Intelligent fusion-based asynchronous transmit-receive split radar networking target tracking method

Abstract

The invention belongs to the technical field of target tracking of phased array networking radar data fusion, and particularly relates to a maneuvering target tracking problem of asynchronous transmit-receive split radar networking. According to the invention, a fusion evaluation module is introduced, and the module can intelligently select a sequential fusion method or a distributed fusion method based on time registration to optimize local target estimation according to actual scene requirements. And for a single receiving node, the obtained bistatic radar measurement data is accurately converted to a Cartesian coordinate system by adopting a depolarization measurement conversion method. For cooperative processing of multiple receiving nodes, measurement data of all the receiving nodes are integrated into a unified matrix, and a bistatic radar maneuvering target tracking algorithm (CTDCMKF-IMM) based on centralized processing is constructed in combination with an interactive multimode algorithm. In combination with a fusion evaluation module, an asynchronous transmit-receive split radar networking target tracking method (AIF-BR) based on intelligent fusion is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of target tracking for phased array networking radar data fusion, and specifically relates to the problem of maneuvering target tracking for asynchronous transmit-receive separated radar networking. Background Art

[0002] Due to the characteristics of transmit-receive separation, the transmit-receive separated radar system has significant advantages in anti-interference and concealment. For an asynchronous transmit-receive separated radar networking system, the use of information fusion technology can improve the target tracking performance. By organically combining the asynchronous measurements or local state estimation results of multiple receiving nodes, this technology overcomes the limitations of a single transmit-receive separated radar system in target tracking, thereby significantly improving the estimation accuracy of the target state.

[0003] In the field of multi-radar data fusion, Bar-Shalom Y. et al. systematically studied the applications of Kalman filtering and its extended forms (Bar-Shalom Y, Willett P K, Tian X. Tracking and data fusion[M]. Storrs, CT, USA: YBS publishing, 2011.). There are three methods for asynchronous radar networking, namely centralized, distributed, and sequential fusion (Deng Zili. Information fusion estimation theory and its applications[M]. Beijing: Science Press, 2012). Centralized fusion processes by registering the measurement information of each radar node to a unified time reference and then transmitting it to the fusion center for processing. It can make full use of the complete measurement data of all nodes, thus obtaining an accurate target state estimate. However, the computational complexity of this method increases significantly with the increase in data volume, posing high requirements on the system's computational resources. In contrast, in the distributed fusion method, each node first preprocesses the acquired measurement data to obtain a local state estimate, and after time registration, it is transmitted to the fusion center for final fusion (Li G, Battistelli G, Chisci L, et al. Distributed multi-target tracking over an asynchronous multi-sensor network[C]. 2020 IEEE Radar Conference (RadarConf20). IEEE, 2020: 1-6.). Although this method has a lower computational complexity and higher system reliability, its fusion result is usually not globally optimal. Sequential fusion adopts a recursive processing method. According to the time order of measurement acquisition, each node uses the registration estimation result of the previous node as the prediction input to obtain the final state estimate of the target for system fusion (Zhang D, Duan Z, Hanebeck U D. Asynchronous multi-radar tracking fusion with converted measurements[C]. 2022 25th International Conference on Information Fusion (FUSION). IEEE, 2022: 1-8.). This method effectively reduces the system's computational burden by reducing the operation dimension.

[0004] In the practical application of asynchronous transceiver separated radar networked target tracking, due to the large amount of measurement data obtained by the system, sequential fusion has a smaller computational burden compared to centralized fusion and can achieve real-time data processing without waiting for all measurement data. When the measurement nonlinearity is relatively low, the tracking performance of sequential fusion is better than that of distributed methods. However, when the measurement nonlinearity is relatively high, the recursive nature of sequential fusion may lead to error accumulation and affect the fusion accuracy. At this time, the distributed fusion method based on time registration can effectively make up for this deficiency and provide more accurate estimation performance for the system.

[0005] Based on this, the present invention introduces a fusion evaluation module, which can intelligently select sequential fusion or the distributed fusion method based on time registration according to the actual scenario requirements to optimize the local target estimation. For a single receiving node, the debiased measurement conversion method is used to accurately convert the obtained bistatic radar measurement data to the Cartesian coordinate system (Marom H, Bar-Shalom Y, Milgrom B. Converting bistatic radar measurements to Cartesian position for tracking [C]. Radar Sensor Technology XXVI. SPIE, 2022, 12108: 88-98.). For the collaborative processing of multiple receiving nodes, the present invention integrates the measurement data of all receiving nodes into a unified matrix and combines the interacting multiple model algorithm (Mazor E, Averbuch A, Bar-Shalom Y, et al. Interacting multiple model methods in target tracking: a survey [J]. IEEE Transactions on Aerospace and Electronic Systems, 1998, 34(1): 103-123.) to construct a bistatic radar maneuvering target tracking algorithm (CTDCMKF-IMM) based on centralized processing. Combining the fusion evaluation module, a method for asynchronous transceiver separated radar networked target tracking based on intelligent fusion (AIF-BR) is proposed. Summary of the Invention

[0006] It is assumed that there are M transmitting nodes and N receiving nodes in the asynchronous transceiver separated radar networked system. At time k, the position of the m-th transmitting node is expressed as The position of the n-th receiving node is The sampling period is At the same time, each receiving node can obtain M measurements. Define the propagation path from the m-th transmitting node to the target and from the target to the n-th receiving node as channel (m,n). The measurement information obtained by the receiving node for a single channel includes the two-way distance and the azimuth measurement and its corresponding measurement noise is zero-mean additive Gaussian white noise, and the measurement variances are respectively and Assume that the coordinate systems of each receiving node are parallel to the unified coordinate system, with only spatial translation changes. A method for target tracking in an asynchronous transceiver-separated radar network based on intelligent fusion starts from time k sta The filtering steps are as follows:

[0007] Step 1: The receiving node n obtains measurements at time k sta Based on the CTDCMKF-BR-IMM algorithm, the measurement data is processed to obtain the state estimation results of the target for all models and its error covariance and the final target state estimation is obtained through the updated model probability where n represents the set of receiving nodes that obtain measurements at the current time.

[0008] Step 1.1 Input interaction estimation. Taking the receiving node n in the set n as an example, the processing of other nodes is similar, and the initial value of k is k sta .

[0009] First, assume that there are J models in the IMM multi-model system, and the motion model is expressed as Calculate the state estimation of the input of each model filter according to the following formula and the covariance

[0010]

[0011] where respectively represent the target state estimation and its error covariance matrix of model l at time k1, represents the model mixing probability, and its calculation method is as follows:

[0012]

[0013] where represents the mixing probability of motion model l at time k-1, represents the Markov model transition probability from model l to model j, is the update probability of motion model l at time k-1, c j is the normalization constant, as shown in the following formula:

[0014]

[0015] Step 1.2 Calculate the state prediction and prediction error covariance of each model.

[0016]

[0017]

[0018] Among them, are the interactive input estimate and its error covariance obtained in Step 1.1 respectively, represent the state transition matrix and noise driving matrix of model j respectively, represents the autocorrelation matrix of the process noise.

[0019] Step 1.3 Based on the measurement results, perform debiased measurement transformation on the M measurements obtained by a single receiving node at time k, and splice the transformed measurements, transformation error covariance with the linear measurement matrix.

[0020]

[0021]

[0022] Among them, represents the vector composed of the debiased measurement transformation results of all channels obtained by node n, is the matrix composed of all transformation error covariances, is the spliced linear measurement matrix,

[0023] Step 1.4 Perform linear Kalman filtering processing and spatial registration on each model to obtain the target state estimate and the error covariance matrix and calculate its likelihood function

[0024]

[0025] Among them, represents the state vector of the nth receiving node.

[0026] Step 1.5 Update the model probability according to the likelihood function of each model.

[0027]

[0028] Among them, C j is the normalization constant, and its expression is as follows:

[0029]

[0030] Step 1.6 Calculate the final target state estimate of node n based on the updated model probability

[0031]

[0032] Step 2: Intelligently select the fusion strategy. When is satisfied, perform sequential fusion and enter Step 3; otherwise, perform distributed fusion and enter Step 4.

[0033]

[0034] where γ th is the threshold value, and n f represents the serial number of the node with the largest number in set n. is calculated as follows:

[0035]

[0036] where and represent the standard deviations of the two-way distance and azimuth measurement errors of receiving node n f respectively, and represents the two-way distance estimate, and the expression is as follows:

[0037]

[0038] where is the position of receiving node n f at time k, is the position of transmitting node M, represents the estimated position of the target.

[0039] Step 3: Perform sequential fusion processing on the asynchronous measurement data obtained by all nodes to obtain the state estimate of the target and its error covariance

[0040] Step 3.1 Determine the moment k sta closest to the current moment k near for each node's measurement moment, and record the corresponding node serial number n.

[0041] near

[0042] Step 3.2 Use as the interactive prediction input of radar i near at time k near , and through the CTDCMKF-IMM algorithm processing as shown in Eqs. (5) and (6), node i near obtains the state estimation results of each motion model of the updated target as shown in Eqs. (7) to (13) and its error covariance where n1 represents the serial number of the node with the smallest number in set n.

[0043] Step 3.3 Repeat Steps 3.1 - 3.2 to obtain k N the state estimation results of the radar N for each motion model of the target at time k and its error covariance

[0044] Step 3.4 Based on the updated model probability, obtain the global target state estimation

[0045] Step 4: Register the asynchronous target state estimations obtained by all nodes to a unified time reference for distributed fusion processing to obtain the state estimation of the target and its error covariance

[0046] Step 4.1 Determine the serial number i of the other nodes except the receiving node in set n rest .

[0047] Step 4.2 Node i rest obtains the local target state estimation through the CTDCMKF - BR algorithm as shown in Eqs. (1) - (18).

[0048] Step 4.3 Register the state estimation and to time k f .

[0049]

[0050] Step 4.4 Based on the distributed fusion method, obtain the fused state estimation of the target.

[0051]

[0052] where L represents the number of nodes in set n.

[0053] Principle of the invention

[0054] For a bistatic multi - radar system, a single receiving node can obtain the two - way range and azimuth measurement information of the target. The measurement equation of the system can be expressed as:

[0055]

[0056] where represents the vector composed of the measurements obtained by all receiving nodes at time k, is the non - linear measurement function, is the corresponding measurement noise vector. Define the propagation path from the m-th transmitting node to the target and from the target to the n-th receiving node as channel (m,n), and its measurement equation is as follows:

[0057]

[0058] where, r k =(x k ,y k ) represents the true position of the target. is a Gaussian white noise with zero mean and variance . can be expressed as:

[0059]

[0060] where, respectively represent the round-trip distance and the azimuth measurement error variance of receiving node n.

[0061] Next, the process of debiased measurement transformation for a single channel is given. To simplify the formula, channel (m,n) is omitted in the following derivation. According to the geometric structure of the bistatic radar, the one-way distance measurement of the target relative to the receiving node can be solved using the cosine theorem of a triangle:

[0062]

[0063] where, L k represents the distance between the receiving node and the transmitting node at time k, and its expression is as follows:

[0064]

[0065] is defined as follows:

[0066]

[0067] So far, based on the one-way distance measurement and the azimuth measurement , combining the conversion relationship between the polar coordinate system and the Cartesian coordinate system, measurement transformation can be performed:

[0068]

[0069] where, x k ,y k represent the true positions of the target in the x and y directions, represents the corresponding measurement transformation error.

[0070] Next, calculate the statistical characteristics of the measurement transformation error . From equation (32), it can be seen that the position measurement is related to both the two-way distance measurement and the azimuth measurement, and they are respectively expanded by the second-order Taylor series at the true values x k and y k The expanded expressions are as follows:

[0071]

[0072] wherein, it is defined that respectively represent the two-way distance and azimuth measurement of the bistatic radar.

[0073] Taking the expectation of the above formula, we can get:

[0074]

[0075] wherein, From Equation (34), it can be seen that there is an additive bias in the measurement vector Therefore, the method of subtraction and bias removal is adopted to eliminate this bias. The debiased measurement can be expressed as:

[0076]

[0077] wherein, represents the position measurement vector after the debiased measurement conversion, represents the linear measurement matrix, and respectively represent the means of the measurement conversion errors and as shown in the second half of Equation (34), The calculation of

[0078]

[0079]

[0080] is the error of the debiased measurement conversion, and its mean value and covariance matrix can be expressed as:

[0081]

[0082] The specific expressions of each element in

[0083]

[0084]

[0085] wherein, the calculation of each first-order derivative is as follows:

[0086]

[0087] and The calculation of the second-order partial derivative is as follows:

[0088]

[0089] So far, the single-channel debiased measurement conversion result and its error covariance matrix can be obtained. By integrating them as shown in Eqs. (7) and (8), and combining the interacting multiple model for Kalman filtering, the state estimation result of the target and its error covariance are obtained, as shown in Eqs. (10) to (18).

[0090] Since there is a strong non-linearity between the measurement information obtained by the bistatic radar and the target motion state, this non-linearity affects the target tracking performance of the system. By introducing the variable as the fusion evaluation index, as shown in Eq. (20), when exceeds the threshold γ th at this time, after the measurement conversion the error no longer approximately follows a Gaussian distribution, resulting in the accumulation of errors in sequential fusion. At this time, the local target state estimations of each node are registered to a unified time reference, and distributed fusion is performed to obtain the fusion estimation result, which will be used as the prediction input for the receiving node at the next measurement moment; conversely, when is less than the threshold γ th at this time, the sequential fusion method can provide a more accurate target state estimation, so the output result is used as the prediction input value for the receiving node at the next measurement moment.

[0091] The present invention proposes a method for tracking networked targets of asynchronous bistatic radars based on intelligent fusion. This method dynamically evaluates the non-linearity of the measurements obtained by the receiving nodes and intelligently selects the optimal fusion strategy, as shown in Eq. (19). When at this time, the system performs sequential fusion; conversely, when the system performs distributed fusion. This intelligent fusion mechanism significantly improves the target tracking performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 is the functional block diagram of the intelligent fusion mechanism of the present invention;

[0093] Figure 2 is the structural block diagram of the CTDCMKF-BR-IMM algorithm;

[0094] Figure 3 is the comparison of the target position RMSE between the algorithm of the present invention and each receiving node in the simulation scenario;

[0095] Figure 4 RMSE comparison of the target speeds of the algorithm of the present invention and each receiving node in the simulation scenario;

[0096] Figure 5 RMSE comparison of the target positions of the fusion algorithm in the simulation scenario;

[0097] Figure 6 RMSE comparison of the target positions of the fusion algorithm in the simulation scenario.

[0098] Figure 7 Probability schematic diagram of the maneuvering target tracking model of the algorithm of the present invention in the simulation scenario.

[0099] Figure 8 Schematic diagram of the intelligent fusion strategy of the algorithm of the present invention in the simulation scenario. Detailed implementation manners

[0100] Consider a scenario of a multi-radar system with asynchronous transmit-receive separation. The system includes four transmitting nodes located at (20 km, 40 km), (20 km, 80 km), (70 km, 40 km), and (70 km, 80 km), and three receiving nodes (20 km, 20 km), (46 km, 10 km), and (70 km, 20 km). Their sampling periods are respectively and Assume that the measurement accuracies of all receiving nodes are the same, the radial distance standard deviation σ r = 50 m, and the azimuth standard deviation σ θ = 0.2°. The target starts moving from the initial position (35 km, 98 km) with an initial velocity of (300 m / s, -80 m / s): it moves in a uniform straight line (CV model) within 0 - 40 s and 60 s - 100 s, and makes a uniform-rate turning motion (CT model) with an angular velocity of -30° / s within 60 s - 75 s. The initial probabilities of both motion models are set to 0.5. The initial state estimation of the radar is obtained using the two-point initialization method, and the Monte Carlo simulation experiment is repeated 500 times.

[0101] The target tracking is realized by using the asynchronous transmit-receive separated radar networking target tracking method based on intelligent fusion (AIF-BR) proposed by the present invention. At the same time, to illustrate the advantages of the algorithm of the present invention, its performance is compared with the individual filtering of receiving nodes 1, 2, and 3 (Receiver-1 / Receiver-2 / Receiver-3), the target tracking method that only uses distributed fusion (ADF-BR), and the target tracking method that only uses sequential fusion (ASF-BR). Among them, the fusion evaluation threshold γ of the AIF-BR algorithm th= 0.23. The root mean square error (RMSE) performance index of position and velocity estimation is adopted below to evaluate the tracking performance of the target.

[0102] Figure 1 The schematic diagram of the intelligent fusion mechanism of the AIF-BR algorithm of the present invention is shown.

[0103] Figure 2 The schematic diagram of the CTDCMKF-IMM algorithm adopted by the present invention is shown.

[0104] Figure 3 and Figure 4 respectively show the comparison diagrams of the target position and velocity RMSE between the AIF-BR algorithm and each receiving node. It can be seen from the figure that the AIF-BR algorithm can significantly improve the filtering performance of a single receiving node.

[0105] Figure 5 and Figure 6 respectively show the comparison diagrams of the target position and velocity RMSE between the AIF-BR algorithm and ADF-BR, ASF-BR. It can be seen from the figure that the AIF-BR algorithm can significantly improve the filtering performance of a single receiving node. The tracking accuracy of the ADF-BR algorithm is relatively low at 7 - 31 s. The reason is that the state initialization accuracy of each asynchronous receiving node for the target is insufficient, resulting in a worse distributed fusion effect than sequential fusion. Although the ASF-BR algorithm has an advantage in the early-stage position estimation of the target, its estimation accuracy for the target velocity is inaccurate. This is because in a bistatic radar system, velocity estimation is highly sensitive to measurement noise, which will cause the estimation error to continuously accumulate and amplify during the sequential processing. In contrast, the AIF-BR algorithm effectively improves the overall performance of target tracking by adaptively optimizing the selection of the fusion strategy. It has the smallest position and velocity RMSE and demonstrates the optimal tracking performance.

[0106] Figure 7 The schematic diagram of the model probability distribution during the tracking process of a maneuvering target by the AIF-BR algorithm is shown. The results show that the model probability distribution of the algorithm effectively reflects the motion law of the target.

[0107] Figure 8 The schematic diagram of the fusion method selection during the target tracking process by the AIF-BR algorithm is shown. It can be observed from the figure that the AIF-BR algorithm can dynamically switch between the sequential fusion and distributed fusion strategies according to the actual scenario, fully reflecting the intelligent feature of the algorithm in the selection of the fusion strategy.

[0108] In summary, the proposed method for multi-radar target tracking with asynchronous transceiver separation based on intelligent fusion in the present invention is an effective target tracking algorithm for an asynchronous transceiver-separated radar networking system.

Claims

1. An asynchronous transceiver separated radar networking target tracking method based on intelligent fusion, and the specific technical solution is as follows: Assume that there are M transmitting nodes and N receiving nodes in a bistatic radar networking system. At time k, the position of the m-th transmitting node is denoted as The position of the n-th receiving node is The sampling period is Each receiving node can obtain M measurements at the same time. Define the propagation path from the m-th transmitting node to the target and from the target to the n-th receiving node as channel (m,n). The measurement information obtained by the receiving node for a single channel includes the round-trip distance and the azimuth measurement The corresponding measurement noise is zero-mean additive Gaussian white noise, and the measurement variances are and Assume that the coordinate systems of the receiving nodes are parallel to the unified coordinate system, with only spatial translation changes. An asynchronous transceiver separated radar networking target tracking method based on intelligent fusion starts from time k sta The filtering steps starting from this moment are as follows: Step 1: The receiving node n obtains measurements at time k sta and processes the measurement data based on the CTDCMKF-BR-IMM algorithm to obtain the state estimation results of the targets of all models and their error covariances and obtains the final target state estimation through the updated model probabilities where n represents the set of receiving nodes that obtain measurements at the current time. Step 1.1 Input interaction estimation. Taking the receiving node n in set n as an example, the processing of other nodes is similar, and the initial value of k is k sta . First, assume that there are J models in the IMM multi-model system, and the motion model is expressed as Calculate the state estimate of the input of each model filter according to the following formula and covariance Among them, respectively represent the target state estimation of model l and its error covariance matrix at time k1, represents the model mixing probability, and its calculation method is as follows: Among them, represents the mixed probability that the motion model at time k - 1 is l, represents the Markov model transition probability that model l transitions to model j, is the update probability of the motion model l at time k - 1, c j is the normalization constant, as shown in the following formula: Step 1.2 Calculate the state prediction and prediction error covariance of each model. wherein, are respectively the interaction input estimation and its error covariance obtained in Step 1.1, respectively represent the state transition matrix and the noise driving matrix of model j, represents the autocorrelation matrix of the process noise. Step 1.3 Based on the measurement results, perform debiased measurement conversion on the M measurements obtained by a single receiving node at time k, and splice the converted measurements, conversion error covariance and the linear measurement matrix. Among them, represents the vector composed of the debiased measurement conversion results of all channels obtained by node n, is the matrix composed of all conversion error covariances, is the spliced linear measurement matrix, Step 1.4 performs linear Kalman filtering and spatial registration on each model to obtain the target state estimate and the error covariance matrix and calculates its likelihood function Among them, represents the state vector of the nth receiving node. Step 1.5 Update the model probability according to the likelihood function of each model. where C j is a normalization constant, and its expression is as follows: Step 1.6 Calculate the final target state estimate of node n based on the updated model probability Step 2: Intelligently select the fusion strategy. When occurs, perform sequential fusion and proceed to Step 3; otherwise, perform distributed fusion and proceed to Step 4. Among them, γ th is the threshold value, and n f represents the serial number of the node with the largest number in the set n, is calculated as follows: Among them, and respectively represent the standard deviations of the two-way distance and azimuth measurement errors of the receiving node n f , and the two-way distance estimation is represented by , and the expression is as follows: The two-way distance estimation is represented by , and the expression is as follows: Among them, is the position of receiving node n at time k f ; is the position of transmitting node M; represents the estimated position of the target. Step 3: Perform sequential fusion processing on the asynchronous measurement data obtained by all nodes to obtain the state estimate of the target and its error covariance Step 3.1 Determine the measurement time of each node and the current time k sta The closest time k near , and record the corresponding node number n near . Step 3.2 Take as k near the interactive prediction input of radar i at time near , as shown in Eqs. (5) and (6), node i near is processed by the CTDCMKF-IMM algorithm, as shown in Eqs. (7) to (13), to obtain the state estimation results of each motion model of the updated target and its error covariance where n1 represents the serial number of the node with the smallest number in set n. Step 3.3 Repeat Steps 3.1 - 3.2 to obtain k N State estimation results of the radar N for each motion model of the target at the moment and their error covariances Step 3.4 Obtain the global target state estimation based on the updated model probability Step 4: Register the asynchronous target state estimates obtained by all nodes to a unified time reference for distributed fusion processing to obtain the state estimate of the target and its error covariance Step 4.1 Determine the sequence number i of other nodes except the receiving nodes in set n rest . Step 4.2 Node i rest Obtain the local target state estimate through the CTDCMKF-BR algorithm As shown in Equations (1) to (18). Step 4.3 Perform state estimation and register it to the k f moment. Step 4.4 Based on the distributed fusion method, obtain the fusion state estimation of the target. Wherein, L represents the number of nodes in the set n.

Images