CN116449360A - Maneuvering target tracking method based on long-short-time memory network - Google Patents
Maneuvering target tracking method based on long-short-time memory network Download PDFInfo
- Publication number
- CN116449360A CN116449360A CN202210930780.4A CN202210930780A CN116449360A CN 116449360 A CN116449360 A CN 116449360A CN 202210930780 A CN202210930780 A CN 202210930780A CN 116449360 A CN116449360 A CN 116449360A
- Authority
- CN
- China
- Prior art keywords
- target
- value
- position information
- network
- maneuvering
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 49
- 230000015654 memory Effects 0.000 title claims abstract description 27
- 239000011159 matrix material Substances 0.000 claims abstract description 73
- 238000005259 measurement Methods 0.000 claims abstract description 61
- 238000001914 filtration Methods 0.000 claims abstract description 42
- 230000008569 process Effects 0.000 claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 22
- 230000006870 function Effects 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000012804 iterative process Methods 0.000 claims description 3
- 238000002360 preparation method Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims 1
- 238000004804 winding Methods 0.000 abstract description 2
- 238000001514 detection method Methods 0.000 description 4
- 230000006399 behavior Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000009194 climbing Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 230000009189 diving Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000008707 rearrangement Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000010396 two-hybrid screening Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/66—Radar-tracking systems; Analogous systems
- G01S13/72—Radar-tracking systems; Analogous systems for two-dimensional tracking, e.g. combination of angle and range tracking, track-while-scan radar
- G01S13/723—Radar-tracking systems; Analogous systems for two-dimensional tracking, e.g. combination of angle and range tracking, track-while-scan radar by using numerical data
- G01S13/726—Multiple target tracking
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a maneuvering target tracking method based on a long-short time memory network, which comprises the steps of firstly training a long-short time memory network LSTM model by radar measurement data to obtain a one-step predicted value, a process noise variance matrix and a measurement noise variance matrix of a target state, then calculating a Kalman filtering gain by Kalman filtering, so as to calculate a filtered value of target position information, and the filtered value of the target position information is calculated without presetting target motion to the model and the noise variance information, thereby improving tracking precision, and simultaneously correcting the filtered gain by a fourth LSTM network in consideration of sudden maneuvering of the target, and improving adaptability of an algorithm to maneuvering. The target track comprises a constant speed, a round winding and a constant speed and a turning back mixed motion track, the characteristic of the maneuvering motion of the target can be more fully reflected, and the filtering value of the target position information finally obtained by model training is more similar to the real target position information.
Description
Technical Field
The invention relates to the technical field of radar tracking, in particular to a maneuvering target tracking method based on a long-and-short-time memory network.
Background
Kalman filtering algorithms are often used for target tracking. However, in the tracking process, the target may maneuver, so that the pre-established mathematical model is not matched with the current motion mode of the target, and thus the tracking precision is reduced, and even the tracking losing phenomenon occurs.
The traditional Kalman filtering algorithm needs to specify a target motion model, a measurement model, noise parameters and the like in advance, and common models comprise a constant speed model, a constant acceleration model, a Singer model, a current statistical model and the like. The maneuvering forms of the aerial targets are various, for example, the bird targets can perform maneuvering behaviors such as cruising, climbing, diving, hovering and the like, maneuvering initiative is high, the time for predicting maneuvering behavior switching cannot be realized, when the targets are tracked, the traditional Kalman filtering algorithm cannot accurately model various complex maneuvering on one hand, and on the other hand, the method is difficult to adapt to various sudden maneuvering, larger peak errors often occur at maneuvering occurrence time, and even track interruption is caused by losing the targets.
Aiming at the problems of reduced tracking precision, tracking divergence and the like caused by target maneuver, the current maneuver tracking algorithm is mainly developed around the uncertainty of the occurrence time of the target maneuver and can be divided into a single model algorithm and a multi-model algorithm based on decision, wherein the multi-model algorithm adopts a plurality of parallel models to describe the motion model of the target, so that the difficult problem of complex maneuver modeling is solved to a certain extent, the algorithm performance is seriously dependent on a selected model set, the design process is complex, and the calculated amount is large; the single model algorithm based on decision combines maneuver detection and identification technology with a single model, adjusts model parameters and filtering parameters according to maneuver detection results, and researches show that if a good maneuver detection method is adopted, the single model algorithm can obtain performance similar to that of a multimode algorithm. However, in the model-based maneuver tracking method, the state transition model needs to be carefully designed and adjusted, so that the difficult problem of modeling the complex maneuver behavior is not solved, and the problems of maneuver detection time delay and the like exist.
Therefore, for radar maneuvering target tracking, a maneuvering target tracking algorithm is needed to solve the problems of difficult modeling of complex maneuvering targets and low tracking accuracy.
Disclosure of Invention
In view of the above, the invention provides a maneuvering target tracking method based on a long-short-time memory network, which can effectively track maneuvering targets, improve tracking accuracy and improve adaptability of an algorithm to target maneuvering.
The invention adopts the following specific technical scheme:
a maneuvering target tracking method based on a long-short time memory network comprises the following steps:
step one, acquiring radar measurement data, namely an observation value of target position information, according to a target track;
training to obtain a maneuvering target tracking network model:
step 201, training an LSTM model by using the radar measurement data to obtain a one-step predicted value of a target state, a process noise variance matrix of the target state and a measurement noise variance matrix of the target state;
step 202, taking the one-step predicted value, the process noise variance matrix and the measurement noise variance matrix as inputs of Kalman filtering to obtain a gain matrix of the Kalman filtering and a normalized distance of an innovation vector;
step 203, performing a stitching process on the Kalman filtered gain matrix and the normalized distance of the innovation vector to obtain a stitching vector, and inputting the stitching vector into a fourth LSTM network to obtain a corrected gain matrix; calculating a filtering value of the target position information by using the corrected gain matrix;
and thirdly, inputting real-time radar measurement data into the maneuvering target tracking network model to obtain a filtering value of the target position information.
Further, in the first step, the target track includes two mixed motion tracks of constant speed, round-robin and constant speed and turn-back;
the radar measurement data acquisition according to the target track comprises the following steps: and adding an oblique distance error and an angle measurement error to the target track to obtain the radar measurement data.
Further, in the step 201, training the long-short-time memory network LSTM model by using the radar measurement data is:
inputting the estimated value of the radar measurement data at the k-1 moment into a first LSTM network to obtain a one-step predicted value of a target state, inputting the estimated value of the radar measurement data at the k-1 moment into a second LSTM network to obtain a process noise variance matrix of the target state, inputting the radar measurement data into a third LSTM network to obtain a measurement noise variance matrix of the target state, wherein k represents the moment and is a positive integer;
for the initial time, the inputs of the first LSTM network and the second LSTM network are the radar measurement data of the initial time; in the iterative process after the initial time, for the kth time, the inputs of the first LSTM network and the second LSTM network are the filter value of the target position information at the kth-1 time, namely the estimated value of the radar measurement data.
Further, in the step 202,
the gain matrix is: k'. k ＝P k|k-1 ·inv(S k )
Wherein K' K represents a gain matrix, pk|k-1 represents a state prediction covariance matrix of the K-1 moment to the K moment prediction, sk represents a innovation covariance matrix obtained by Kalman filtering, and inv (&) represents matrix inversion;
the normalized distance of the innovation vector is as follows:
wherein D is k Representing normalized distance of innovation vector e k Representing the innovation vector obtained by the kalman filter,representing the transpose of the innovation vector.
Further, in the step 203, the filtering value of the target position information obtained by calculating using the corrected gain matrix is:
wherein,,filtered value representing the target position information at the kth moment, is->A filter value representing the target position information predicted at the kth time at the kth-1 time, K k Representing a modified gain matrix, e k Representing the innovation vector obtained by the kalman filter.
Further, in the second step, a loss function is adopted to optimize the maneuvering target tracking network model training process:
where N is the time step, y k For the k-th moment data true value, namely the position information in the target track,for the network predicted output value at the kth moment, i.e. the filtered value of the target position information at the kth moment,/->A filter value representing target position information predicted at the kth time at the kth-1 time, W being a trainable weight of the neural network, lambda 1 Is LSTM f Coefficients of the deviation term of the output predicted value, i.e. the filtered value of the target position information, from the true value, lambda 2 For regularization coefficients, k represents the time of day, is a positive integer which is used for the preparation of the high-voltage power supply, I 2 Representing a two-norm operation.
Further, the first LSTM network, the second LSTM network, the third LSTM network, and the fourth LSTM network each include only two layers of gate structures: an input gate and an output gate.
The beneficial effects are that:
(1) A maneuvering target tracking method based on a long-short time memory network firstly utilizes radar measurement data to train a long-short time memory network LSTM model to obtain a one-step predicted value, a process noise variance matrix and a measurement noise variance matrix of a target state, and then calculates Kalman filtering gain through Kalman filtering, so that a filtering value of target position information is calculated, a target motion is not required to be preset for the model and the noise variance information, tracking precision is improved, meanwhile, a fourth LSTM network is utilized to correct the filtering gain in consideration of sudden maneuvering of the target, and adaptability of an algorithm to maneuvering is improved.
(2) The target track comprises a constant speed, a round winding and a constant speed and a turning back mixed motion track, so that the characteristic of the maneuvering motion of the target can be more fully reflected, and the filtering value of the target position information finally obtained by model training is more approximate to the real target position information.
(3) The LSTM network only comprises a two-layer door structure of an input door and an output door, information inflow can be weakened and strengthened through complementary forms, when long-time memory information is weakened, current input information is strengthened, three doors are combined into two doors, and training speed is faster.
Drawings
FIG. 1 is a flow chart of a maneuvering target tracking method based on a long-short time memory network according to the invention.
FIG. 2 is a network structure diagram of a maneuvering target tracking method based on a long-short time memory network according to the invention.
Fig. 3 is a diagram of a modified LSTM network structure.
FIG. 4 is a graph showing the relation of root mean square error results of round maneuver positions.
FIG. 5 is a graph showing the root mean square error results for the retracing maneuver position.
FIG. 6 is a graph showing the root mean square error results for maneuver positions.
Detailed Description
A maneuvering target tracking method based on a long-short time memory network firstly utilizes radar measurement data to train a long-short time memory network LSTM model to obtain a one-step predicted value, a process noise variance matrix and a measurement noise variance matrix of a target state, and then calculates Kalman filtering gain through Kalman filtering, so that a filtering value of target position information is calculated, a target motion is not required to be preset for the model and the noise variance information, tracking precision is improved, meanwhile, a fourth LSTM network is utilized to correct the filtering gain in consideration of sudden maneuvering of the target, and adaptability of an algorithm to maneuvering is improved.
The invention will now be described in detail by way of example with reference to the accompanying drawings.
As shown in fig. 1, the maneuvering target tracking method based on the long-short time memory network of the invention specifically comprises the following steps:
step one, obtaining radar measurement data, namely an observed value of target position information, according to a target track.
Generating a plurality of maneuvering target tracks, obtaining measurement data, preprocessing the measurement data, and dividing a training set and a testing set according to a proportion;
the generation of the target track is as follows: setting track parameters, constructing two mixed motion tracks of uniform speed, round, uniform speed and turn back, wherein track information consists of position coordinates in a two-dimensional plane, and the position coordinates are used as true values of a data set;
obtaining radar measurement data: the radar measurement consists of an inclined distance and an angle, the motion trail under the rectangular coordinate system is converted into a spherical coordinate system of the radar, so as to simulate the real situation as far as possible, and a normal distribution inclined distance error with the standard deviation of 50m and a normal distribution angle measurement error with the standard deviation of 0.02 DEG are added, and the filtering tracking process is usually carried out under the rectangular coordinate system, so that the radar measurement with the added error is projected under the rectangular coordinate system to obtain position coordinate information, and the position coordinate information is used as an input value of a data set; the radar measurement data is an observation value of the target position information.
Data rearrangement and data set partitioning: to ensure uniform sample distribution, the two hybrid machines are randomly rearranged, and the data set is divided into a training set and a testing set according to a certain proportion (7:3);
data normalization: each dimension of the data has different distribution ranges, the neural network is very sensitive to the difference, the training is influenced, the input data and the true value of the data set are required to be normalized, dimensionless data are obtained, and the normalization is specifically as follows: both the truth data and the metrology data were normalized.
Wherein z' in Representing the input data to be normalized,sum sigma in Respectively represent the mean value and standard deviation, z of input data in Representing normalized input data, the normalized input value being the network input data, y' representing the truth value data to be normalized,>sum sigma y′ And respectively representing the mean value and standard deviation of the truth value data, wherein y represents the normalized truth value data, and the normalized truth value data is used for evaluating the network performance.
Training to obtain a maneuvering target tracking network model.
FIG. 2 is a network structure diagram of a maneuvering target tracking method based on a long-short-time memory network according to the invention.
Step 201, training an LSTM model of a long-short-time memory network by using radar measurement data to obtain a one-step predicted value of a target state, a process noise variance matrix of the target state and a measurement noise variance matrix of the target state.
The long-time memory network LSTM model training is carried out by utilizing radar measurement data, and the training is as follows:
inputting the estimated value of the radar measurement data into a first LSTM network to obtain a one-step predicted value of the target state, inputting the estimated value of the radar measurement data into a second LSTM network to obtain a process noise variance matrix of the target state, and inputting the radar measurement data into a third LSTM network to obtain a measurement noise variance matrix of the target state.
I.e. the first long and short time memory network LSTM f Output value of network at time k-1As input, output is taken as one-step predictive value x of k moment state k|k-1 . Second long-term memory network LSTM Q Output value of network at time k-1As input, take output as k moment process noise variance matrix Q k . Third long and short time memory network LSTM R Network input value z at time k k Taking the output as the k moment measurement noise variance matrix R as the input k . Fourth long short time memory network LSTM K The input of (a) is a gain matrix K' k Normalized innovation D k Is output as a corrected gain matrix K k 。
The input of the first LSTM network and the second LSTM network at the initial moment is radar measurement data at the initial moment; in the iterative process after the initial time, for the kth time, the inputs of the first LSTM network and the second LSTM network are the filter value of the target position information at the kth-1 time, namely the estimated value of radar measurement data, and k represents the time and is a positive integer.
The radar measurement has low-dimensional characteristics, a fully-connected network is designed as an encoder, the input data is subjected to weight change, the high-order characteristics of the measurement data can be extracted from the input data, and the characteristics of the input data at each moment are extracted by z k ＝FC(z in ) FC (·) represents a fully connected network.
Step 202, taking a one-step predicted value, a process noise variance matrix and a measurement noise variance matrix as inputs of Kalman filtering to obtain a gain matrix of the Kalman filtering and a normalized distance of an innovation vector;
the gain matrix is: k'. k ＝P k|k-1 ·inv(S k )
Wherein K' K represents a gain matrix, pk|k-1 represents a state prediction covariance matrix of the K-1 moment to the K moment prediction, sk represents a innovation covariance matrix obtained by Kalman filtering, and inv (&) represents matrix inversion;
the normalized distance of the innovation vector is:
wherein D is k Representing normalized distance of innovation vector e k Representing the innovation vector obtained by the kalman filter,representing the transpose of the innovation vector.
Performing Kalman filtering, and calculating a state prediction covariance matrix P by using the obtained state prediction value and noise variance information k|k-1 Information vector e k New information covariance matrix S k Gain matrix K' k The method comprises the following steps:
the state covariance matrix of the previous moment of Pk-1|k-1, inv (·) represents matrix inversion, the innovation vector ek can be used for representing target maneuver, if the target does not maneuver, the innovation is zero-mean Gaussian white noise, the distance function Dk is subjected to X2 distribution with the degree of freedom being the measurement dimension nz, and when the target maneuver occurs, the innovation characteristic and the characteristic of the distance function thereof change, so that the Dk can be used for correcting the filtering gain.
Step 203, performing splicing processing on the normalized distance between the Kalman filtered gain matrix and the innovation vector to obtain a spliced vector, and inputting the spliced vector into a fourth LSTM network to obtain a corrected gain matrix; calculating and obtaining a filtering value of the target position information by using the corrected gain matrix;
the filtered value of the target position information is obtained by calculating the corrected gain matrix:
wherein,,filtered value representing the target position information at the kth moment, is->A filter value representing the target position information predicted at the kth time at the kth-1 time, K k Representing a modified gain matrix, e k Representing the innovation vector obtained by the kalman filter.
Taking into account the targeted burst maneuver, utilizing LSTM K And correcting the gain obtained by the Kalman filtering, and adjusting the gain calculated by the Kalman filtering by taking the spliced vector of the gain and the innovation distance function as an input.
Obtaining corrected gain matrix, calculating filtering state and covariance matrix thereof, including
Where I represents an identity matrix of the same dimension as the gain matrix.
It should be noted that the LSTM modules involved each comprise only two layers of gate structures (input gate and output gate), and only the input gate i is reserved t In 1-i t As a forgetting door f t The information inflow is weakened and strengthened through the complementary form, when the memory information is weakened for a long time, the current input information is strengthened, and the three gates are combined into two gates, so that the training speed is faster. A single network module, i.e. a modified LSTM network structure, is shown in fig. 3.
In the model training process, a loss function is adopted to optimize the maneuvering target tracking network model training process:
where N is the time step, y k For the k-th moment data true value, namely the position information in the target track,for the network predicted output value at the kth moment, i.e. the filtered value of the target position information at the kth moment,/->A filter value representing target position information predicted at the kth time at the kth-1 time, W being a trainable weight of the neural network, lambda 1 Is LSTM f Coefficients of the deviation term of the output predicted value, i.e. the filtered value of the target position information, from the true value, lambda 2 For regularization coefficients, k represents the time of day, is a positive integer which is used for the preparation of the high-voltage power supply, I 2 Representing a two-norm operation.
Setting network parameters including learning rate, iteration times, hidden layer node number, time step and the like, selecting a loss function of the network model, wherein the loss function can be used for evaluating the performance of the model and providing a parameter optimization direction.
From the above equation, the loss function consists of three parts: the first part is the mean square error between the network output value and the true value, representing the error degree of the final estimated value and the true value, and the second part is LSTM f And the third part is a regularization term used for controlling the complexity of the model and preventing the overfitting.
The invention optimizes parameters by utilizing an Adam optimization algorithm, sets the attenuation rate, and further optimizes the learning rate by combining the attenuation of a fixed step mode.
The decay rate was set to 0.9, and the learning rate was multiplied by the decay rate every 3 steps to obtain a new learning rate.
And thirdly, inputting real-time radar measurement data into a maneuvering target tracking network model to obtain a filtering value of the target position information.
To verify the validity of the maneuvering target tracking method described above. Based on simulation experimental data, the invention adopts the maneuvering target tracking algorithm combining the long-short-time memory network and the Kalman filtering to track turn-back and round maneuvering scenes respectively; and then, under the same maneuvering scene, tracking the target by utilizing a traditional Kalman filtering algorithm based on a model, and taking the position root mean square error as an evaluation index, wherein RMSE curves of the two methods are shown in fig. 4 to 6, wherein the RMSE curves of the two methods are round maneuvering position root mean square error, the RMS error of the round maneuvering position is shown in fig. 4, the RMS error of the turning maneuvering position is shown in fig. 5, and the RMS error of the maneuvering position is shown in fig. 6. The results of the statistical averaging of the individual segments RMSE are shown in table 1.
Table 1 comparison of three methods under multiple motion scenarios
Therefore, the method of the invention improves the tracking precision of the maneuvering target, enhances the adaptability of the algorithm to maneuvering and verifies the effectiveness.
In summary, the present invention firstly uses three long-short time memory network modules to model the state transfer function, the process noise variance matrix and the measurement noise variance matrix; then carrying out Kalman filtering by utilizing relevant components obtained from a large amount of data, calculating Kalman gain, and correcting the Kalman gain by utilizing an LSTM network in consideration of the possibility of sudden maneuver of a target; and finally, completing the estimation of the state and the error variance matrix by using the corrected gain.
The above specific embodiments merely describe the design principle of the present invention, and the shapes of the components in the description may be different, and the names are not limited. Therefore, the technical scheme described in the foregoing embodiments can be modified or replaced equivalently by those skilled in the art; such modifications and substitutions do not depart from the spirit and technical scope of the invention, and all of them should be considered to fall within the scope of the invention.
Claims (7)
1. A maneuvering target tracking method based on a long-short-time memory network, which is characterized by comprising the following steps:
step one, acquiring radar measurement data, namely an observation value of target position information, according to a target track;
training to obtain a maneuvering target tracking network model:
step 201, training an LSTM model by using the radar measurement data to obtain a one-step predicted value of a target state, a process noise variance matrix of the target state and a measurement noise variance matrix of the target state;
step 202, taking the one-step predicted value, the process noise variance matrix and the measurement noise variance matrix as inputs of Kalman filtering to obtain a gain matrix of the Kalman filtering and a normalized distance of an innovation vector;
step 203, performing a stitching process on the Kalman filtered gain matrix and the normalized distance of the innovation vector to obtain a stitching vector, and inputting the stitching vector into a fourth LSTM network to obtain a corrected gain matrix; calculating a filtering value of the target position information by using the corrected gain matrix;
and thirdly, inputting real-time radar measurement data into the maneuvering target tracking network model to obtain a filtering value of the target position information.
2. The method of claim 1, wherein in the first step, the target track includes a mixed motion track of constant velocity + circle around and constant velocity + fold back;
the radar measurement data acquisition according to the target track comprises the following steps: and adding an oblique distance error and an angle measurement error to the target track to obtain the radar measurement data.
3. The maneuvering target tracking method according to claim 1, wherein in step 201, training the long-time memory network LSTM model using the radar measurement data is as follows:
inputting the estimated value of the radar measurement data at the k-1 moment into a first LSTM network to obtain a one-step predicted value of a target state, inputting the estimated value of the radar measurement data at the k-1 moment into a second LSTM network to obtain a process noise variance matrix of the target state, inputting the radar measurement data into a third LSTM network to obtain a measurement noise variance matrix of the target state, wherein k represents the moment and is a positive integer;
for the initial time, the inputs of the first LSTM network and the second LSTM network are the radar measurement data of the initial time; in the iterative process after the initial time, for the kth time, the inputs of the first LSTM network and the second LSTM network are the filter value of the target position information at the kth-1 time, namely the estimated value of the radar measurement data.
4. The method of maneuvering target tracking as claimed in claim 1, wherein in step 202,
the gain matrix is: k'. k ＝P k|k-1 ·inv(S k )
Wherein, K' k Representing a gain matrix, P k|k-1 A state prediction covariance matrix representing a prediction of a kth moment from a kth moment-1, S k Representing a innovation covariance matrix obtained by Kalman filtering, wherein inv (·) represents matrix inversion;
the normalized distance of the innovation vector is as follows:
wherein D is k Representing normalized distance of innovation vector e k Representing the innovation vector obtained by the kalman filter,representing the transpose of the innovation vector.
5. The maneuvering target tracking method according to claim 1, wherein in step 203, the filtering value of the target position information obtained by calculation using the corrected gain matrix is:
wherein,,filtered value representing the target position information at the kth moment, is->A filter value representing the target position information predicted at the kth time at the kth-1 time, K k Representing a modified gain matrix, e k Representing the innovation vector obtained by the kalman filter.
6. The maneuvering target tracking method according to claim 1, wherein in the second step, a loss function is used to optimize the training process of the maneuvering target tracking network model:
where N is the time step, y k For the k-th moment data true value, namely the position information in the target track,for the network predicted output value at the kth moment, i.e. the filtered value of the target position information at the kth moment,/->A filter value representing target position information predicted at the kth time at the kth-1 time, W being a trainable weight of the neural network, lambda 1 Is LSTM f Coefficients of the deviation term of the output predicted value, i.e. the filtered value of the target position information, from the true value, lambda 2 For regularization coefficients, k represents the time of day, is a positive integer which is used for the preparation of the high-voltage power supply, I 2 Representing a two-norm operation.
7. The maneuvering target tracking method of claim 3 wherein the first LSTM network, the second LSTM network, the third LSTM network, and the fourth LSTM network each include only two layers of gate structures: an input gate and an output gate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210930780.4A CN116449360A (en) | 2022-08-04 | 2022-08-04 | Maneuvering target tracking method based on long-short-time memory network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210930780.4A CN116449360A (en) | 2022-08-04 | 2022-08-04 | Maneuvering target tracking method based on long-short-time memory network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116449360A true CN116449360A (en) | 2023-07-18 |
Family
ID=87129017
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210930780.4A Pending CN116449360A (en) | 2022-08-04 | 2022-08-04 | Maneuvering target tracking method based on long-short-time memory network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116449360A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117788511A (en) * | 2023-12-26 | 2024-03-29 | 兰州理工大学 | Multi-expansion target tracking method based on deep neural network |
-
2022
- 2022-08-04 CN CN202210930780.4A patent/CN116449360A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117788511A (en) * | 2023-12-26 | 2024-03-29 | 兰州理工大学 | Multi-expansion target tracking method based on deep neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111351488B (en) | Intelligent trajectory reconstruction reentry guidance method for aircraft | |
| CN113110592A (en) | Unmanned aerial vehicle obstacle avoidance and path planning method | |
| CN103383261B (en) | A kind of modified can't harm the indoor moving targets location method of Kalman filtering | |
| CN111985093A (en) | Adaptive unscented Kalman filtering state estimation method with noise estimator | |
| CN111460636B (en) | Hybrid interactive strong tracking filtering method for maneuvering extended target under drive of incomplete measurement data | |
| CN111798494B (en) | Maneuvering target robust tracking method under generalized correlation entropy criterion | |
| CN114842681A (en) | Airport scene flight path prediction method based on multi-head attention mechanism | |
| CN114819068A (en) | Hybrid target track prediction method and system | |
| CN116449360A (en) | Maneuvering target tracking method based on long-short-time memory network | |
| CN105424043A (en) | Motion state estimation method based on maneuver judgment | |
| Zhao et al. | Adaptive non-linear joint probabilistic data association for vehicle target tracking | |
| Riviere et al. | Neural tree expansion for multi-robot planning in non-cooperative environments | |
| Meng et al. | A Fully-Autonomous Framework of Unmanned Surface Vehicles in Maritime Environments Using Gaussian Process Motion Planning | |
| Lin et al. | A convolutional neural network particle filter for UUV target state estimation | |
| CN116047495B (en) | State transformation fusion filtering tracking method for three-coordinate radar | |
| CN116432539A (en) | Time consistency collaborative guidance method, system, equipment and medium | |
| Jiang et al. | Trainable kalman filter based on recurrent neural network and its application in aviation surveillance | |
| CN111948601B (en) | Single-station pure-angle target positioning and tracking method under non-Gaussian noise condition | |
| CN111416595B (en) | Big data filtering method based on multi-core fusion | |
| Liu et al. | A deep neural network based maneuvering-target tracking algorithm | |
| CN110649911B (en) | Distributed nonlinear Kalman filtering method based on alpha divergence | |
| Xu et al. | Xgb Assisted Self-Learning Kalman Filter for Uwb Localization | |
| Farooq et al. | Hybrid Deep Neural Network for Data-Driven Missile Guidance with Maneuvering Target. | |
| CN115128597B (en) | Maneuvering target tracking method under non-Gaussian noise based on IMM-STEKF | |
| Zhang et al. | Multiple Sensor Track Fusion Algorithm Based on LSTM Network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |