Target tracking method and system based on classification information assistance

Abstract

The application relates to a target tracking method and system based on classification information assistance, which comprises the following steps: firstly, calculating target motion state prediction and target existence state prediction at the current time; then, combining spatial information of sensor measurement at the current time to calculate associated spatial probability and associated prior probability of the target and the measurement; then, calculating an associated initial confidence set of the target and the measurement according to the two probabilities; then, extracting features from radar range-Doppler spectrum information by using a radar range-Doppler spectrum feature extraction model; then, calculating an associated confidence set of the target and the measurement after fusing classification confidence obtained by using a target classification model and the associated initial confidence of the target and the measurement; then, calculating an update message of the measurement according to the associated confidence set and the spatial information of the measurement; and finally, calculating motion state estimation and existence state estimation of the target at the current time according to the update message and the target motion state prediction and the target existence state prediction. The application can significantly improve the performance and robustness of target tracking.

Description

Technical Field

[0001] This invention relates to the field of radar data processing technology, and in particular to a target tracking method and system based on classification information assistance. Background Technology

[0002] Radar target tracking is a crucial component of radar data processing, widely applied in fields such as civil aviation and long-range early warning. The purpose of radar target tracking is to estimate the target's motion over a given period of time based on measurement data provided by the radar.

[0003] In existing technologies, radar target tracking methods typically employ a detection-after-tracking framework. First, the detector sets a detection threshold and generates candidate measurements. Then, the tracker estimates the target's motion state to achieve target tracking. However, radar target tracking algorithms only use spatial information from radar measurements for target tracking, neglecting other feature information from radar signal processing. In complex cluttered environments such as clouds, rain, land, and sea, using a high detection threshold may lead to missed target detections, while a low threshold may result in a large amount of clutter, severely degrading target tracking performance. Summary of the Invention

[0004] Therefore, it is necessary to provide a target tracking method and system based on classification information that can significantly improve the performance and robustness of target tracking, addressing the aforementioned technical problems.

[0005] In a first aspect, the present invention provides a target tracking method based on classification information assistance, comprising the following steps:

[0006] Based on the target motion state estimate and target existence state estimate of the previous moment, calculate the target motion state prediction and target existence state prediction of the current moment.

[0007] The spatial probability of the association between the target and the measurement is predicted and calculated by combining the spatial information measured by the sensor at the current moment with the target motion state at the current moment, and the prior probability of the association between the target and the measurement is predicted and calculated based on the target existence state at the current moment.

[0008] Calculate the initial confidence set of the association between the target and the measurement based on the spatial probability and prior probability of association between the target and the measurement;

[0009] The radar range-Doppler spectrum feature extraction model is used to extract features from the radar range-Doppler spectrum information to obtain the feature vector of the radar range-Doppler spectrum information;

[0010] The association confidence set between the target and the measurement is calculated by integrating the classification confidence obtained from the target classification model with the initial association confidence between the target and the measurement;

[0011] Based on the association confidence set between the target and the measurement and the spatial information of the measurement, calculate the update message of the measurement;

[0012] Based on the measurement update messages and the target motion state prediction and target existence state prediction at the current moment, calculate the target motion state estimate and existence state estimate at the current moment.

[0013] In one embodiment, calculating the association confidence set between the target and the measurement after fusing the classification confidence obtained using the target classification model with the initial association confidence between the target and the measurement includes the following steps:

[0014] The feature vectors of radar range-Doppler spectrum information and the initial confidence set of target-measurement association are input into the target classification model to obtain classification confidence;

[0015] The initial confidence sets for the association between classification confidence and target and measurement are fused using the DS fusion rule;

[0016] The iterative correlation confidence set is solved based on the fusion result of the initial confidence set of the correlation between classification confidence and target and measurement;

[0017] Determine whether the number of times the classification confidence and the initial confidence set of the target and measurement association have been fused has reached the maximum number of iterations. If not, use the iterative association confidence set as the initial confidence set of the target and measurement association and return the step of inputting the feature vector of radar range-Doppler spectrum information and the initial confidence set of the target and measurement association into the target classification model to obtain the classification confidence. If yes, use the iterative association confidence set as the confidence set of the target and measurement association.

[0018] In one embodiment, calculating the association confidence set between the target and the measurement based on the fusion result of the classification confidence set and the initial confidence set of the association between the target and the measurement includes:

[0019] The confidence set of the associated events where the measurement is clutter is replaced in the initial confidence set of the association between the target and the measurement with the fusion result of the classification confidence and the initial confidence set of the association between the target and the measurement, and the replaced initial confidence set of the association is obtained.

[0020] The replaced initial confidence set is input into the confidence propagation algorithm to obtain the iterative confidence set.

[0021] In one embodiment, the target motion state estimate and target existence state estimate at the previous moment are used to calculate the target motion state prediction and target existence state prediction at the current moment. This is done by calculating the target motion state prediction at the current moment based on the target motion state estimate and target motion model at the previous moment, and calculating the target existence state prediction at the current moment based on the target existence state estimate and target existence state transition equation at the previous moment.

[0022] In one embodiment, calculating the initial confidence set for the association between the target and the measurement based on the spatial probability and prior probability of association between the target and the measurement includes:

[0023] Calculate the product of the spatial probability of the association between the target and the measurement and the prior probability of the association;

[0024] The product of the spatial probability of the association between the target and the measurement and the prior probability of the association is used as the input of the belief propagation algorithm to solve for the initial confidence set of the association between the target and the measurement.

[0025] In one embodiment, based on the association confidence set between the target and the measurement and the spatial information of the measurement, the update message of the measurement is calculated by weighting the target's measurement using the association confidence set between the target and the measurement to calculate the equivalent measurement of the target, thereby obtaining the update message of the target's motion state, and calculating the update message of the target's existence state based on the statistical dependency of the target's existence state on the association confidence set between the target and the measurement.

[0026] In one embodiment, the motion state estimate and existence state estimate of the target at the current moment are calculated based on the measurement update message and the target motion state prediction and the target existence state prediction at the current moment. This is achieved by inputting the measurement update message and the target motion state prediction at the current moment into a Kalman filter to obtain the target motion state estimate at the current moment, and by multiplying the measurement update message and the target existence state prediction at the current moment and normalizing the result to obtain the target existence state estimate at the current moment.

[0027] In one embodiment, the radar range-Doppler spectrum feature extraction model includes three convolutional layers, one flattening layer, and three fully connected layers sequentially from input to output, while the target classification model includes three fully connected layers.

[0028] Secondly, the present invention also provides a target tracking system based on classification information, including: a state prediction calculation module, used to calculate the target motion state prediction and the target existence state prediction at the current moment based on the target motion state estimation and the target existence state estimation at the previous moment;

[0029] The associated prior probability calculation module is used to predict and calculate the associated spatial probability between the target and the measurement by combining the spatial information measured by the sensor at the current moment with the target motion state at the current moment, and to predict and calculate the associated prior probability between the target and the measurement based on the target existence state at the current moment.

[0030] The module for calculating the initial confidence set of association is used to calculate the initial confidence set of association between the target and the measurement based on the spatial probability and prior probability of association between the target and the measurement.

[0031] The feature extraction module is used to extract features from radar range-Doppler spectrum information using a radar range-Doppler spectrum feature extraction model, and obtain the feature vector of radar range-Doppler spectrum information.

[0032] The fusion module is used to fuse the classification confidence obtained by the target classification model with the initial confidence of the association between the target and the measurement, and then calculate the association confidence set between the target and the measurement.

[0033] The update message calculation module is used to calculate the update message of the measurement based on the association confidence set between the target and the measurement and the spatial information of the measurement.

[0034] The state estimation calculation module is used to calculate the target's motion state estimate and existence state estimate at the current moment based on the measurement update message and the target motion state prediction and target existence state prediction at the current moment.

[0035] The beneficial effects of this invention are as follows: This invention utilizes the characteristic information of radar signal processing, such as radar range-Doppler spectrum, to determine the classification confidence that the measurement is generated by the target or by clutter. Then, the fusion result of the classification confidence and the initial confidence set associated with the measurement replaces the associated events of the measurement as clutter in the initial confidence set associated with the target and the measurement. The classification confidence, as additional information for target tracking, achieves the suppression of clutter, thereby significantly improving the performance and robustness of target tracking. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a target tracking method based on classification information provided in an embodiment of the present invention;

[0037] Figure 2 This is a flowchart illustrating a target tracking method based on classification information in another embodiment of the present invention.

[0038] Figure 3 This is a flowchart illustrating a target tracking method based on classification information in another embodiment of the present invention.

[0039] Figure 4 This is a flowchart illustrating a target tracking method based on classification information in another embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of the factor flow provided in an embodiment of the present invention;

[0041] Figure 6 This is a schematic diagram of the radar range-Doppler spectrum feature extraction model provided in an embodiment of the present invention;

[0042] Figure 7 This is a schematic diagram of the target classification model provided in an embodiment of the present invention;

[0043] Figure 8 This is a map showing the actual target trajectory in the RD domain;

[0044] Figure 9 Two sets of sea clutter echo maps in the time domain and RD domain;

[0045] Figure 10 This is a schematic diagram of the detection results of the constant false alarm rate (CFAR) detector, where, Figure 10 (a) is a schematic diagram of the detection results at -8dB. Figure 10 (b) is a schematic diagram of the detection results at -4dB. Figure 10 (c) is a schematic diagram of the detection results at 0dB. Figure 10 (d) is a schematic diagram of the detection results at 4dB;

[0046] Figure 11 This is a schematic diagram illustrating the tracking results of NEMP and MP, where, Figure 11 (a) is a schematic diagram of the NEMP tracking results. Figure 11 (b) is a schematic diagram of the MP tracking results;

[0047] Figure 12 A schematic diagram illustrating the Monte Carlo mean root square error of target distance and velocity estimation over time;

[0048] Figure 13 A schematic diagram illustrating the Monte Carlo mean OSPA for estimating the target's motion state over time.

[0049] Figure 14 Comparison of target tracking performance for different SCRs. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] In one embodiment, such as Figure 1 As shown, Figure 1 This is one of the flowcharts of a target tracking method based on classification information assisted by an embodiment of the present invention. The target tracking method based on classification information assisted by this embodiment includes the following steps:

[0052] S101. Based on the target motion state estimate and target existence state estimate of the previous moment, calculate the target motion state prediction and target existence state prediction of the current moment.

[0053] The target motion state includes the target's position, velocity, and acceleration, while the target existence state includes both existence and non-existence.

[0054] S102. Calculate the spatial probability of the target and the measurement by combining the spatial information measured by the sensor at the current moment with the target motion state at the current moment, and calculate the prior probability of the target and the measurement based on the target existence state at the current moment.

[0055] Among them, the associated spatial probability represents the likelihood probability of measurement generated by the target, and the associated prior probability represents the likelihood probability of measurement generated by clutter given the state of the target.

[0056] S103. Calculate the initial confidence set of the association between the target and the measurement based on the spatial probability and prior probability of the association between the target and the measurement.

[0057] S104. Use the radar range-Doppler spectrum feature extraction model to extract features from the radar range-Doppler spectrum information and obtain the feature vector of the radar range-Doppler spectrum information.

[0058] Specifically, the radar range-Doppler spectrum feature extraction model consists of three convolutional layers, one flattening layer, and three fully connected layers along the input-to-output direction.

[0059] S105. After integrating the classification confidence obtained from the target classification model with the initial confidence of the association between the target and the measurement, calculate the association confidence set between the target and the measurement.

[0060] S106. Based on the association confidence set between the target and the measurement and the spatial information of the measurement, calculate the update message of the measurement. The update message is used to calculate the motion state estimate and existence state estimate of the target at the current time.

[0061] S107. Calculate the target's motion state estimate and existence state estimate at the current moment based on the measurement update message and the target motion state prediction and target existence state prediction at the current moment.

[0062] This embodiment of the target tracking method based on classification information combines classification information obtained by classifying measurements using radar range-Doppler spectrum information with the initial confidence of the association between the target and the measurement. Then, based on the combined confidence set of the association between the target and the measurement and the spatial information of the measurement, it calculates the motion state estimate and existence state estimate of the target at the current moment, thereby achieving target tracking. The classification information obtained by classifying measurements using radar range-Doppler spectrum information, as supplementary information, can enhance clutter suppression and data association, thereby improving the performance and robustness of target tracking.

[0063] In one embodiment, such as Figure 2 As shown, Figure 2This is one of the flowcharts of a target tracking method based on classification information provided in this embodiment of the invention. This embodiment involves how to integrate the classification confidence obtained by using the target classification model with the initial confidence of the association between the target and the measurement, and then calculate the association confidence set between the target and the measurement. Based on the above embodiment, step S105 includes the following steps:

[0064] S201. Input the feature vector of radar range-Doppler spectrum information and the initial confidence set of target-measurement association into the target classification model to obtain classification confidence.

[0065] Specifically, the target classification model includes three fully connected layers.

[0066] S202. Use the DS fusion rule to fuse the initial confidence set of classification confidence and the association between the target and the measurement;

[0067] S203. Solve for the iterative correlation confidence set based on the fusion result of the initial confidence set of the classification confidence and the correlation between the target and the measurement.

[0068] S204. Determine whether the number of times the classification confidence and the initial confidence set of the target and measurement association have been fused has reached the maximum number of iterations. If not, use the iterative association confidence set as the initial confidence set of the target and measurement association and return the step of inputting the feature vector of the radar range-Doppler spectrum information and the initial confidence set of the target and measurement association into the target classification model to obtain the classification confidence. If yes, use the iterative association confidence set as the confidence set of the target and measurement association.

[0069] In one embodiment, such as Figure 3 As shown, Figure 3 This is one of the flowcharts of a target tracking method based on classification information provided in this embodiment of the invention. This embodiment involves how to calculate the association confidence set between the target and the measurement based on the fusion result of the classification confidence and the initial confidence set of the association between the target and the measurement. Based on the above embodiment, step S203 includes:

[0070] S301. Replace the confidence set of the associated events where the measurement is clutter in the initial confidence set of the association between the target and the measurement with the fusion result of the classification confidence and the initial confidence set of the association between the target and the measurement, and obtain the replaced initial confidence set of the association.

[0071] S302. Input the replaced initial confidence set of association into the confidence propagation algorithm to obtain the iterative confidence set of association.

[0072] In one embodiment, the target motion state estimate and target existence state estimate at the previous moment are used to calculate the target motion state prediction and target existence state prediction at the current moment. This is done by calculating the target motion state prediction at the current moment based on the target motion state estimate and target motion model at the previous moment, and calculating the target existence state prediction at the current moment based on the target existence state estimate and target existence state transition equation at the previous moment.

[0073] In one embodiment, such as Figure 4 As shown, Figure 4 This is one of the flowcharts of a target tracking method based on classification information provided in this embodiment of the invention. This embodiment involves how to calculate the initial confidence set of the association between the target and the measurement based on the spatial probability and prior probability of the association between the target and the measurement. Based on the above embodiment, step S103 includes:

[0074] S401. Calculate the product of the spatial probability of the association between the target and the measurement and the prior probability of the association.

[0075] S402. Use the product of the spatial probability of the association between the target and the measurement and the prior probability of the association as the input of the belief propagation algorithm to solve for the initial confidence set of the association between the target and the measurement.

[0076] In one embodiment, based on the association confidence set between the target and the measurement and the spatial information of the measurement, the update message of the measurement is calculated by weighting the target's measurement using the association confidence set between the target and the measurement to calculate the equivalent measurement of the target, thereby obtaining the update message of the target's motion state, and calculating the update message of the target's existence state based on the statistical dependency of the target's existence state on the association confidence set between the target and the measurement.

[0077] In one embodiment, the motion state estimate and existence state estimate of the target at the current moment are calculated based on the measurement update message and the target motion state prediction and the target existence state prediction at the current moment. This is achieved by inputting the measurement update message and the target motion state prediction at the current moment into a Kalman filter to obtain the target motion state estimate at the current moment, and by multiplying the measurement update message and the target existence state prediction at the current moment and normalizing the result to obtain the target existence state estimate at the current moment.

[0078] In one specific embodiment, the target tracking method based on classification information assistance of the present invention includes the following steps:

[0079] Step 1: Based on the target motion state estimate and target existence state estimate of the previous moment, calculate the target motion state prediction and target existence state prediction of the current moment.

[0080] Step 1.1: Model the target's motion state and existence state at time k. At time k, there are N targets in the target tracking scenario.T There are n objectives, i∈{1,...,N} T} represents the index identifier of the target; let x i,k Let s represent the motion state of target i, which includes the target's position, velocity, and acceleration; i,k s represents the existence state of target i. i,k Let ∈{0,1} be a binary random variable, s i,k =1 indicates the target exists; otherwise, it indicates the target does not exist. Let and Let K represent the set of motion states and the set of existence states of all targets at time k, respectively.

[0081] Step 1.2: Estimation of the target motion state at time k-1, b(X) k-1 ) and existence state b(S) k-1 The estimation of ) is used for modeling. This method adopts a recursive processing framework. When executed at time k, such as Figure 5 As shown, Figure 5 This is a schematic diagram of the factor flow provided in an embodiment of the present invention. The algorithm input is the motion state estimate b(X) of all targets calculated at time k-1. k-1 ) and existence state estimate b(S) k-1 It should be noted that b(X0) and b(S0) can be obtained using a general track initialization algorithm, which will not be elaborated here. Let in, x represents i,k-1 Follow the mean The variance is P i,k-1 Gaussian distribution, b(s) i,k-1 ) represents s i,k-1 Follows a Bernoulli distribution;

[0082] Step 1.3: Estimate b(X) based on the target's motion state at time k-1. k-1 And the evolution information of the target motion state at time k-1, such as Figure 5 As shown, the motion state prediction of the target is calculated. in F is the target motion state transition equation, Q k Let T be the error covariance matrix of the noise during the target motion process, and let T be the matrix transpose.

[0083] Step S1.4: Estimate b(S) based on the existence state of the target at time k-1. k-1 And the evolutionary information of the target's state, such as Figure 5 As shown, the existence state prediction of the target is calculated. Wherein, b(s) i,k|k-1 )=Tb(s i,k-1Let T be the state transition equation for the objective, and [T] be the state transition equation for the objective. 1,1 =[T] 2,2 =0.85, [T] 1,2 =[T] 2,1 =0.15;

[0084] Step 2: Calculate the spatial probability of the association between the target and the measurement by combining the spatial information measured by the sensor at the current moment with the target motion state at the current moment, and calculate the prior probability of the association between the target and the measurement based on the target existence state at the current moment. Calculate the initial confidence set of the association between the target and the measurement based on the spatial probability and prior probability of the association.

[0085] Step 2.1: Model the spatial information and data association event set measured at time k. At time k, the sensor detects N... M,k Each measurement is used, with j∈{1,...,N} M,k} represents the index identifier of the measurement; let z j,k =[r j,k ,d j,k ] Τ This represents the spatial location information of measurement j, where r j,k and d j,k Representing radial distance measurement and Doppler frequency measurement respectively; let This represents the set of spatial location information measured at time k. (Definition) This represents the set of data association events between the measurement and the target; if measurement j is generated by target i, then a i,j,k =1, otherwise a i,j,k =0; if target i is missed, then a i,0,k =1, otherwise a i,0,k =0; if measurement j is clutter, then a 0,j,k =1, otherwise a 0,j,k =0.

[0086] Step 2.2: As Figure 5 As shown, prediction is based on the target's motion state. And the measured spatial information Z k Calculate the spatial probability of data association

[0087]

[0088] Where H is the measurement equation for the radar's spatial position, and R... k Let p be the covariance matrix of the measurement error. FA This represents the probability of a false alarm; such as Figure 5 As shown, based on the existence state prediction Calculate the prior probability of data association

[0089]

[0090] Where, p d Let be the radar detection probability.

[0091] Step 2.3: Iteratively calculate the following two messages

[0092]

[0093]

[0094] If the maximum number of iterations is reached, the iteration terminates. Then, the initial confidence set for data association is calculated. in, Initial confidence for data association.

[0095] Step 3: Use the radar range-Doppler spectrum feature extraction model to extract features from the radar range-Doppler spectrum information to obtain the feature vector of the radar range-Doppler spectrum information;

[0096] Step 3.1: Construct the RD spectrum information of the measurement. Measurement j in step S2.1 also possesses RD spectrum information. Where m∈{1,...,N m} represents the distance cell index, p∈{1,...,N p} represents the pulse index. The signal amplitude. The RD spectrum information of all measurements at time k is denoted as...

[0097] Step 3.2: Construct the neural network g M (·) is used for RD spectral feature extraction. g M The structure of (·) is as follows Figure 6 As shown, it mainly consists of three convolutional layers and three fully connected layers. Feature extraction is required for all measurements at time k. For measurement j, g... M The input of (·) is the RD spectrum information M. j,k The output is the corresponding feature vector h. j,k .like Figure 5 As shown, the RD spectrum characteristics of all measurements at time k are denoted as...

[0098] Step 4: After integrating the classification confidence obtained from the target classification model with the initial confidence of the association between the target and the measurement, calculate the association confidence set between the target and the measurement. This step requires iteratively running steps 4.1 to 4.3 three times. The processing flow for the t-th iteration is as follows:

[0099] Step 4.1: Construct the neural network g c(·) is used to classify measurements generated by a target or clutter. g c The structure of (·) is as follows Figure 7 As shown, it mainly consists of three fully connected layers. It is necessary to classify all measurements at time k. For measurement j, g... c The input of (·) is the RD spectrum feature vector h. j,k The confidence b associated with the data from iteration t-1 (t-1) (a 0,j,k The output is the classification confidence of the measurement generated by the target or clutter generation after t iterations. like Figure 5 As shown, the classification confidence scores for all measurements at time k are recorded as follows:

[0100] Step 4.2: Use DS rules to fuse the data association confidence sets from t-1 iterations. Classification confidence after t iterations First, define a set. in The measurement is generated by clutter, and h indicates that the measurement is generated by the target. This indicates that the measurement is generated by both clutter and the target. This indicates that the measurement may be generated by either clutter or the target; secondly, for measurement j, a basic probability assignment function m based on data association confidence is defined. j,k,1 ,in m j,k,1 (h)=b (0) (a 0,j,k =1), And define the basic assignment function m based on classification confidence. j,k,2 ,in, m j,k,2 (h)=ω j,k , Next, use the DS rule to analyze m. j,k,1 and m j,k,2 The fusion yields the fusion basic assignment function m. j,k =f DS (m j,k,1 ,m j,k,2 The calculation formula is:

[0101]

[0102] in, The normalized coefficients are used; finally, the fusion confidence function is calculated.

[0103]

[0104] Where |X| represents the number of elements in set X. For example... Figure 5 As shown, the set of fusion confidence functions at time k is denoted as

[0105] Step 4.3: Use the fusion confidence function Optimize the data association confidence set for iteration t-1. make And iteratively calculate the following two messages

[0106]

[0107]

[0108] If the results of two consecutive iterations are less than the threshold or the maximum number of iterations is reached, the iteration terminates. Then, the data association confidence set for the tth iteration is calculated. in, Confidence in the data association for t iterations.

[0109] Step 5: Based on the association confidence set between the target and the measurement and the spatial information of the measurement, calculate the update message of the measurement, and calculate the motion state estimate and existence state estimate of the target at the current moment based on the update message of the measurement and the target motion state prediction and target existence state prediction at the current moment.

[0110] Step 5.1: Based on the data association confidence set calculated in Step 4.3 And the measured spatial information Z k Calculate measurement update message in,

[0111]

[0112] as well as And based on the data association confidence set calculated in step S4.3 Calculate the probability update message of target existence

[0113] Step 5.2: Predict the target's motion state based on the calculation in Step 1.3 Measurement update message calculated in step 5.1 Update the target's motion state estimate at time k The calculation formula is as follows:

[0114]

[0115] The above equation can be solved using Kalman filtering, i.e. Predicting the existence state of the target based on the calculation in step 1.4 Measurement update message calculated in step S5.1 Update the existence state estimation of the target at time k The calculation formula is as follows:

[0116]

[0117] Once the target's motion state estimation and existence state estimation are completed, target tracking is finished.

[0118] In this embodiment, training and testing datasets are also constructed to train the neural network composed of the radar range-Doppler spectrum feature extraction model and the classification model. The specific training process is as follows:

[0119] (1) Construct training and testing datasets. In the training dataset, sea clutter backgrounds are randomly selected from real data, and simulated moving targets are introduced in each radar scan. First, a Fast Fourier Transform (FFT) is performed on the sea clutter echoes to obtain the range Doppler spectrum. Then, a constant false alarm rate (CFAR) detector is used to obtain candidate detectors, with a false alarm probability set to P. FA =0.28. To cluster closely spaced candidate detections into the same detection region, the DBSCAN clustering algorithm is used, where the range partitioning clustering threshold is R. th The Doppler partitioning clustering threshold is D. th Then, the spatial measurement value is calculated using the amplitude-weighted measurement center point estimation algorithm, and the calculation formula is as follows:

[0120]

[0121] Among them, A τ The amplitude is measured in the τth initial detection, and N is the number of initial detections in detection area j. To facilitate training the designed neural network, the RD (Radar Range-Doppler) spectrum is stretched to 0–255 dB while retaining the maximum and minimum values. In the training set, we set the SCR (Signal-to-Clutter Ratio) to -20–20 dB in 4 dB intervals, with 40 Monte Carlo tests per SCR. Therefore, we constructed a training set consisting of 722 target RD spectra and 2838 clutter RD spectra. For the test set, we used the same method as the training set.

[0122] (2) Complete target measurement and clutter labeling. If the distance between the measured value and any actual ground location is less than or equal to T. dist Then the ground real label The value is 1 if it is 1, otherwise it is 0.

[0123] (3) Design the classification loss function. We used the following binary cross-entropy loss to measure classification in both training steps.

[0124]

[0125] Where, ω j It is a classification confidence and It is the basic true label for each measurement.

[0126] It is an adjustment parameter. Adjusting the parameter ∈ addresses the imbalance problem in learning-based binary classification. This problem arises because object detectors often produce more false positives than true measurements, since missing an object is usually more important than generating a false positive. Adjusting the parameter ∈ helps handle this class imbalance problem during training;

[0127] (4) Complete network training. Assuming there is a training set consisting of an RD map and its corresponding labels, the proposed NN training process will employ a supervised method. Training is conducted in two steps. The first step mainly involves training... Figure 5 The feature extraction network shown is used as input. Figure 7 The second and third linear layers of the classification network. The second step mainly involves training the classification network by inputting the output of the feature extraction network and the belief of the FG part. Figure 6 The network in the diagram. During training, all parameters of the NN (neural network) are updated through backpropagation, which calculates the gradient of the loss function.

[0128] To verify the significant improvement in target tracking performance and robustness of this invention, in one embodiment, a simulation experiment was conducted using a real dataset from the CSIR (Center for Scientific and Industrial Research) collected during a sea clutter measurement experiment in Pretoria, South Africa. The sea clutter background contained M = 96 ranging zones and P = 10,000 pulses. The simulation lasted 150 seconds with a 10-second step size and included 15 radar scans.

[0129] Simulation target: The motion state of each target is represented by x. i,k The data includes the target's radial range, radial velocity, and radial acceleration. The target motion model is uniformly accelerated motion, with radial lengths sampled uniformly between 5 and 30 meters. The total radar echo is the simulated target echo with a specified SCR added to real sea clutter data. The RD spectrum is obtained by performing FFT on the echo data. Figure 8 It displays the actual target trajectory in the RD domain. Figure 9 Two sets of sea clutter echo maps in the time domain and RD domain are shown.

[0130] Training and Test Sets: The same method was used for both the training and test sets. Sea clutter backgrounds were randomly selected from 28 sets of real CSIR sea clutter data, and simulated targets were introduced in each radar scan. First, the sea clutter echoes were processed using FFT to obtain the RD spectrum. Then, a candidate detection list was generated using a constant false alarm rate (CFAR) detector, and the DBSCAN algorithm was used to cluster closely spaced candidate detections into the same detection group. Finally, the amplitude-weighted method was used to calculate the spatial information of the measurements. To facilitate neural network training, the RD spectrum was stretched to 0-255 dB. In the training set, the SCR was set to 20-20 dB with 4 dB intervals, generating 40 datasets for each SCR.

[0131] Algorithm parameters: For confidence propagation data association, the iterative convergence threshold is... The maximum number of iterations is N T Represents the maximum number of targets, if p(s) i,k =1) At least 3 out of 5 consecutive scans are greater than δ t =0.5, then the track is confirmed as a true track. If in three consecutive scans, p(s) i,k =1) less than δ t If the probability is 0.5, the flight path terminates. If the target is visible, the detection probability P is set. d (s i,k =1) =0.9; If the target is not visible, the detection probability P d (s i,k =0) =0.01. The target birth probability is set to p. b =0.15, the target survival probability is set to p s =0.85. The initial state of the target is set to f. s =0.5.

[0132] Performance evaluation: Average Multi-Target Tracking Accuracy (AMOT), Data Association Switching Count (IDS); Tracking Fragment, Root Mean Square Error (RMSE) of Target Motion State, and Average OSPA. Performance metrics are the average of 40 Monte Carlo tests. The proposed NEMP (Classification-Assisted Target Tracking) method is compared with two other methods: one using unclassified measurements, called MP, and the other using measurements after clutter suppression via an RD spectral classifier, called MP-NN.

[0133] Test results: Figure 10 The detection results of the constant false alarm rate detector are shown, among which, Figure 10Images (a) through 10(d) show target detection results at -8dB, -4dB, 0dB, and 4dB, respectively. These results indicate that the detector can measure targets, but it also generates a large amount of sea clutter, posing a significant challenge to target tracking. Furthermore, some targets are located within the main clutter spectrum. To improve data correlation performance, the RD spectrum within the frame is used as supplementary information about the measurement. Figure 11 As shown in (a), NEMP successfully tracked all four targets without generating any erroneous tracks. The tracking results of NEMP are as follows: Figure 11 As shown in (b), MP can track four targets, but it generates more erroneous tracks. Comparing NEMP and MP reveals that NEMP significantly suppresses erroneous tracks by utilizing RD spectral feature information. Table 1 presents a performance comparison of NEMP, MP, and MP-NN at SCR = 0 dB. The results show that, for AMOT, NEMP performs best, followed by MP-NN, and MP performs worst. This indicates that NEMP performs excellently in true track tracking, false track suppression, and data association. In terms of IDS and Frag, NEMP's results are similar to those of MP and MP-NN. Figure 12 The Monte Carlo average RMSE of target motion state estimation is shown. The RMSE-p index of NEMP is slightly smaller than that of MP and MP-NN, while the RMSE-v index of NEMP is almost the same as that of MP and MP-NN. Figure 13 The diagram shows the relationship between OSPA metrics and time. It can be seen that NEMP is superior to MP and MP-NN. Therefore, NEBP performs better in terms of positioning error and erroneous track suppression.

[0134] Table 1 Performance Comparison

[0135]

[0136] Figure 14 The performance comparison of different SCRs is presented. As can be seen from the figures, NEMP outperforms other algorithms for AMOT. This improvement is attributed to NEMP's incorporation of classification information into the target tracking architecture, effectively suppressing erroneous trajectories. For IDS and Frag, NEMP's results are similar to MP and MP-NN. For OSPA, NEMP outperforms MP-NN, while MP performs the worst. Comparing the RMSE of NEMP and other target tracking methods, their performance is comparable. Overall, NEMP outperforms other algorithms. This superior performance can be attributed to the use of RD spectral features, which enhances the robustness and effectiveness of target tracking. These results demonstrate that, compared to MP and MP-NN methods, the proposed classification-assisted target tracking method (NEMP) of this invention has advantages in erroneous trajectory suppression and overall target tracking performance.

[0137] Based on the same inventive concept, embodiments of the present invention also provide a classification-assisted target tracking system for implementing the classification-assisted target tracking method described above. The solution provided by this device is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more classification-assisted target tracking system embodiments provided below can be found in the limitations of the classification-assisted target tracking method described above, and will not be repeated here.

[0138] In one embodiment, a target tracking system based on classification information includes:

[0139] The state prediction calculation module is used to calculate the target motion state prediction and target existence state prediction at the current moment based on the target motion state estimation and target existence state estimation at the previous moment.

[0140] The associated prior probability calculation module is used to predict and calculate the associated spatial probability between the target and the measurement by combining the spatial information measured by the sensor at the current moment with the target motion state at the current moment, and to predict and calculate the associated prior probability between the target and the measurement based on the target existence state at the current moment.

[0141] The module for calculating the initial confidence set of association is used to calculate the initial confidence set of association between the target and the measurement based on the spatial probability and prior probability of association between the target and the measurement.

[0142] The feature extraction module is used to extract features from radar range-Doppler spectrum information using a radar range-Doppler spectrum feature extraction model, and obtain the feature vector of radar range-Doppler spectrum information.

[0143] The fusion module is used to fuse the classification confidence obtained by the target classification model with the initial confidence of the association between the target and the measurement, and then calculate the association confidence set between the target and the measurement.

[0144] The update message calculation module is used to calculate the update message of the measurement based on the association confidence set between the target and the measurement and the spatial information of the measurement.

[0145] The state estimation calculation module is used to calculate the target's motion state estimate and existence state estimate at the current moment based on the measurement update message and the target motion state prediction and target existence state prediction at the current moment.

[0146] The target tracking system based on classification information of the present invention has high target tracking performance and robustness.

[0147] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A target tracking method based on classification information assistance, characterized in that, The method comprises the following steps: According to the target motion state estimation and the target existence state estimation of the last moment, the target motion state prediction and the target existence state prediction of the current moment are calculated; The spatial information of the sensor measurement of the current moment is used to calculate the associated spatial probability of the target and the measurement in combination with the target motion state prediction of the current moment, and the associated prior probability of the target and the measurement is calculated according to the target existence state prediction of the current moment; The associated initial confidence set of the target and the measurement is calculated according to the associated spatial probability and the associated prior probability of the target and the measurement; The radar range-Doppler spectrum information is extracted by using a radar range-Doppler spectrum feature extraction model to obtain a feature vector of the radar range-Doppler spectrum information; The associated confidence set of the target and the measurement is calculated after fusing the classification confidence obtained by using the target classification model and the associated initial confidence of the target and the measurement; The update message of the measurement is calculated according to the associated confidence set of the target and the measurement and the spatial information of the measurement; The target motion state estimation and the existence state estimation of the current moment are calculated according to the update message of the measurement and the target motion state prediction and the target existence state prediction of the current moment. The associated confidence set of the target and the measurement is calculated after fusing the classification confidence obtained by using the target classification model and the associated initial confidence of the target and the measurement, which comprises the following steps: The feature vector of the radar range-Doppler spectrum information and the associated initial confidence set of the target and the measurement are input into the target classification model to obtain the classification confidence; The classification confidence and the associated initial confidence set of the target and the measurement are fused by using the DS fusion rule; The iterative associated confidence set is solved according to the fusion result of the classification confidence and the associated initial confidence set of the target and the measurement; It is judged whether the number of times of fusing the classification confidence and the associated initial confidence set of the target and the measurement reaches the maximum iteration number, if not, the iterative associated confidence set is taken as the associated initial confidence set of the target and the measurement, and the step of inputting the feature vector of the radar range-Doppler spectrum information and the associated initial confidence set of the target and the measurement into the target classification model to obtain the classification confidence is returned, if yes, the iterative associated confidence set is taken as the associated confidence set of the target and the measurement.

2. The target tracking method based on classification information assistance according to claim 1, characterized in that, The associated confidence set of the target and the measurement is calculated according to the fusion result of the classification confidence and the associated initial confidence set of the target and the measurement, which comprises: The confidence set of the associated event of the measurement as clutter in the associated initial confidence set of the target and the measurement is replaced by the fusion result of the classification confidence and the associated initial confidence set of the target and the measurement to obtain a replaced associated initial confidence set; The replaced associated initial confidence set is input into the confidence propagation algorithm to obtain the iterative associated confidence set.

3. The target tracking method based on classification information assistance according to claim 2, characterized in that, The target motion state prediction and the target existence state prediction of the current moment are calculated according to the target motion state estimation and the target existence state estimation of the last moment, which comprises: The target motion state prediction of the current moment is calculated according to the target motion state estimation of the last moment and a target motion model, and the target existence state prediction of the current moment is calculated according to the target existence state estimation of the last moment and a target existence state transition equation.

4. The target tracking method based on classification information assistance according to claim 3, characterized in that, The method comprises the following steps: calculating the product of the associated spatial probability and the associated prior probability of the target and the measurement; solving the initial associated confidence set of the target and the measurement by taking the product of the associated spatial probability and the associated prior probability of the target and the measurement as the input of the belief propagation algorithm.

5. The target tracking method based on classification information assistance according to claim 4, characterized in that, According to the associated confidence set of the target and the measurement and the spatial information of the measurement, the update message of the measurement is calculated by weighting the measurement of the target with the associated confidence set of the target and the measurement to calculate the equivalent measurement of the target, and obtaining the update message of the target motion state, and calculating the update message of the target existence state according to the statistical dependence of the target existence state on the associated confidence set of the target and the measurement.

6. The target tracking method based on classification information assistance according to claim 5, characterized in that, According to the update message of the measurement and the target motion state prediction and the target existence state prediction at the current time, the motion state estimation and the existence state estimation of the target at the current time are calculated by inputting the update message of the measurement and the target motion state prediction at the current time into the Kalman filter to obtain the motion state estimation of the target at the current time, and multiplying the update message of the measurement and the target existence state prediction at the current time and normalizing to obtain the target existence state estimation at the current time.

7. The target tracking method based on classification information assistance according to claim 6, characterized in that, The radar range-Doppler spectrum feature extraction model comprises three convolutional layers, one flattening layer and three fully connected layers in sequence from input to output, and the target classification model comprises three fully connected layers.

8. A classification information assistance-based target tracking system for implementing the classification information assistance-based target tracking method according to any one of claims 1 to 7, characterized by The method comprises the following steps: a state prediction calculation module for calculating the target motion state prediction and the target existence state prediction at the current time according to the target motion state estimation and the target existence state estimation at the previous time; an associated prior probability calculation module for calculating the associated spatial probability of the target and the measurement by using the spatial information of the sensor measurement at the current time in combination with the target motion state prediction at the current time, and calculating the associated prior probability of the target and the measurement according to the target existence state prediction at the current time; an initial associated confidence set calculation module for calculating the initial associated confidence set of the target and the measurement according to the associated spatial probability and the associated prior probability of the target and the measurement; a feature extraction module for performing feature extraction on the radar range-Doppler spectrum information by using the radar range-Doppler spectrum feature extraction model to obtain the feature vector of the radar range-Doppler spectrum information; a fusion module for calculating the associated confidence set of the target and the measurement after fusing the classification confidence obtained by using the target classification model and the initial associated confidence of the target and the measurement; an update message calculation module for calculating the update message of the measurement according to the associated confidence set of the target and the measurement and the spatial information of the measurement; a state estimation calculation module for calculating the motion state estimation and the existence state estimation of the target at the current time according to the update message of the measurement and the target motion state prediction and the target existence state prediction at the current time.

Images