CN116106890A - Radar target track starting method based on quantum particle swarm and LGBM - Google Patents
Radar target track starting method based on quantum particle swarm and LGBM Download PDFInfo
- Publication number
- CN116106890A CN116106890A CN202310084045.0A CN202310084045A CN116106890A CN 116106890 A CN116106890 A CN 116106890A CN 202310084045 A CN202310084045 A CN 202310084045A CN 116106890 A CN116106890 A CN 116106890A
- Authority
- CN
- China
- Prior art keywords
- track
- model
- real
- initial
- optimal
- 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
Images
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Networks & Wireless Communication (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention provides a radar target track starting method based on a quantum particle swarm and an LGBM, which comprises the following steps: obtaining and marking original data; the feature extraction of the data, the initial feature vector is established, and the further extraction of the target features is carried out from the space-time dimension respectively; establishing and training a LightGbm model, and selecting parameters of the model by using grid searching and cross verification; optimizing a probability threshold, establishing a function of an objective function on probability threshold parameters, and carrying out parameter optimization by using a QPSO algorithm to obtain a probability threshold when the objective function converges to a maximum value; deployment application of the LightGbm model. The method is based on extracting multidimensional features of an initial track of a target, establishing a LightGbm model for false discrimination of the target track, and optimizing probability threshold parameters by using a QPSO algorithm to maximize the initial efficiency of the track.
Description
Technical Field
The invention relates to a radar target track starting method, in particular to a radar target track starting method based on a quantum particle swarm and an LGBM.
Background
The radar target tracking technology is widely applied to the fields of military, civil and the like, and the track initiation is one of important links in the target tracking technology, and the good track initiation can lay a foundation for the stable tracking of subsequent targets. The track initiation is a process of automatically establishing a track for a target entering a radar power zone, and mainly comprises three aspects of candidate track generation, track initiation judgment and formal track establishment as shown in fig. 1. The quality of track initiation is mainly evaluated from the following two aspects: firstly, the real track is established as accurately as possible, and secondly, the false track is avoided being established as far as possible. This means that high quality track initiation corresponds to low leakage rate and low false alarm rate, which is important for achieving multi-target fast and stable tracking. Because of the influence of sea clutter interference, enemy electronic interference, trace loss, target maneuver and the like, the full-automatic track initiation still has great difficulty, and further track information excavation and establishment of a more suitable initiation judgment model are required.
Classical track initiation methods fall into two main categories: sequential processing techniques and batch processing techniques. The basic idea of the method is to process echo data obtained by each scanning one by one, and judge whether to establish a track according to a point track association result in a certain time window. The sequential processing technology has lower calculation amount and is suitable for the situation that the background clutter is weaker. The representative method of the batch processing technology is a Hough transformation method and an improved algorithm thereof, and the main idea is that echo data obtained by multiple scanning are processed in a combined way, and the Hough transformation is utilized to realize incoherent accumulation of echo signals, so that the track starting performance is improved. The batch processing technology is suitable for the situation of strong clutter background, but is often used for offline processing due to larger calculated amount and weaker real-time performance. Under a complex background, both algorithms have difficulty in achieving real-time full-automatic track initiation under a complex background.
The machine learning algorithm is initially applied to the track initiation problem, and the method regards the track initiation problem as a problem of judging whether a target is real or not according to radar observation combination data, and belongs to the classification category. Compared with the traditional decision method, the machine learning algorithm has better adaptability and decision accuracy. However, no deep mining is performed on the initial track characteristic information at present, and different influences caused by the missing condition rate and the false alarm rate are not considered.
Disclosure of Invention
The invention aims to: the invention aims to solve the technical problem of providing a radar target track starting method based on a quantum particle swarm and an LGBM (land based on beam) aiming at the defects of the prior art.
In order to solve the technical problems, the invention discloses a radar target track starting method based on a quantum particle swarm and an LGBM, which comprises the following steps:
step 1: recording initial track information meeting basic starting conditions, marking, and forming an initial feature matrix and a corresponding label value through preliminary processing;
step 2: acquiring features on the time and space dimensions of an initial track through statistical calculation, and combining the features with an initial feature matrix to form final input feature data;
step 3: according to the final input characteristic data and the corresponding labels, a LightGbm model is established, model training is carried out, parameter selection of the model is carried out by utilizing grid search and cross verification, and the optimal LightGbm model in training is stored;
step 4: calculating the accuracy P1 of the real track and the accuracy P2 of the false track according to the probability output and the probability threshold of the optimal LightGbm model, designing a target function, and optimizing the target function by using a QPSO algorithm to obtain the optimal probability threshold;
step 5: performing online real-time prediction by using a LightGbm model, and performing statistical calculation on the time predicted by the LightGbm model to verify the online real-time prediction capability of the LightGbm model;
step 6: and (3) periodically updating the LightGbm model, checking whether the time accumulation reaches a LightGbm model updating period or whether a forecast error of the LightGbm model reaches a condition of the LightGbm model updating, and updating if the time accumulation reaches the updating period or the updating condition.
The beneficial effects are that:
(1) Aiming at the situations of radar in complex environments, such as Jiang Hai clutter, artificial electronic interference and the like, the problems of multiple false batches, track interruption, more jump points and the like in the track starting process are reduced.
(2) The method has the advantages that the track starting problem is converted into the common classification problem in machine learning, the characteristic information in the initial track is deeply mined, the intelligent model LightGbm is used for prediction classification, and compared with the traditional method, the LightGbm model is fully combined with various characteristics to perform nonlinear judgment, so that a better effect can be achieved.
(3) The machine learning method is used for replacing the prior knowledge in the traditional method to manually select the classification threshold, so that the dependence on prior information is obviously reduced, and the self-adaptability of the model is improved.
(4) The objective function distinguishes different costs brought by the correct initial real track and the false track, and the QPSO algorithm is utilized to obtain the optimal probability threshold under the given cost weighted combination condition, so that the optimal effect is realized.
Drawings
The foregoing and/or other advantages of the invention will become more apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings and detailed description.
FIG. 1 is a schematic block diagram of a target track initiation flow.
Fig. 2 is a flow chart of an intelligent full-automatic track initiation method based on feature extraction in a complex background environment of the invention.
FIG. 3 is a full-automatic initial tracking screenshot of an example implementation of the invention in a radar tracking system.
Detailed Description
The invention provides an intelligent full-automatic track initiation method of a radar target in a complex environment, namely a radar target track initiation method based on a quantum particle swarm and an LGBM (Light Gradient Boosting Machine, namely a light Gbm, namely a lightweight gradient elevator), which is used for improving the track initiation efficiency of the radar target in the complex background environment, and comprises the following specific steps:
step 1: recording initial track information meeting basic starting conditions through display control software, manually marking, and forming an initial feature matrix and a label value corresponding to the initial feature matrix through preliminary processing;
step 2: the method comprises the steps of obtaining features of an initial track in time and space dimensions through statistical calculation, and combining the features with the initial features to form final input feature data;
step 3: building a LightGbm integrated learning model (refer to Qi M. LightGBM: AHighly Efficient Gradient Boosting Decision Tree [ C ]// Neural Information Processing systems. Curran Associates Inc. 2017.), training the model, performing parameter selection of the model by utilizing grid search and cross verification, and storing an optimal model;
step 4: the accuracy P1 of a real track and the accuracy P2 of a false track are calculated according to the model probability output and the probability threshold, an objective function is designed according to the P1 and the P2, and an QPSO (Quantum-behaved Particle Swarm Optimization, quantum particle swarm) algorithm is used for optimizing (refer to Sun J, xu W, feng B.A global search strategy of Quantum-behaved particle swarm optimization [ C ]// Proc IEEE Conference on Cybernetics & Intelligent systems.2004.) objective function to obtain an optimal probability threshold;
step 5: performing online real-time prediction by using the model, and performing model prediction time statistical calculation on a corresponding platform in order to verify the online real-time prediction capability of the model;
step 6: and (5) updating the model regularly, and checking the model prediction error regularly to see whether the condition of updating the model is met.
In step 1; the condition of meeting basic starting points means that the number of the associated points contained in the initial track reaches the starting points requirement, the starting points requirement can be specified according to actual requirements, and the starting points requirement is set to be 5 when the method is applied to engineering;
in step 1; the manual labeling method is mainly characterized in that the manual observation is adopted, and because the initial target is set to basically accord with uniform linear motion in the starting stage, the original video information, the continuity and the jitter of each period of track can be referred to determine the real track and the false track, so that each track is labeled;
in step 1; the initial feature matrix information obtained by the preliminary processing mainly comprises the following two aspects of information:
on the one hand, the latest N pieces of point trace information are contained, N is set as a starting point requirement, and each piece of point trace information contains: distance r, azimuth a, amplitude amp, distance width re, azimuth width ae, and time frame number t.
On the other hand, the radar working mode comprises other information such as radar working mode and weather environment envi information, and when the radar working mode is used, the working mode and the weather environment need to be subjected to independent heat coding, for example, when the radar working mode comprises three types of short range, medium range and long range, the three working modes are respectively [1, 0], [0,1,0], [0, 1] through independent heat coding; the weather environment information can comprise sunny days, rainy days, snowy days and cloudy days, and the same three weather information is respectively [1, 0], [0,1,0], [0, 1] through independent heat coding.
The initial feature matrix is:
M1＝[r1,r2,r3,r4,r5,a1,a2,a3,a4,a5,amp1,amp2,amp 3 ,amp4,amp5,re1,re2,re3,re4,re5,ae1,ae2,ae3,ae4,ae5,t1,t2,t3,t4,t5,workmode,envi]；
in step 1; the tag values are 0 and 1,0 represents a false track, and 1 represents a real track;
in step 2; statistical analysis calculation is carried out on the original data from two dimensions of space and time, and further obtained characteristic information comprises the following steps:
time dimension: amplitude variance varAmp, amplitude mean avgAmp, distance width variance reVar, distance width mean avgRe, azimuth width variance varAe, azimuth width mean avgAe, average associated distance avgConnDis, associated probability cibnnProba;
spatial dimension: radial velocity vRad, tangential velocity vran, radial velocity variation Δvrad, tangential velocity variation Δvtan, and dot-navigation correlation distance connDis;
the calculation formula of the characteristics is as follows:
amplitude mean:
/>
amplitude variance:
distance width mean:
distance width variance:
azimuth width mean:
azimuth width variance:
ith correlation distance:
correlating the distance average:
probability of association:
radial velocity:
tangential velocity:
adjacent radial velocity difference: deltavRad j ＝|vRad j+1 -vRad j | j≤N-2
Adjacent tangential velocity difference Δvtan j ＝|vTan j+1 -vTan j | j≤N-2
Wherein DeltaR i Representing the difference in radial distance, ΔA, at the time of the ith trace correlation i Represents the difference of azimuth when the ith point trace is associated, r represents the distance precision of the radar, a represents the azimuth precision of the radar, dotCnt represents the number of associated points, period represents the processing cycle number, and DeltavRad j Represents the difference between the j+1th radial velocity and the j-th radial velocity, deltavTan j Representing the difference between the j+1th tangential velocity and the j-th tangential velocity.
In step 3, the parameter setting method of the LightGbm model is as follows: parameters determined by the model are fixed according to experience knowledge, and parameters to be optimized are adjusted by using a grid searching and cross-validation method; wherein the determinable model parameter reference settings are shown in Table 1:
table 1 table of determined LightGbm model parameters
| Parameters (parameters) | Value taking |
| Frame type (boosting_type) | Gradient lifting tree (gbdt) |
| Task (object) | Two-class (binary) |
| Evaluation function (metrics) | binary_logloss |
| Using feature_fraction | 0.8 |
The parameters of the model to be adjusted and optimized and the optimization range are shown in table 2:
TABLE 2 LightGbm model parameter Table to be optimized
| Parameters (parameters) | Optimization scope |
| Learning rate (learning_rate) | [0.01,0.02,0.05,0.1] |
| Number of base learners (n_estimators) | [100,200,500,800] |
| Tree maximum depth (max_depth) | [3,4,5,6] |
Optimizing parameter interpretation:
learning_rate: the model is trained slowly with smaller learning rate, and model performance with better stability can be obtained by generally selecting smaller learning rate;
n_evastiators: the number of the base learners is also the iteration number of boosting, and generally, larger iteration number is selected, so that the efficiency is better, but the model is easy to be over-fitted if the number of the base learners is too large;
max_depth: the maximum depth of the tree model is generally set to be between 3 and 6, and the parameters are important parameters for preventing the model from being overfitted, and have decisive influence on the model performance and generalization capability;
selecting optimal parameters by using a five-fold cross-validation method (refer to SVM mill load prediction [ J ]. Chinese test, 2017 (1)) based on grid search and cross-validation through the given parameter combinations, and storing a corresponding optimal model;
in step 4, the steps of calculating the accuracy rate P1 of the real track and the accuracy rate P2 of the false track according to the model probability output and the probability threshold value are as follows:
in step 4-1, we can obtain the probability output proba of the kth training sample according to the trained optimal model k To obtain the final classification result labelpred k Further determination of the probability threshold α is required, and discrimination is performed according to the following formula:
in step 4-2, according to the real labels and the predicted labels of all samples, P1 and P2 are obtained by the following calculation:
wherein TP represents the number of real tracks in all training samples that are model predicted as real tracks, TN represents the number of false tracks in all training samples that are model predicted as false tracks, FP represents the number of false tracks in all training samples that are model predicted as real tracks, FN represents the number of real tracks in all training samples that are model predicted as false tracks;
in step 4, the design of the objective function is mainly based on the track-initiated objective: firstly, the real track is built as accurately as possible, and secondly, the false track is built as little as possible. These two costs need to be taken into account, and the final objective function is formed according to P1 and P2 obtained in the previous step, as follows:
Fitness function ＝(λ*P1+P2)/(λ+1),λ>0
wherein, lambda is a weight coefficient, and the larger lambda represents the more important accuracy of the real track, namely, the more the real track is expected to normally start as much as possible, and the lower the leakage rate is; the smaller λ represents the more accurate we see for the heavy false track, i.e. we prefer as few false track starts as possible, representing as low a false alarm rate as possible. In combination with the requirement of the user, the user more hopefully does not miss any real tracks on the premise of a certain false alarm rate, namely, a small amount of false tracks are allowed to start, and because the small amount of false tracks can be judged and deleted by a track evaluation module after further data accumulation, the influence is smaller, so lambda can be set to a larger value, the lambda is set to 5 in engineering application, and the final optimization objective function can be as follows:
Fitness function ＝(5*P1+P2)/(5+1)
Fitness function ＝(5*TP/(TP+FN)+TN/(TN+FP)/6
in step 4, in order to obtain the optimized threshold α, the QPSO algorithm is used to perform maximum value solution on the objective function, and the setting of QPSO model parameters is shown in table 3:
TABLE 3 QPSO model parameter settings
| Parameters (parameters) | Value taking |
| Particle number | 50 |
| Particle dimension | 1 |
| Control coefficient | 0.7 |
| Maximum number of iterations | 100 |
| Probability threshold alpha max | 1 |
| Probability threshold alpha minimum | 0 |
Obtaining a probability threshold alpha when the objective function converges to a maximum 0 ；
In step 5, the step of calling the model on line in real time to realize classification prediction comprises the following steps:
in step 5-1, loading a trained optimal model;
in step 5-2, for each new track sample, processing according to the original data to obtain an initial feature matrix, a time dimension feature matrix and a space dimension feature matrix, and combining to obtain a final model input feature matrix;
in step 5-3, the model is called to obtain the output of the prediction probability, and the prediction probability is obtained by the method and the optimized threshold alpha 0 In contrast, if the predictive probability output is greater than the threshold α 0 Then judge as the true track, if the prediction is probableThe rate output is less than the threshold alpha 0 Judging the navigation path as a false track;
in step 5, in order to verify the real-time performance of the model, the total time required for statistically predicting 5000 samples on the localization platform is 250ms, and then the time required for predicting each sample is 0.05ms, thereby meeting the real-time performance requirement.
In step 6, the periodic update of the model mainly judges two conditions, namely whether the time accumulation reaches the model update period or the model prediction accuracy of the last period of time is counted.
The invention relates to an intelligent full-automatic track starting method based on feature extraction, which is characterized in that a track distinguishing LightGbm model is established by combining with manually extracting multidimensional features of an initial track, an objective function considers different costs brought by real track distinguishing accuracy and false track distinguishing accuracy, and a QPSO algorithm is used for optimizing and obtaining an optimal probability threshold value, so that the optimal full-automatic track starting is finally realized, dependence on priori knowledge is greatly reduced, the self-adaptability of the algorithm is improved, and the track starting performance under a complex environment is finally improved.
Examples:
with reference to fig. 2, the intelligent full-automatic track initiation method based on feature extraction in a complex background environment of the invention comprises the following steps:
firstly, establishing a full-automatic area in a sea clutter area through display control software, and recording all initial track information meeting basic starting conditions of the radar in different working modes and different working environments. The condition of meeting the basic starting point means that the number of the associated points contained in the initial track reaches the starting point requirement, and the starting point requirement can be specified according to the actual requirement and is set to be 5 in the project.
By manual observation, judging whether the target basically accords with the uniform linear motion state by combining the original video information with the continuity and the jitter of each period of the track, thereby judging the authenticity of the track, if the track is the authentic track, marking a label 1, and if the track is the fake track, marking a label 0;
the initial characteristic matrix is formed by initially processing the track original information, and the initial characteristic matrix information mainly comprises the following information:
on the one hand, the latest N pieces of point trace information are contained, N is set as a starting point requirement, and each piece of point trace information contains: distance r, azimuth a, amplitude amp, distance width re, azimuth width ae, and time frame number t.
On the other hand, the radar working mode comprises other information such as radar working mode and weather environment envi information, and when the radar working mode is used, the working mode and the weather environment need to be subjected to independent heat coding, for example, when the radar working mode comprises three types of short range, medium range and long range, the three working modes are respectively [1, 0], [0,1,0], [0, 1] through independent heat coding; the weather environment information can comprise sunny days, rainy days, snowy days and cloudy days, and the same three weather information is respectively [1, 0], [0,1,0], [0, 1] through independent heat coding.
The initial feature matrix is sorted as:
M1＝[r1,r2,r3,r4,r5,a1,a2,a3,a4,a5,amp1,amp2,amp3,amp4,amp5,re1,re2,re3,re4,re5,ae1,ae2,ae3,ae4,ae5,t1,t2,t3,t4,t5,workmode,envi]。
secondly, further statistical analysis is carried out on the acquired original data from two dimensions of space and time, and feature information is further extracted, wherein the feature information mainly comprises the following features:
time dimension: amplitude variance varAmp, amplitude mean avgAmp, distance width variance reVar, distance width mean avgRe, azimuth width variance varAe, azimuth width mean avgAe, average associated distance avgConnDis, associated probability connProba;
spatial dimension: radial velocity vRad, tangential velocity vran, radial velocity variation Δvrad, tangential velocity variation Δvtan, and dot-navigation correlation distance connDis;
the calculation formula of the characteristics is as follows:
amplitude mean:
amplitude variance:
distance width mean:
/>
distance width variance:
azimuth width mean:
azimuth width variance:
ith correlation distance:
correlating the distance average:
probability of association:
radial velocity:
tangential velocity:
adjacent radial velocity difference: deltavRad j ＝|vRad j+1 -vRad j | j≤N-2
Adjacent tangential velocity difference Δvtan j ＝|vTan j+1 -vTan j | j≤N-2
Wherein DeltaR i Representing the difference in radial distance, ΔA, at the time of the ith trace correlation i Represents the difference of azimuth when the ith point trace is associated, r represents the distance precision of the radar, a represents the azimuth precision of the radar, dotCnt represents the number of associated points, period represents the processing cycle number, and DeltavRad j Represents the difference between the j+1th radial velocity and the j-th radial velocity, deltavTan j Representing the difference between the j+1th tangential velocity and the j-th tangential velocity.
And combining the features and the initial features obtained in the second step to obtain a final input feature matrix.
Thirdly, establishing a LightGbm model, fixing parameters determined by the model according to experience knowledge, and adjusting the parameters to be optimized by using a grid searching and cross-validation method; wherein the determinable parameter reference settings are shown in table 4:
table 4 table of determined LightGbm model parameters
| Parameters (parameters) | Value taking |
| Frame type (boosting_type) | Gradient lifting tree (gbdt) |
| Task (object) | Two-class (binary) |
| Evaluation function (metrics) | binary_logloss |
| Using feature_fraction | 0.8 |
The parameters and optimization ranges to be adjusted are shown in table 5:
TABLE 5 LightGbm model parameter Table to be optimized
| Parameters (parameters) | Optimization scope |
| Learning rate (learning_rate) | [0.01,0.02,0.05,0.1] |
| Number of base learners (n_estimators) | [100,200,500,800] |
| Tree maximum depth (max_depth) | [3,4,5,6] |
Introduction of optimization parameters:
learning_rate: the model is trained slowly with smaller learning rate, and model performance with better stability can be obtained by generally selecting smaller learning rate;
n_evastiators: the number of the base learners is also the iteration number of boosting, and generally, larger iteration number is selected, so that the efficiency is better, but the model is easy to be over-fitted if the number of the base learners is too large;
max_depth: the maximum depth of the tree model is generally set to be between 3 and 6, and the parameters are important parameters for preventing the model from being overfitted, and have decisive influence on the model performance and generalization capability;
using five-fold cross validation (cross-validation), traversing the given parameter combinations by a grid search (grid search) method to select optimal parameters, and storing a corresponding optimal model;
fourth, we can get the probability output proba of the kth training sample according to the trained optimal model k To obtain the final classification result labelPred k Further determination of the probability threshold α is required, and discrimination is performed according to the following formula:
according to the real labels and the predicted labels of all the samples, the accuracy P1 of the real track and the accuracy P2 of the false track can be obtained through the following calculation:
wherein TP represents the number of the real tracks predicted by the model, TN represents the number of the false tracks predicted by the model, FP represents the number of the false tracks predicted by the model, FN represents the number of the false tracks predicted by the model;
in the design of objective functions, two main objective tasks of track initiation need to be considered: firstly, the real track is built as accurately as possible, and secondly, the false track is built as little as possible, so that the two costs need to be considered to form a final objective function, as follows:
Fitness function ＝(λ*P1+P2)/(λ+1),λ>0
wherein, lambda is a weight coefficient, and the larger lambda represents the more important accuracy of the real track, namely, the more the real track is expected to normally start as much as possible, and the lower the leakage rate is; the smaller λ represents the more accurate we see for the heavy false track, i.e. we prefer as few false track starts as possible, representing as low a false alarm rate as possible. In combination with the requirement of the user, the user more hopefully does not miss any real tracks on the premise of a certain false alarm rate, namely, a small amount of false tracks are allowed to start, and because the small amount of false tracks can be judged and deleted by a track evaluation module after further data accumulation, the influence is smaller, so lambda can be set to a larger value, the lambda is set to 5 in engineering application, and the final optimization objective function can be as follows:
Fitness function ＝(5*P1+P2)/(5+1)
Fitness function ＝(5*TP/(TP+FN)+TN/(TN+FP)/6
wherein TP, TN, FN, FP are all functions of the probability threshold α;
after the objective function is obtained, the QPSO algorithm is used to optimize the objective function to obtain an optimal probability threshold. The general parameter settings for the QPSO algorithm are shown in Table 6:
TABLE 6 QPSO model parameter settings
| Parameters (parameters) | Value taking |
| Particle number | 50 |
| Particle dimension | 1 |
| Control coefficient | 0.7 |
| Maximum number of iterations | 100 |
| Probability threshold alpha max | 1 |
| Probability threshold alpha minimum | 0 |
Obtaining a probability threshold alpha when the objective function converges to a maximum 0 ；
Fifthly, carrying out online real-time prediction by using the model, wherein the method mainly comprises the following steps of:
1. loading a trained optimal model;
2. for each new track sample, processing according to the original data to obtain an initial feature matrix, a time dimension feature matrix and a space dimension feature matrix, and combining to obtain a final model input feature matrix;
3. the model is called to predict to obtain probability output, and the probability output is obtained through the prediction with a probability threshold alpha 0 Comparing, judging the real track if the output is larger than the threshold value, and judging the false track if the output is smaller than the threshold value;
in addition, in order to verify the real-time performance of the model to meet the engineering application requirements, the total time required for predicting 5000 samples is counted to be 250ms on a localization platform, and then the time required for predicting each sample is 0.05ms, so that the engineering application real-time performance requirements are met.
Sixth, the model is updated regularly. Judging whether the time accumulation reaches a model updating period or calculating the prediction efficiency of the time model according to the observed real data and the model prediction result, and if the time accumulation reaches the model updating or the prediction error is larger than a set prediction error threshold value, retraining the model by using all accumulated data.
The invention will be further illustrated by means of an example of implementation of the application in engineering and its evaluation of effect, described below in connection with fig. 3. FIG. 3 is a diagram showing a radar display screen of a pair of sea radars during fully automatic start and tracking of sea clutter regions, where TVXXX is the track number, e.g., number TV0005 indicates that the track with lot number 0005 is in a normal tracking state. Through multi-period manual observation and confirmation, targets which are already tracked in batches in a picture are real targets, and false targets are basically avoided in the whole process, so that the method can realize full-automatic starting of target tracks in a complex background environment, and can greatly reduce the operation load of radar operators.
In a specific implementation, the application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and the computer program can run the invention content of the radar target track starting method based on the quantum particle swarm and the LGBM and part or all of the steps in each embodiment when being executed by the data processing unit. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), a random-access memory (random access memory, RAM), or the like.
It will be apparent to those skilled in the art that the technical solutions in the embodiments of the present invention may be implemented by means of a computer program and its corresponding general hardware platform. Based on such understanding, the technical solutions in the embodiments of the present invention may be embodied essentially or in the form of a computer program, i.e. a software product, which may be stored in a storage medium, and include several instructions to cause a device (which may be a personal computer, a server, a single-chip microcomputer, MUU or a network device, etc.) including a data processing unit to perform the methods described in the embodiments or some parts of the embodiments of the present invention.
The invention provides a thought and a method for starting a radar target track based on a quantum particle swarm and an LGBM, and the method and the way for realizing the technical scheme are numerous, the above description is only a preferred embodiment of the invention, and it should be pointed out that a plurality of improvements and modifications can be made by those skilled in the art without departing from the principle of the invention, and the improvements and modifications are also regarded as the protection scope of the invention. The components not explicitly described in this embodiment can be implemented by using the prior art.
Claims (10)
1. The radar target track starting method based on the quantum particle swarm and the LGBM is characterized by comprising the following steps of:
step 1: recording initial track information meeting basic starting conditions, marking, and forming an initial feature matrix and a corresponding label value through preliminary processing;
step 2: acquiring features on the time and space dimensions of an initial track through statistical calculation, and combining the features with an initial feature matrix to form final input feature data;
step 3: according to the final input characteristic data and the corresponding labels, a LightGbm model is established, model training is carried out, parameter selection of the model is carried out by utilizing grid search and cross verification, and the optimal LightGbm model in training is stored;
step 4: calculating the accuracy P1 of the real track and the accuracy P2 of the false track according to the probability output and the probability threshold of the optimal LightGbm model, designing a target function, and optimizing the target function by using a QPSO algorithm to obtain the optimal probability threshold;
step 5: performing online real-time prediction by using a LightGbm model, and performing statistical calculation on the time predicted by the LightGbm model to verify the online real-time prediction capability of the LightGbm model;
step 6: and (3) periodically updating the LightGbm model, checking whether the time accumulation reaches a LightGbm model updating period or whether a forecast error of the LightGbm model reaches a condition of the LightGbm model updating, and updating if the time accumulation reaches the updating period or the updating condition.
2. The method for initiating a radar target track based on a population of particles and LGBM according to claim 1, wherein in step 1:
the number of associated points contained in the initial track meets the starting point requirement, the starting point requirement is specified according to actual requirements, and the number is set to be 5 when the method is applied in engineering;
the labeling in the step 1, namely manual labeling, comprises the following steps: determining a real track and a false track by manual observation with reference to original video information and the continuity and jitter of each period track, so as to label each track;
the preliminary processing described in step 1 is a process of forming an initial feature matrix, where the initial feature matrix M1 is as follows:
M1＝[r 1 ,r 2 ,…,r i ,…,r N ,a 1 ,a 2 ,…,a i ,…,a N ,amp 1 ,amp 2 ,…,amp i ,…,amp N ,
re 1 ，re 2 ，...，re i ，...，re N ，ae 1 ，ae 2 ，...，ae i ，...，ae N ，t 1 ，t 2 ，…，t i ，…，t N ，
workmode，envi]
wherein N is the set starting point requirement, namely the initial characteristic matrix M1 contains the latest N pieces of point trace information in the initial track, and r i Represents the distance of the ith trace, a i Representing the azimuth of the ith trace, amp i Representing the amplitude, re, of the ith trace i Represents the distance width of the ith trace, ae i The azimuth width of the ith trace, t i Representing the number of time frames of the ith trace, workmode representing the radar operating mode, and envi representing the weather environment;
carrying out independent heat coding on a radar working mode workmode and a weather environment envi;
the tag values described in step 1 are used to represent false tracks and true tracks.
3. The method of claim 2, wherein the initial track time and space dimension features of step 2 comprise:
features in the initial track time dimension, including: amplitude variance varAmp, amplitude mean avgAmp, distance width variance reVar, distance width mean avgRe, azimuth width variance varAe, azimuth width mean avgAe, average associated distance avgConnDis, associated probability connProba;
features in the initial track space dimension, including: radial velocity vRad, tangential velocity vran, radial velocity variation Δvrad, tangential velocity variation Δvtan, and dot-navigation correlation distance connDis.
4. A method for initiating a radar target track based on a group of particles and LGBM according to claim 3, wherein the acquiring features in the initial track time and space dimensions by statistical calculation in step 2 specifically includes:
amplitude mean:
amplitude variance:
distance width mean:
distance width variance:
azimuth width mean:
azimuth width variance:
ith correlation distance:
correlating the distance average:
probability of association:
radial velocity:
tangential velocity:
adjacent radial velocity difference: deltavRad j ＝|vRad j+1 -vRad j | j≤N-2
Adjacent tangential velocity difference Δvtan j ＝|vTan j+1 -uTan j | j≤N-2
Wherein DeltaR i Representing the difference in radial distance, ΔA, at the time of the ith trace correlation i Represents the difference of azimuth when the ith point trace is associated, r represents the distance precision of the radar, a represents the azimuth precision of the radar, dotCnt represents the number of associated points, period represents the processing cycle number, and DeltavRad j Represents the difference between the j+1th radial velocity and the j th radial velocity, and Δvtanj represents the difference between the j+1th tangential velocity and the j th tangential velocity.
5. The method for initiating radar target tracks based on the quantum particle swarm and the LGBM according to claim 4, wherein the parameters of the model performed by using grid search and cross-validation in the step 3 specifically comprises:
selecting model parameters of optimization to be adjusted and an optimization range thereof from the LightGbm model, wherein the model parameters of the optimization to be adjusted comprise: learning rate, number of base learners, and maximum depth of tree model;
and traversing the parameter combinations by using five-fold cross verification through a grid searching method to select optimal parameters, and storing a corresponding model as an optimal model.
6. The method for starting radar target tracks based on the quantum particle swarm and the LGBM according to claim 5, wherein the calculating the accuracy P1 of the real track and the accuracy P2 of the false track according to the probability output and the probability threshold of the optimal LightGbm model in step 4 comprises the following steps:
step 4-1, obtaining a probability output proba of a kth training sample according to the trained optimal LightGbm model k Final classification result labelPred k Discrimination is performed according to the following formula:
wherein α is a probability threshold;
step 4-2, according to the real labels and the predicted labels of all training samples, the following calculation is performed to obtain the accuracy P1 of the real track and the accuracy P2 of the false track:
P1＝TP/(TP+FN)
P2＝TN/(TN+FP)
wherein TP represents the number of real tracks predicted to be real by the optimal LightGbm model in all training samples, TN represents the number of false tracks predicted to be false by the optimal LightGbm model in all training samples, FP represents the number of false tracks predicted to be real by the optimal LightGbm model in all training samples, and FN represents the number of false tracks predicted to be false by the optimal LightGbm model in all training samples.
7. The method of claim 6, wherein the objective function Fitness in step 4 function The expression is as follows:
Fitness function ＝(λ*P1+P2)/(λ+1)，λ＞0
wherein λ is a weight coefficient, and the final optimization objective function is:
Fitness function ＝(λ*TP/(TP+FN)+TN/(TN+FP))/(λ+1)
all of which are a function of the TP, TN, FN and FP average probability threshold α.
8. The method for radar target track initiation based on a population of quantum particles and LGBM according to claim 7, wherein said method for optimizing an objective function using QPSO algorithm in step 4 comprises: maximum solving of objective function using QPSO algorithm, probability threshold alpha when objective function converges to maximum 0 I.e. the optimal probability threshold.
9. The method for starting radar target tracks based on the quantum particle swarm and the LGBM according to claim 8, wherein the online real-time prediction by applying the LightGbm model in step 5 comprises the following specific steps:
step 5-1, loading a trained optimal LightGbm model;
step 5-2, for each new track sample, processing according to the original data to obtain an initial feature matrix, a time dimension feature matrix and a space dimension feature matrix, and combining to obtain a final model input feature matrix;
step 5-3, calling the loaded model to obtain a prediction probability output and an optimized threshold alpha 0 By contrast, if the predictive probability output is greater than the threshold α 0 Judging the real track, otherwise judging the false track;
and 5-4, counting the total time of predicting a preset number of samples by using the loaded model, calculating the time consumed by predicting each sample, and completing the verification of the online real-time prediction capability of the LightGbm model.
10. The method for starting a radar target track based on a group of quantum particles and LGBM according to claim 9, wherein the starting point requirement N in step 1 is set to 5; the weight coefficient λ described in step 4 is set to 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310084045.0A CN116106890A (en) | 2023-02-08 | 2023-02-08 | Radar target track starting method based on quantum particle swarm and LGBM |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310084045.0A CN116106890A (en) | 2023-02-08 | 2023-02-08 | Radar target track starting method based on quantum particle swarm and LGBM |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116106890A true CN116106890A (en) | 2023-05-12 |
Family
ID=86267001
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310084045.0A Pending CN116106890A (en) | 2023-02-08 | 2023-02-08 | Radar target track starting method based on quantum particle swarm and LGBM |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116106890A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116359907A (en) * | 2023-06-02 | 2023-06-30 | 西安晟昕科技股份有限公司 | Track processing method of radar data |
-
2023
- 2023-02-08 CN CN202310084045.0A patent/CN116106890A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116359907A (en) * | 2023-06-02 | 2023-06-30 | 西安晟昕科技股份有限公司 | Track processing method of radar data |
| CN116359907B (en) * | 2023-06-02 | 2023-08-15 | 西安晟昕科技股份有限公司 | Track processing method of radar data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20220188568A1 (en) | Methods and systems for mining minority-class data samples for training a neural network | |
| US7664329B2 (en) | Block-based Gaussian mixture model video motion detection | |
| CN111161315B (en) | Multi-target tracking method and system based on graph neural network | |
| CN103473540B (en) | The modeling of intelligent transportation system track of vehicle increment type and online method for detecting abnormality | |
| CN109239669B (en) | Self-evolution radar target detection algorithm based on deep learning | |
| CN111950810B (en) | Multi-variable time sequence prediction method and equipment based on self-evolution pre-training | |
| CN112363896B (en) | Log abnormality detection system | |
| CN105844501A (en) | Consumption behavior risk control system and method | |
| CN114022904B (en) | Noise robust pedestrian re-identification method based on two stages | |
| CN116106890A (en) | Radar target track starting method based on quantum particle swarm and LGBM | |
| CN116244647A (en) | Unmanned aerial vehicle cluster running state estimation method | |
| CN110849372A (en) | Underwater multi-target track association method based on EM clustering | |
| CN112714130A (en) | Big data-based adaptive network security situation sensing method | |
| Liu et al. | Grey-based approach for estimating software reliability under nonhomogeneous Poisson process | |
| CN115063664A (en) | Model learning method, training method and system for industrial vision detection | |
| CN116862079B (en) | Enterprise pollutant emission prediction method and prediction system | |
| CN117630860A (en) | Gesture recognition method of millimeter wave radar | |
| CN112418522A (en) | Industrial heating furnace steel temperature prediction method based on three-branch integrated prediction model | |
| CN111984514A (en) | Prophet-bLSTM-DTW-based log anomaly detection method | |
| CN111797935A (en) | Semi-supervised deep network picture classification method based on group intelligence | |
| CN115015908A (en) | Radar target data association method based on graph neural network | |
| CN113469816A (en) | Digital currency identification method, system and storage medium based on multigroup technology | |
| CN113011597A (en) | Deep learning method and device for regression task | |
| CN110988803A (en) | Radar radiation source individual identification system with high confidence degree dynamic adjustment | |
| Zhang et al. | Pondering About Task Spatial Misalignment: Classification-Localization Equilibrated Object Detection |
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 |