A vehicle localization method based on deep neural network under Toeplitz and sparse prior

Abstract

The present invention relates to a vehicle positioning method based on a deep neural network under Toeplitz and sparse prior, which comprises: using a large-scale uniform linear array in three cooperative base stations arranged in a right triangle to receive vehicle positioning signals, determining the angle of arrival (DOA) estimation signal form under each cooperative base station antenna array; calculating the covariance matrix of each cooperative base station antenna array receiving data, and obtaining DOA estimation by using a deep convolutional network after linear shrinkage estimation and sparse vector representation in turn; based on the obtained DOA estimation results of the three cooperative base stations, the vehicle is reliably positioned based on the prejudgment average criterion and the cross positioning principle. Compared with the existing technical solutions, the method of the present invention not only has low computational complexity, but also has significantly improved resolution and vehicle positioning accuracy for vehicles with very close spatial spacing.

Description

Technical Field

[0001] The present invention relates to the field of vehicle networking and intelligent transportation, and in particular to a vehicle positioning method based on a deep neural network under Toeplitz and sparse prior. Background Art

[0002] Internet of Vehicles technology has become a hot topic in the current development of the automotive industry. It can not only provide a wealth of services, such as real-time traffic information, remote wake-up, Internet of Vehicles, lane-level autonomous driving, etc., but also effectively improve traffic efficiency and driving safety. As one of the key technologies supporting Internet of Vehicles applications, high-performance vehicle positioning has attracted much attention in recent years.

[0003] Positioning technology based on the Global Navigation Satellite System (GNSS) has been widely adopted by the industry. However, due to system errors and obstructions caused by tunnels and clouds, GNSS cannot provide reliable and continuous positioning performance all day long. Especially in densely built urban areas, GNSS signals are easily affected by interference and blockage. As a result, its positioning accuracy is severely reduced, with an accuracy of only 10-50 meters. In this case, it is difficult to achieve safe driving relying solely on GNSS.

[0004] At present, some auxiliary positioning methods have been widely proposed to overcome the limitations of GNSS, such as radar, lidar, camera positioning, dead reckoning, and collaborative positioning (CP) using base stations (BSs) or roadside units (RSUs). Among them, CP has been widely studied due to its advantages in cost, latency, reliability, etc. According to the positioning principle, CP positioning is mainly divided into four categories: positioning based on received signal strength (RSS) or differential RSS (DRSS), positioning based on time of arrival (TOA), positioning based on time difference of arrival (TDOA), and positioning based on angle of arrival (DOA). Compared with the other three methods, the DOA-based method does not require the spatial fading characteristics of the signal or the perfect synchronization of the clock, so it has become one of the most competitive solutions for high-performance vehicle positioning.

[0005] Existing DOA estimation algorithms can be roughly divided into three categories:

[0006] Methods based on characteristic subspace include the multiple signal classification (MUSIC) algorithm, the subspace rotation invariance technique (ESPRIT) algorithm, the covariance-corrected MUSIC (R-MUSIC) algorithm, and the phase-compensated ESPRIT algorithm.

[0007] Methods based on sparse signal reconstruction, such as l1-SVD method, l0-norm approximation method, sparse Bayesian method, etc. They have good robustness to noise and can obtain DOA estimation under small sample conditions; however, the main problem of such solutions is their high computational complexity.

[0008] Deep learning (DL) based methods, such as deep neural network (DNN) or deep convolutional network (DCN) based methods, and dual one-dimensional convolutional neural network (D1D-CNN) based methods are good examples. These methods have the advantages of high computational efficiency, high resolution and robustness to non-ideal conditions.

[0009] However, it is worth emphasizing that efficient and super-resolution DOA estimation is still a challenge. Especially for large-scale arrays and small samples, most existing methods have problems with insufficient resolution or high computational burden in this case. On the other hand, in some practical IoV scenarios, the number of available samples may be quite limited due to real-time and accessibility requirements. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a vehicle positioning method based on a deep convolutional neural network under Toeplitz and sparse priors, which can avoid insufficient resolution, reduce computational complexity, and improve vehicle positioning accuracy.

[0011] The technical solution adopted by the present invention is a vehicle positioning method based on a deep neural network under Toeplitz and sparse prior, the method comprising the following steps:

[0012] Step 1: Arrange the three cooperative base stations in a right triangle, and set the angles between the antenna arrays of the three cooperative base stations and the x-axis in the rectangular coordinate system to β1, β2 and β3 respectively;

[0013] Step 2: Use the antenna arrays in the three cooperative base stations to receive the vehicle positioning signal and determine the estimated signal form of the angle of arrival (DOA) under the antenna array of each cooperative base station. The specific process is as follows:

[0014] Assume that the mutually unrelated narrowband positioning signals emitted by K vehicles are all incident on the antenna arrays of the three cooperative base stations. Assuming that the number of antennas in the antenna arrays of the three cooperative base stations is M and the array element spacing is d, the received data of the array of the i-th cooperative base station at the t-th sampling sample is expressed as:

[0015] y i (t)=A(θ)s(t)+n(t)

[0016] Among them, y i (t) = [y i,1(t),...,y i,M (t)] T , i = 1, 2, 3, A(θ) represents the array steering matrix, A(θ) = [a(θ1), ..., a(θ K )], the kth column of A(θ) is expressed as: s(t) represents the mutually uncorrelated narrowband positioning signal vector, s(t) = [s1(t), ..., s K (t)] T , n(t) represents the Gaussian white noise vector, n(t)=[n1(t),...,n M (t)] T , λ represents the carrier wavelength, satisfying λ ≥ 2d, and the superscript T represents the transposition operation;

[0017] Step 3: Calculate the covariance matrix of the data received by each cooperative base station antenna array: Where N represents the total number of samples, and the superscript H represents the conjugate transpose operation; the covariance matrix After linear shrinkage estimation and sparse vector representation, the DOA estimation result is obtained using a deep convolutional network;

[0018] Step 4: Based on the DOA estimation results of the three cooperative base stations obtained in step 3, the vehicle positioning is completed based on the prediction average criterion and the cross positioning principle.

[0019] Preferably, in step 3, the covariance matrix After linear shrinkage estimation and sparse vector representation, the specific steps of using deep convolutional network to obtain DOA estimation results include:

[0020] Step 3.1: For the covariance matrix Perform Toeplitz correction to obtain the target matrix R T , expressed as

[0021]

[0022] Among them, J m represents the M×M-dimensional shift matrix, J m Only the mth diagonal element is 1, the other elements are 0, and J -m =(J m ) T , J 0 =I M represents the identity matrix, Represents taking the matrix traces;

[0023] Step 3.2: Calculate the linear shrinkage coefficient α as

[0024]

[0025] Step 3.3: Get the improved covariance matrix estimate as

[0026]

[0027] Step 3.4: Perform eigenvalue decomposition to obtain K large eigenvalues and the mean of MK small eigenvalues Then we can calculate the unbiased estimate of the noise variance

[0028]

[0029] in, c = M / N;

[0030] Step 3.5: In the overcomplete basis matrix Φ = [b(φ1), ..., b(φ L )], we vectorize R and subtract the noise term to get the noise-free sparse representation model:

[0031]

[0032] Among them, vec(·) means to vectorize the matrix in the brackets by column, b(φ l )=vec(a(φ l ) H (φ l )), η=[η1,η2,...,η L ] T is a K sparse column vector;

[0033] Step 3.6: Construct sparse spectrum The sparse spectrum is substituted as an input value into a deep convolutional network for network training and DOA estimation; the deep convolutional network consists of an input layer, five hidden layers and an output layer; each of the hidden layers includes a one-dimensional convolutional layer, a batch normalization layer and an activation function, and the activation function is f ReLU (x)=max(0,x) in the nonlinear activation function layer, the numbers of convolution kernels of the five hidden layers are 24, 20, 12, 5, and 1, respectively, and the lengths of the convolution kernels are 21, 15, 11, 5, and 3, respectively.

[0034] Preferably, in step 4, the specific process of completing the reliable positioning of the vehicle based on the prediction average criterion and the cross positioning principle includes the following steps:

[0035] Step 4.1: The three cooperative base stations are base station 1, base station 2 and base station 3. The initial position information of the vehicle is calculated based on the DOA estimation results of base station 1 and base station 2 or base station 2 and base station 3 using the cross-positioning principle. The specific process is as follows: let the coordinates of base station 1, base station 2 and base station 3 be (0, Y), (0, 0) and (X, 0) respectively, and the DOA estimation values ​​of the kth vehicle arriving at the three cooperative base stations are and Then the estimated position of the kth vehicle obtained using the DOA estimation results of base station 1 and base station 2 is

[0036]

[0037] The estimated position of the kth vehicle obtained using the DOA estimation results of base stations 2 and 3 is

[0038]

[0039] The estimated position of the kth vehicle obtained using the DOA estimation results of base stations 1 and 3 is

[0040]

[0041] in

[0042] Step 4.2: Use the position estimate of the k-th vehicle to determine whether the k-th vehicle is located on the straight line between the adjacent base stations 1 and 3; if so, average the position estimates of the k-th vehicle obtained by base station 1 and base station 2 and base station 2 and base station 3 to obtain the final positioning result of the k-th vehicle; if not, average the position estimates of the k-th vehicle obtained by each pair of base stations to obtain the final positioning result of the k-th vehicle.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: the method of the present invention firstly utilizes the Toeplitz prior of the covariance matrix to enhance the estimation of the sample covariance matrix through linear shrinkage estimation, and then jointly utilizes sparse prior, deep convolutional neural network, predictive average criterion and cross-positioning principle to complete reliable vehicle positioning. The mechanism of offline training and online estimation of deep convolutional network not only makes the complexity of the method of the present invention low, but also the combined application of Toeplitz and sparse prior makes the method of the present invention have significantly improved resolution and vehicle positioning accuracy for vehicles with close spatial spacing. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the structure of the arrangement of three cooperative base stations in the present invention;

[0045] Figure 2 It is a model architecture diagram of the deep convolutional network described in the present invention;

[0046] Figure 3 This is a flow chart of using deep convolutional network data for network training and DOA estimation in the present invention;

[0047] Figure 4 It is a flow chart of a vehicle positioning method based on a deep convolutional neural network under Toeplitz and sparse priors of the present invention. DETAILED DESCRIPTION

[0048] The invention will be further described below with reference to the accompanying drawings and in combination with specific implementations, so that those skilled in the art can implement the invention with reference to the description. The protection scope of the invention is not limited to the specific implementations.

[0049] In order to overcome the problems of existing DOA estimation methods under large-scale arrays and small samples, the present invention provides a super-resolution DOA estimation scheme based on deep convolutional network and covariance matrix linear shrinkage estimation technology under three cooperative base stations, and completes the reliable positioning of vehicles in the case of close spatial spacing based on the predicted average criterion and cross-positioning principle.

[0050] The embodiment of the present invention provides a vehicle positioning method based on a deep convolutional neural network under Toeplitz and sparse prior, and the implementation steps are as follows:

[0051] Step 1: The three cooperative base stations are arranged in a right triangle, and the angles between the antenna arrays in the three cooperative base stations and the x-axis in the rectangular coordinate system are set to β1, β2 and β3 respectively, as shown in Figure 1 As shown;

[0052] Step 2: Use the large-scale uniform linear arrays in the three cooperative base stations to receive the vehicle positioning signal and determine the angle of arrival (DOA) estimation signal form under the antenna array of each cooperative base station. The specific process is as follows:

[0053] Assume that the mutually unrelated narrowband positioning signals emitted by K vehicles are all incident on the array antennas of the three cooperative base stations. Assuming that the number of array antennas is M and the array element spacing is d, the received data of the i-th (i=1, 2, 3) base station array at the t-th sampling sample is expressed as:

[0054] y i (t)=A(θ)s(t)+n(t)

[0055] Among them, y i (t) = [y i,1 (t),...,y i,M (t)] T represents the output data of the ith base station array antenna, A(θ)=[a(θ1),...,a(θ K)] represents the array steering matrix, and its kth column is expressed as s(t)=[s1(t),...,s K (t)] T represents the mutually uncorrelated narrowband positioning signal vector, n(t)=[n1(t),...,n M (t)] T is a Gaussian white noise vector, λ represents the carrier wavelength, satisfies λ≥2d, and the superscript T represents the transposition operation;

[0056] Step 3: Calculate the covariance matrix of the data received by each cooperative base station antenna array: Where N is the total number of samples, and the superscript H represents the conjugate transpose operation. After linear shrinkage estimation and sparse vector representation, DOA estimation is obtained using a deep convolutional network;

[0057] Step 4: Based on the DOA estimation results of the three cooperative base stations obtained in step 3, the reliable positioning of the vehicle is completed based on the prediction average criterion and the cross-positioning principle.

[0058] Preferably, in step 3, After linear shrinkage estimation and sparse vector representation, the specific steps of obtaining DOA estimation using a deep convolutional network include:

[0059] Step 3.1: Perform Toeplitz correction to obtain the target matrix R T , expressed as

[0060]

[0061] Among them, J m is an M×M-dimensional shift matrix, J m Only the mth diagonal element is 1, the other elements are 0, and J -m =(J m ) T , J 0 =I M is the identity matrix, Represents the matrix Trace of.

[0062] Step 3.2: Calculate the linear shrinkage coefficient α as

[0063]

[0064] Step 3.3: Get the improved covariance matrix estimate as

[0065]

[0066] Step 3.4: Perform eigenvalue decomposition to obtain K large eigenvalues and the mean of MK small eigenvalues Then we can calculate the unbiased estimate of the noise variance

[0067]

[0068] in, c = M / N;

[0069] Step 3.5: In the overcomplete basis matrix Φ = [b(φ1), ..., b(φ L )]Next Vectorize and subtract the noise term to get the noise-free sparse representation model:

[0070]

[0071] Among them, vec(·) means to vectorize the matrix in the brackets by column, b(φ l )=vec(a(φ l ) H (φ l )), η=[η1,η2,...,η L ] T is a K-sparse column vector.

[0072] Step 3.6: Construct sparse spectrum And substitute it as input value into the designed deep convolutional network for network training and DOA estimation. The designed deep convolutional network consists of an input layer, five hidden layers and an output layer. For each hidden layer, it includes a one-dimensional convolution layer, a batch normalization layer and an activation function. The activation function is f ReLU (x) = max(0, x) nonlinear activation function layer, and for the setting of convolution operation, the number of convolution kernels of the five hidden layers is 24, 20, 12, 5, 1, and the length of convolution kernel is 21, 15, 11, 5, 3 respectively;

[0073] based on Figure 2 The specific process of constructing a deep convolutional neural network and conducting network training and DOA estimation is as follows: Figure 3 As shown, first a large number of training samples are used for preprocessing to form Secondly, construct a sparse spectrum And perform network training, repeat forward propagation, loss calculation, and back propagation to update network parameters. Finally, use the trained network and verification sample sparse spectrum Estimation of DOA is performed on unknown data samples.

[0074] In step 4, the specific process of completing the reliable positioning of the vehicle based on the prediction average criterion and the cross positioning principle is as follows:

[0075] Step 4.1: Based on the DOA estimation results of base station 1 and base station 2 or base station 2 and base station 3, the initial position information of the vehicle is calculated using the cross-positioning principle.

[0076] Assume that the coordinates of base stations 1-3 are (0,Y), (0,0) and (X,0), respectively. The estimated DOA values ​​of the kth vehicle arriving at base stations 1-3 are and Then the kth vehicle position estimation obtained using the DOA estimation results of base station 1 and base station 2 is

[0077]

[0078] The k-th vehicle position estimation obtained using the DOA estimation results of base station 2 and base station 3 is

[0079]

[0080] The k-th vehicle position estimation obtained using the DOA estimation results of base station 1 and base station 3 is

[0081]

[0082] in

[0083] Step 4.2: Use the position estimate of the k-th vehicle to determine whether the k-th vehicle is located on the straight line between the adjacent base stations 1 and 3; if so, average the position estimates of the k-th vehicle obtained by base station 1 and base station 2 and base station 2 and base station 3, and obtain the final positioning result of the k-th vehicle based on the average value obtained by the averaging operation; if not, average the position estimates of the k-th vehicle obtained by each pair of base stations, and obtain the final positioning result of the k-th vehicle based on the average value obtained by the averaging operation.

[0084] The following simulation experiments are used to analyze the positioning performance and computational effectiveness of the vehicle positioning method based on deep convolutional neural network under Toeplitz and sparse prior proposed in the present invention.

[0085] In the experiment, the positions of the three cooperative base stations are B1 (0m, 100m), B2 (0m, 0m), and B3 (100m, 0m), where m represents the unit meter, β1 = β2 = β3 = π / 4, and the number of array elements in each cooperative base station is the same and is 40. The positions of the two vehicles that are close to each other are set to (49m, 29m) and (51m, 29m), respectively. The classic MUSIC method and the deep convolutional network method based only on sparse priors (DCNSP) are selected for comparison. The absolute positioning error obtained from 300 independent Monte Carlo tests is selected as the performance indicator, which is defined as:

[0086]

[0087] in, represents the estimated position of the k-th vehicle in the c-th experiment.

[0088] Table 1 is the simulation result of the absolute error of vehicle positioning obtained by the method of the present invention as the signal-to-noise ratio changes, wherein the number of samples is fixed at 70, and the signal-to-noise ratio changes from -10 decibels (dB) to 10dB; Table 2 is the simulation result of the absolute error of vehicle positioning obtained by the method of the present invention as the number of sampling samples changes, wherein the signal-to-noise ratio is fixed at -5dB, and the number of sampling samples changes from 50 to 250. It can be seen from the simulation results that the method of the present invention can obtain positioning performance that is significantly better than the comparison method. In particular, from Table 1 when the signal-to-noise ratio is ≥5dB, and from Table 2 when the number of samples is ≥200, the method of the present invention can provide centimeter-level positioning accuracy. Table 1 is the simulation result of the absolute error of vehicle positioning as the signal-to-noise ratio changes using the classic MUSIC method, the DCNSP method, and the method of the present invention;

[0089]

[0090] Table 2 shows the simulation results of the absolute error of vehicle positioning with the change of signal-to-noise ratio using the classic MUSIC method, the DCNSP method and the method of the present invention;

[0091]

[0092]

Claims

1. A vehicle localization method based on deep neural network under Toeplitz and sparse prior, characterized by: The method comprises the following steps: Step 1: Arrange the three cooperative base stations in a right triangle, and set the angles between the antenna arrays of the three cooperative base stations and the x-axis in the rectangular coordinate system to β1, β2 and β3 respectively; Step 2: Use the antenna arrays in the three cooperative base stations to receive the vehicle positioning signal and determine the estimated signal form of the arrival angle under the antenna array of each cooperative base station. The specific process is as follows: Assume that the mutually unrelated narrowband positioning signals emitted by K vehicles are all incident on the antenna arrays of the three cooperative base stations. Assuming that the number of antennas in the antenna arrays of the three cooperative base stations is M and the array element spacing is d, the received data of the array of the i-th cooperative base station at the t-th sampling sample is expressed as: y i (t)=A(θ)s(t)+n(t) Among them, y i (t) = [y i,1 (t),...,y i,M (t)] T , i = 1, 2, 3, A(θ) represents the array steering matrix, A(θ) = [a(θ1), ..., a(θ K )], the kth column of A(θ) is expressed as: s(t) represents the mutually uncorrelated narrowband positioning signal vector, s(t) = [s1(t), ..., s K (t)] T , n(t) represents the Gaussian white noise vector, n(t)=[n1(t),...,n M (t)] T , λ represents the carrier wavelength, satisfying λ ≥ 2d, and the superscript T represents the transposition operation; Step 3: Calculate the covariance matrix of the data received by each cooperative base station antenna array: Where N represents the total number of samples, and the superscript H represents the conjugate transpose operation; the covariance matrix After linear shrinkage estimation and sparse vector representation, the DOA estimation result is obtained using a deep network; Step 4: Based on the DOA estimation results of the three cooperative base stations obtained in step 3, the vehicle positioning is completed based on the prediction average criterion and the cross positioning principle; In step 3, the covariance matrix After linear shrinkage estimation and sparse vector representation, the specific steps of using deep convolutional network to obtain DOA estimation results include: Step 3.1: For the covariance matrix Perform Toeplitz correction to obtain the target matrix R T , expressed as Among them, J m represents the M×M-dimensional shift matrix, J m Only the mth diagonal element is 1, the other elements are 0, and J -m =(J m ) T , J 0 =I M represents the identity matrix, Represents taking the matrix traces; Step 3.2: Calculate the linear shrinkage coefficient α as: Step 3.3: Get the improved covariance matrix estimate as: Step 3.4: Perform eigenvalue decomposition to obtain K large eigenvalues and the mean of MK small eigenvalues Then the unbiased estimate of the noise variance is calculated: in, c = M / N; Step 3.5: In the overcomplete basis matrix Φ = [b(φ1), ..., b(φ L )]Next Vectorization and subtracting the noise term yields a noise-free sparse representation model: Among them, vec(·) means to vectorize the matrix in the brackets by column, b(φ l )=vec(a(φ l ) H (φ l )), η=[η1,η2,...,η L ] T is a K sparse column vector; Step 3.6: Construct sparse spectrum The sparse spectrum is substituted as an input value into a deep convolutional network for network training and DOA estimation; the deep convolutional network consists of an input layer, five hidden layers and an output layer; each of the hidden layers includes a one-dimensional convolutional layer, a batch normalization layer and an activation function, and the activation function is f ReLU A non-linear activation function layer with (x)=max(0,x).

2. A vehicle positioning method based on deep neural network under Toeplitz and sparse prior according to claim 1, characterized in that: In step 4, the specific process of completing the reliable positioning of the vehicle based on the prediction average criterion and the cross positioning principle includes the following steps: Step 4.1: The three cooperative base stations are base station 1, base station 2 and base station 3. The initial position information of the vehicle is calculated based on the DOA estimation results of base station 1 and base station 2 or base station 2 and base station 3 using the cross positioning principle; The specific process is as follows: Assume that the coordinates of base station 1, base station 2, and base station 3 are (0, Y), (0, 0), and (X, 0), respectively. The estimated DOA values ​​of the kth vehicle arriving at the three cooperative base stations are and Then the estimated position of the kth vehicle obtained using the DOA estimation results of base station 1 and base station 2 is: The estimated position of the kth vehicle obtained using the DOA estimation results of base stations 2 and 3 is: The estimated position of the kth vehicle obtained using the DOA estimation results of base station 1 and base station 3 is: in Step 4.2: Use the position estimate of the k-th vehicle to determine whether the k-th vehicle is located on the straight line between the adjacent base stations 1 and 3; if so, average the position estimates of the k-th vehicle obtained by base station 1 and base station 2 and base station 2 and base station 3 to obtain the final positioning result of the k-th vehicle; if not, average the position estimates of the k-th vehicle obtained by each pair of base stations to obtain the final positioning result of the k-th vehicle.

Images