T c,t+1 * =q j * , then c ij =c ij +1, where T c,t * represents the classification label corresponding to time t; let Indicates that from q i * The total number of transfers departing from the region, the regional switching probability matrix is ​​expressed as: Among them, m ij The regional state ∈S;

[0127] S4, optimizing the sub-model switching probability in the IMM algorithm according to the regional transfer probability;

[0128] In the IMM algorithm, the model transition probability matrix is ​​defined as: P = [p ij ] 3×3 , where p ij The regional state ∈S, then the updated model transition probability after the update is: P m =ω p P+ω m M=[p m,ij ] 3×3 ; Among them, ω p With ω m They respectively represent the weight coefficient of the regional switching probability and the weight coefficient of the model transition probability; the weight coefficients obtained in this embodiment are: regional switching probability weight: 0.5 and IMM sub-model transition probability weight: 0.5;

[0129] S5. Reselect measurement information according to the area identification and division results;

[0130] Specifically, re-measurement selection is performed based on the regional recognition result. If the recognition result is digital label 1, GNSS information is selected as the measurement; if the recognition result is 2, GNSS / UWB information is selected as the measurement. When selecting measurements in the indoor and outdoor transition areas, measurement weight information should be added to better select the measurements and make adjustments based on experience. In this embodiment, there are 365 points assigned label 2. Based on experience, the weight ratio of the first 30% of the points is 0.75*GNSS information + 0.25*UWB information; the middle 40% of the points are based on: 0.5*GNSS information + 0.5*UWB information; the last 30% of the points are based on: 0.35*GNSS information + 0.65*UWB information; if the recognition result is 3, UWB information is selected as the measurement;

[0131] S6, inputting the updated sub-model switching probability and measurement information into the extended Kalman filter of the IMM algorithm to perform fusion relocation of the vehicle;

[0132] Specifically, the measurement information updated after regional identification is used as the measurement information of the interactive multi-model, and the IMU measurement value is used as the state input. The outdoor model (ODR), semi-outdoor model (I-OTR), and indoor model (IDO) are introduced as regional positioning sub-models to describe the motion state of the vehicle. The sub-model switching probability in the IMM algorithm is optimized according to the regional transfer probability, and the extended Kalman filter is used as the filter of the interactive multi-model, which is beneficial to the filtering of the nonlinear motion model. The present invention enables the vehicle to provide high-precision positioning during indoor and outdoor movement, and compared with traditional IMM positioning and single model filtering, the positioning accuracy of the present invention is improved. Compared with the positioning trajectory of a single sensor, such as Figure 5 As shown, the error comparison diagram of the present invention and a single sensor in the X direction is as follows Figure 6 As shown in the figure, the error comparison in the Y direction is as follows Figure 7 As shown; the present invention compares the traditional IMM positioning with the filtering positioning trajectory of a single model as shown Figure 8 As shown in FIG. 1 , the error comparison diagram of the present invention, the IMM algorithm and the single model filtering in the X direction is shown in FIG. Fig. 9 As shown in the figure, the error comparison in the Y direction is as follows Fig.10 In addition, the invention of this paper also performs environmental recognition, which can better adapt to complex environments.

Claims

1. A method for seamless indoor and outdoor positioning of unmanned vehicles based on area recognition, characterized in that: The following steps are involved: S1. Build a PSO-SVM / HMM joint prediction model, collect GNSS environmental feature values ​​through unmanned vehicles, and input the GNSS environmental feature values ​​into the joint prediction model for training; The GNSS environmental characteristic values ​​include: the total number of visible satellites, the mean signal-to-noise ratio, and the environmental labels corresponding to the known environmental areas; S2, connect the GNSS information as an outdoor positioning source to the unmanned vehicle, provide the vehicle posture through the IMU, and arrange 4 UWB base stations in the indoor rectangular area as indoor positioning sources, build three positioning models and solve the vehicle positioning for the first time through the interactive multi-model algorithm; The GNSS information includes: the longitude of the unmanned vehicle and the latitude of the unmanned vehicle; The three positioning models include: outdoor model, indoor-outdoor transition model and indoor model; S3, inputting the GNSS environmental feature values ​​collected during the movement of the unmanned vehicle into the joint prediction model, performing regional identification and division on the movement trajectory of the unmanned vehicle, and outputting the regional transfer probability; S4, optimizing the sub-model switching probability in the interactive multi-model algorithm according to the regional transfer probability; S5. Reselect measurement information according to the area identification and division results; S6. Input the updated sub-model switching probability and measurement information into the extended Kalman filter of the interactive multi-model algorithm, perform fusion relocation on the vehicle, and obtain the final result.

2. The method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition according to claim 1, characterized in that: The steps of constructing the PSO-SVM / HMM joint prediction model are as follows: S1.1, through the PSO-SVM prediction model, the optimal objective function is obtained, as follows: The objective function to be optimized is expressed as: Where n is the total number of samples; C represents the penalty parameter; ξ (k) represents the slack variable, Used to handle the situation where samples are linearly inseparable; ω (k) represents the normal vector of the hyperplane in the kth SVM classifier; b (k) represents the intercept; x i Represents the sample to be tested; any sample point (x i ,y i (k) )middle, is the sample x to be tested i The corresponding regional status label category, Regional state ∈ S, S represents the regional state set and; the sample to be tested x i It is a second-order vector, which contains two pieces of information: the total number of visible satellites of GNSS and the mean signal-to-noise ratio; Introducing the Lagrange multiplier α into the objective function i (k) and α ii (k) , construct the Lagrangian function, the expression is as follows: In the formula, α i (k) represents the first Lagrange multiplier; α ii (k) represents the second Lagrange multiplier; y i (k) and ii (k) Respectively represent the sample x to be tested i and the sample x to be tested ii The regional state label category, where i, ii∈n, i≠ii; is the sample x to be tested i Mapping in high-dimensional space; is the sample x to be tested ii Mapping in high-dimensional space; is the kernel function; In the PSO algorithm, the kernel parameter γ is introduced and used together with the penalty parameter C as the position vector of the particle to be input into the particle swarm algorithm for optimization. The above dual problem is solved to obtain the optimal α (k)* , and then we get ω (k) and b (k) ; S1.2, construct HMM prediction model; A hidden Markov model based on indoor and outdoor area recognition is established, and the expression is as follows: λ=(S,O,A,B,π) Where S represents the regional state set, including outdoor area 1, indoor-outdoor transition area 2 and indoor area 3, S∈(1,2,3); O={o1,o2} represents the observation vector composed of the total number of visible satellites and the mean signal-to-noise ratio; A represents the probability matrix of transferring to the other two areas starting from the current location of the unmanned vehicle; B represents the probability of obtaining observation values ​​from three hidden states after obtaining a set of observation values ​​O, including: outdoor area 1, indoor-outdoor transition area 2 and indoor area 3; π represents the probability of the unmanned vehicle being in different areas at the initial moment; S1.3, determine the model weight information and build a PSO-SVM / HMM joint prediction model; There are n samples, and the true category label is Y = {y1,y2,...,y i ,...,y n }, where y i The regional state ∈S; then the classification result of PSO-SVM on the verification set is: in, The regional state ∈S of HMM is: in, The regional state ∈S; Definition N k is the number of real samples of category k in Y, and the region state of k∈S, N s,k N is the number of samples correctly classified by PSO-SVM for the validation samples, h,k is the number of samples correctly classified by HMM for the validation samples, P S,K represents the classification accuracy of PSO-SVM, P H,K Represents the classification accuracy of HMM, and the total accuracy is recorded as P K =P S,K +P H,K ; Defined in joint classification prediction and They represent the classification prediction result weight coefficient of PSO-SVM and the classification prediction weight coefficient of HMM in k respectively; The classification prediction includes: the case where the results in the indoor area are inconsistent and the case where the results in the indoor area are consistent.

3. The method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition according to claim 1, characterized in that: The GNSS environmental feature values ​​collected during the movement of the unmanned vehicle are input into the joint prediction model, the unmanned vehicle movement trajectory is divided into regions, and the region transfer probability is output. The steps are as follows: S3.

1. Divide the results of S2 into a training set and a test set in an 8:2 ratio; train the model using the training set, and use the trained model to test the test set; S3.2, output PSO-SVM classification prediction results; For the sample x to be tested i , respectively calculate the hyperplane f when the regional state ∈S of k k (x) = (ω (k) ) T x i +b (k) , the sample to be tested is classified as f k (x) is the largest category, and the final classification result is defined as: V * = {V1 * ,V2 * ,...V i * ,...,V n * },V i * The regional state ∈S; Sample to be tested x i After the hyperplane is input, the result will be calculated. If the result is greater than 0, it is classified as positive, and if it is less than 0, it is classified as negative. In this way, the binary classification of a single model can be achieved, so the hyperplane is a decision function; S3.3, output HMM classification prediction results; Solve the hidden sequence q with the highest probability in the test set by using the Viterbi algorithm t * =ψ t+1 (q t+1 * ) where q (t+1) represents the hidden sequence at the next moment; ψ t+1 Indicates that the backtracking pointer at time t+1 is used; the resulting hidden sequence is recorded as: H * ={S1 * ,S2 * ,...,S i * ,...,S n * }, where S i * The regional state ∈S; S3.4, output the PSO-SVM joint classification prediction result; Get the prediction result V of PSO-SVM * The prediction H with HMM * After that, the classifier weight information is combined and The final joint classification result is recorded as: S3.5, output region transfer probability; The regional transfer matrix is ​​C T =[c ij ] 3×3 , where c ij Represents the area q after calculation i * Transfer to other areas after calculation q j * The number of transfers: If T c,t * =q i * And T c,t+1 * =q j * , then c ij =c ij +1, where T c,t * represents the classification label corresponding to time t; let Indicates that from q i * The total number of transfers departing from the region, the regional switching probability matrix is ​​expressed as: Among them, m ij The regional state ∈S.

4. The method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition according to claim 1, characterized in that: In the optimization of the sub-model switching probability in the interactive multi-model algorithm according to the regional transition probability, the model transition probability matrix in the interactive multi-model algorithm is defined as: Among them, p ij The regional state ∈S, then the updated model transition probability after the update is: P m =ω p P+ω m M=[p m,ij ] 3×3 ; Among them, ω p With ω m They represent the weight coefficient of the region switching probability and the weight coefficient of the model transition probability respectively.

5. The method for seamless indoor and outdoor positioning of an unmanned vehicle based on area recognition according to claim 1, characterized in that: According to the regional identification and division results, the measurement information is reselected and the measurement selection is re-performed according to the regional identification results. If the identification result is 1, GNSS information is selected as the measurement; if the identification result is 2, GNSS / UWB information is selected as the measurement. When selecting the measurement in the indoor and outdoor transition area, the measurement weight information is added to better select the measurement and adjust it according to experience; if the identification result is 3, UWB information is selected as the measurement; Among them, UWB means ultra-wideband obtained through a UWB base station.

Images