Automatic target tracking method and device

Through technologies such as automatic target tracking method and multi-head self-attention mechanism, the contradiction between filtering accuracy and maneuver detection sensitivity in the target tracking system is solved, and the continuous, accurate and stable tracking of the target is achieved, and the system performance is improved.

CN115420293BActive Publication Date: 2024-11-01NAVAL AVIATION UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211059934.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-31
Publication Date
2024-11-01
Estimated Expiration
2042-08-31

AI Technical Summary

Technical Problem

There is a contradiction between the balanced filtering accuracy and maneuver detection sensitivity. High filtering accuracy requires more measurement points, but too long smoothing time will lead to a decrease in maneuver detection sensitivity and it is difficult to achieve continuous, accurate and stable goals. track.

Method used

Using automatic target tracking method, the target admission method, the tracking network and network input generation method and the output conversion method are used to generate and process measurement and historical status information through the target admission method, the tracking network and the network input generation method and the output conversion method are used to design the tracking network of the multi-head self-attention mechanism and the feedforward network to generate and process measurement and historical status information to achieve the goal. Continuous tracking and stationary state estimation, combined with network training and optimization, improve the performance of the tracking system.

Benefits of technology

The continuous, accurate and stable tracking of the target is achieved, the filtering accuracy and maneuver detection sensitivity are improved, the probability of error tracking is reduced, and the technical requirements of the International Maritime Organization for the target tracking system are met.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115420293B_ABST
    Figure CN115420293B_ABST
Patent Text Reader

Abstract

The embodiment of the present invention provides an automatic target tracking method and device, the target tracking method is composed of a target recording method f A , Tracking Network T , network input generation method f I and the network output conversion method f O Joint composition, where the target admission method f A Using the measurement information at multiple times, the initial state information of the tracking target is obtained through the track initiation, and the tracking network is started in different tracking channels for the target to be tracked. T Then, the current measurement information and the target historical state information are used to generate the method f through network input I , get the tracking network f T Input, processed by the network, get the network output, and then pass the network output conversion method f O The forward conversion obtains the real-time status information of the tracking target, thereby realizing continuous tracking of the target. The embodiment of the present invention can effectively balance the contradictions in target tracking to realize continuous, accurate and stable tracking of the target.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to target tracking technology, and more specifically, to an automatic target tracking method and device, which is applicable to single-target or multi-target tracking problems in radar, optical, infrared and other early warning detection equipment. Background Art

[0002] Target detection is a technical activity that uses radar, optical, infrared and other early warning detection equipment to discover, locate and track targets. The results can be used as input and basis for further high-level processing such as attribute identification, threat assessment and combat decision-making. Target tracking is an important part of the target detection process. Its purpose is to connect the detection information from the same target at different times, and form time-series target state information with the same identity mark through filtering estimation, that is, target track, so as to achieve real-time, continuous and accurate grasp of individual targets. The input of target tracking is target detection data at different times, and the output of target tracking is multiple target tracks, which correspond to different targets in the actual environment. Each track is composed of multiple target state information at different times and has a unique identity mark. In actual tracking, a fixed-length digital string is generally used for track identity marking, also known as track batch number.

[0003] Target tracking technology is widely used in many fields, such as air traffic control, maritime navigation, maritime surveillance and military air defense warning. Taking maritime navigation as an example, the special equipment that realizes automatic target tracking, echo calculation and display, and collision prediction and avoidance is called Automatic Radar Plotting Aid (ARPA). The International Maritime Organization (IMO) requires ARPA as follows: "Observers can automatically obtain information about the target to be plotted, thereby reducing their workload, so that they can complete the plotting of several individual targets as if they were manually plotting a single target." In the MSC.192 (79) resolution adopted on December 6, 2004, IMO cancelled the term ARPA and replaced it with automatic "target tracking", which includes two stages: target admission and tracking filtering. Target admission describes the start of the target tracking process, which can be manual or automatic. When the target is admitted, the data collection and tracking related to the target are started. Once the tracking filter is activated, the target tracking system will continue to follow the target until the tracking is canceled or the target is "lost". Target tracking is mainly achieved through two major technical links: track interconnection and track filtering. The technical points are: the core of track interconnection lies in the selection of interconnection gates. Too large a gate will lead to too many clutter points entering, which is easy to cause interconnection errors, while too small a gate will reduce the number of interconnection points that meet the requirements, which is easy to cause interconnection failure. When the target is first captured, a larger gate must be set because it is uncertain in which direction the target will move; the improvement of tracking accuracy is mainly achieved through target measurement point accumulation and track filtering. A large number of continuous measurements can improve the accuracy of target position, speed and heading. However, if the smoothing time is too long, the automatic tracking system will become insensitive to track changes. As a result, small maneuvers may not be detected, and there will be a long delay before large maneuvers become obvious. On the other hand, if the smoothing time is very short, the output data will fluctuate. Obviously, the track filtering algorithm needs to comprehensively consider the smoothing amount and smoothing time, and needs to comprehensively coordinate the sensitivity of target maneuvers and data stability.In Resolution MSC.192(79), IMO's main technical requirements for target tracking systems are as follows: ① Whether it is manual or automatic registration, a simple target tracking device should have at least 20 to 30 tracking channels, and a complete device should have at least 40 tracking channels; ② When a target has been registered, the target tracking device should be able to display the target movement trend within 1 minute, with a circular positioning accuracy of 1 nautical mile, and give a target movement trend prediction within 3 minutes, with a circular positioning accuracy of 0.3 nautical miles; ③ It should be able to automatically track the target from beginning to end, and update the position, speed and heading of all registered targets in real time. and other information; ④ If the target is clearly displayed on the display 5 times in 10 antenna scans, or displayed in equal proportions, the target tracking device should be able to track such targets continuously and uninterruptedly; ⑤ The target tracking device should have a filtering function that can smooth the radar measurements and obtain a stable target state estimate, that is, obtain more accurate target position, speed and heading data. At the same time, the target tracking equipment should also have a target maneuver detection function that can detect the maneuver in the target in a timely and early manner; ⑥ The target tracking device should minimize the probability of false tracking, including the probability of target track crossing and exchanging.

[0004] In summary, there are contradictory problems that need to be solved in the target tracking system. For example, high filtering accuracy requires smoothing more measurement points, while too long smoothing time will lead to a decrease in the sensitivity of maneuver detection. Therefore, it is necessary to continuously study target tracking technology in order to build a target tracking system with strong tracking continuity, high filtering accuracy, and sensitive maneuver detection. Summary of the invention

[0005] The purpose of the present invention is to propose an automatic target tracking method and device, aiming to balance the contradictory problems in target tracking so as to achieve continuous, accurate and stable tracking of the target. The specific technical solution is as follows:

[0006] In a first aspect, an embodiment of the present invention provides an automatic target tracking method, the method comprising:

[0007] Automatic target tracking method by target admission method f A , Tracking Network T , network input generation method f I and the network output conversion method f O Joint composition, where the target admission method f A Using the measurement information at multiple times, the initial state information of the tracking target is obtained through the track initiation, and the tracking network is started in different tracking channels for the target to be tracked. T Then, the current measurement information and the target historical state information are used to generate the method f through network input I , get the tracking network f TInput, processed by the network, get the network output, and then pass the network output conversion method f O Forward conversion, to obtain the real-time status information of the tracking target, and then to achieve continuous tracking of the target, the measurement information at different times at least includes time and position information, and the status information of the tracking target at least includes time, position and speed information;

[0008] Tracking Network T There are two inputs, one is the stable measurement within the wave gate, and the other is generated by the network input method f I The other is the target historical steady state, which is obtained by accumulating the steady state output of the tracking network with time information. The output of the tracking network is the target steady state, which is further converted by the network output conversion method f O Forward conversion can finally get the current target state, where stability means not including time accumulation, stable measurement and stable state mean not including pure time and pure position information, and network input generation method f I According to the measurement gate set, the stable measurement in the output gate is calculated by using the stable state output and target state of the tracking network at the previous moment, and the network output conversion method f O It includes forward conversion and reverse conversion. The forward conversion calculates the current target state according to the current target steady state and the target state at the previous moment. The reverse conversion calculates the current target steady state according to the target state at the previous moment and the current target state.

[0009] Tracking Network T It consists of a measurement feature extraction part, a state feature extraction part, a unified feature generation part, and a stable state estimation part. The measurement feature extraction part is designed with a multi-head self-attention mechanism, a feedforward network, and a residual connection. The input is a stable measurement within the wave gate, and the output is a measurement feature. The state feature extraction part is designed with a feedforward network and a residual connection. The input is the historical stable state of the target, and the output is the state feature. The unified feature generation part is designed with a multi-head attention mechanism, a feedforward network, and a residual connection. The input is the measurement feature and the state feature, and the output is a unified feature. The measurement feature is used to generate the K matrix and the V matrix in the multi-head attention mechanism, and the state feature is used to generate the Q matrix of the first multi-head attention mechanism. The stable state estimation is designed with a feedforward network, the input is a unified feature, and the output is the current stable state of the target.

[0010] Tracking Network T The loss function of is composed of the steady state estimation loss and the network weight penalty term, where the steady state estimation loss is represented by the regression loss function, which is calculated by the steady state estimation and the true steady state. The true steady state can be obtained by reverse conversion using the network output conversion method.

[0011] The measurement data of early warning detection equipment and the true state value of the target contained therein are collected to construct a tracking network training data set. Each sample includes three parts: stable measurement input within the wave gate, target historical stable state input and target stable state output. T The two inputs and one output of the gate correspond to each other, where the stable measurement input is based on the measurement data collected by the early warning detection device at a certain time k and the state true value of a certain target at the corresponding time. By calling the network input generation method f I It is calculated that the historical steady state input of the target is based on the state truth value of the corresponding target k at multiple moments before, and the network output conversion method f O The target steady state output is obtained by reverse conversion based on the true value of the corresponding target state at time k, which is obtained by reverse conversion through the network output conversion method. The neural network optimization method is further used to optimize the tracking network f based on the tracking network training data set. T Train, optimize and debug until the tracking network performance meets the requirements, and finally output the trained tracking network f T ;

[0012] Deploy an automatic target tracking method in the early warning detection equipment, receive the measurement data generated by real-time detection, and adopt the target admission method f A Start the track and get the initial state information of the target. For each starting target, start the trained tracking network f under the limit of the number of tracking channels. T , real-time tracking is performed until tracking is cancelled or the target is lost.

[0013] In a second aspect, an embodiment of the present invention further provides an automatic target tracking device, the device comprising:

[0014] The target admission module is used to obtain the initial state of the target to be tracked, and then start the tracking network to continuously track the target. It can be started manually, that is, the target to be tracked is started at the operator's choice, or it can automatically start tracking the target in the entire detection area or in a limited area. One tracking network can track one target. For multiple targets started, the target admission module can start the corresponding number of tracking networks to track multiple targets;

[0015] Tracking network module, the main body of target tracking, consists of measurement feature extraction part, state feature extraction part, unified feature generation part, and steady state estimation part. It is started by the target acquisition module and automatically processes the measurement obtained by the detection device according to the target initial state input by the target acquisition module to achieve continuous tracking of the target;

[0016] An auxiliary computing module, including a network input generation method and a network output conversion method, is used to assist the tracking network in generating the required input and converting the output during the construction of the tracking network training data set and during the use of the tracking network;

[0017] The network training module includes two parts: tracking network training data set construction and tracking network training optimization, which are used to train, optimize and debug the tracking network.

[0018] An automatic target tracking method and device proposed in an embodiment of the present invention can provide a single-target or multi-target automatic tracking system suitable for radar, optical, infrared and other early warning detection equipment, which has the advantages of unified structure, balanced performance, good practical effect, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a flow chart of an automatic target tracking method provided by an embodiment of the present invention;

[0021] Figure 2 It is a schematic diagram of a tracking network structure provided by an embodiment of the present invention;

[0022] Figure 3 It is a structural schematic diagram of an automatic target tracking device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In order to construct a target tracking method with strong tracking continuity, high filtering accuracy and sensitive maneuvering detection, an embodiment of the present invention provides an automatic target tracking method, such as Figure 1 and Figure 2 As shown, the method comprises the following steps:

[0025] Step 1: Automatic target tracking method by target admission method f A , Tracking Network T, network input generation method f I and the network output conversion method f O Joint composition, where the target admission method f A Using the measurement information at multiple times, the initial state information of the tracking target is obtained through the track initiation, and the tracking network is started in different tracking channels for the target to be tracked. T Then, the current measurement information and the target historical state information are used to generate the method f through network input I , get the tracking network f T Input, processed by the network, get the network output, and then pass the network output conversion method f O Forward conversion, to obtain the real-time status information of the tracking target, and then to achieve continuous tracking of the target, the measurement information at different times at least includes time and position information, and the status information of the tracking target at least includes time, position and speed information;

[0026] Step 1.1: Automatic target tracking method by target admission method f A , Tracking Network T , network input generation method f I and the network output conversion method f O Joint composition, where the target admission method f A Using the measurement information at multiple times, the initial state information of the tracking target is obtained through the track initiation, and the tracking network is started in different tracking channels for the target to be tracked. T Then, the current measurement information and the target historical state information are used to generate the method f through network input I , get the tracking network f T Input, processed by the network, get the network output, and then pass the network output conversion method f O Forward conversion, to obtain the real-time status information of the tracking target, and then realize the continuous tracking of the target;

[0027] Step 1.2: The measurement information at different times includes at least time and position information. The state information of the tracking target includes at least time, position and speed information. The detection data obtained by the detection device in one detection cycle is set as Z k ={z k,1 ,z k,2 ,...,z k,n(k)}, contains n(k) measurement point information z k,i , each measurement information z k,i At least contains the detection time and location information, that is, z k,i =[t k,i ,x k,i ,y k,i ] T, and can also include target speed, target height information and attribute information. For radar, the attribute information can be the target echo, and for optical and infrared, the attribute information can be the target appearance characteristics. The state information of the tracking target at time k is set as It includes at least time, position and speed information, and may also include attribute information, but it must be consistent with the attribute information in the measurement information. The target position can be described in a local rectangular coordinate system, a geodetic rectangular coordinate system or a geographic coordinate system. For optical and infrared early warning detection equipment, it can also be described as a reference unified pixel coordinate, where the superscript T in the formula indicates transposition;

[0028] Step 1.3: Target Admission Method A The target is started based on the track start rule. A reasonable track start rule is that within n detection cycles, if the measurement point sequence of m different detection cycles {z O(1),I(1) ,z O(2),I(2) ,...,z O(m),I(m)}, satisfying the constraints of formula (1), formula (2) and formula (3), then the measurement point sequence is taken as the initial track of the target to be tracked, and the target initial state of the output target in the O(m)th detection cycle is calculated according to formula (4): The measurement point subscript O(i), I(i) indicates that the i-th measurement point in the sequence is the I(i)-th measurement obtained by the detection device in the O(i)-th detection cycle. At the same time, the detection cycle distribution range of m measurements cannot exceed n, that is, O(m)-O(1)≤n-1. Generally, n is set to 4 and m is set to 3. Based on the track start rule, manual target admission can be performed, that is, the first point z in the measurement point sequence can be manually determined. O(1),I(1) , other points are obtained through search, and can also be automatically entered, that is, in the entire area or a limited area, a measurement point sequence that meets the target starting principle is searched. During the search process, a dynamic programming method can be used to improve the search speed;

[0029]

[0030]

[0031]

[0032]

[0033] Where V min ,V max ,a max ,ω max They are the minimum speed threshold, maximum speed threshold, maximum acceleration threshold and maximum angle speed threshold.

[0034] Step 2: Trace the network T There are two inputs, one is the stable measurement within the wave gate, and the other is generated by the network input method f I The other is the target historical steady state, which is obtained by accumulating the steady state output of the tracking network with time information. The output of the tracking network is the target steady state, which is further converted by the network output conversion method f O Forward conversion can finally get the current target state, where stability means not including time accumulation, stable measurement and stable state mean not including pure time and pure position information, and network input generation method f I According to the measurement gate set, the stable measurement in the output gate is calculated by using the stable state output and target state of the tracking network at the previous moment, and the network output conversion method f O It includes forward conversion and reverse conversion. The forward conversion calculates the current target state according to the current target steady state and the target state at the previous moment. The reverse conversion calculates the current target steady state according to the target state at the previous moment and the current target state.

[0035] Step 2.1: Trace the network T There are two inputs, one is the stable measurement within the wave gate, and the other is generated by the network input method f I The other is the target historical steady state, which is obtained by accumulating the steady state output of the tracking network with time information. The output of the tracking network is the target steady state, which is further converted by the network output conversion method f O Forward conversion can finally get the current target state, where stability means not including time accumulation, stable measurement and stable state mean not including pure time and pure position information, and network input generation method f I According to the measurement gate set, the stable measurement in the output gate is calculated by using the stable state output and target state of the tracking network at the previous moment, and the network output conversion method f O It includes forward conversion and reverse conversion. The forward conversion calculates the current target state according to the current target steady state and the target state at the previous moment. The reverse conversion calculates the current target steady state according to the target state at the previous moment and the current target state.

[0036] Step 2.2: Trace the network T The input is measured steadily within the gate of the kth detection cycle. Can be set to It is composed of N gate-internal stable measurements, which is an N×5 matrix. Each gate-internal stable measurement can be set to It is a 1×5 size matrix, and the target historical steady state input can be set in The target steady-state output of the tracking network can be set to be 1×5M, which is the accumulation of M steady-state outputs with added time information. They represent the velocity estimation in the X direction, the velocity estimation in the Y direction, the acceleration estimation in the X direction, and the acceleration estimation in the Y direction, respectively, and are 1×4 matrices;

[0037] Step 2.3: Generate method f in the network input I In , a feasible measurement gate setting method is that the distance between the measurement point position and the target predicted position at the corresponding moment is less than the set threshold, that is, Where g is the set threshold, G(i) represents the measurement sequence number corresponding to the i-th measurement that satisfies the gate constraint, and x k,G(i) ,y k,G(i) is the position of the measurement point, The predicted position of the target is x, which is the position of the target state at the previous moment. k-1 and speed And the acceleration contained in the steady-state output of the tracking network Calculated, that is Δt k,G(i) is the difference between the time of the measurement point and the target state time at the previous moment, and is represented by Δt k,G(i) =t k,G(i) -t k-1 Calculated, then the stable measurement can be calculated by formula (5) Finally, based on the stable measurement in N wave gates, If the number of measurement points that meet the requirements is less than N, a 1×5 zero matrix can be used as a stationary measurement for filling. If the number of measurement points that meet the requirements is greater than N, select Δr k,G(i) Minimum N measurement points to calculate stable measurement;

[0038]

[0039] Step 2.4: Output the transformation method f in the network O In the example, a reasonable forward transformation to the target state x at time k is k Time t k By k-1 Add the average time difference measured in the wave gate to get, that is, Where N0 represents the number of non-zero stationary measurements, and the target state x at time k k The velocity component in is output by the target steady state The velocity component in is directly obtained, and the target state x at time k k The position component in is obtained by adding the previous position to the current velocity multiplied by time, that is, A reasonable inverse transformation to the target steady-state output The velocity component in is directly determined by the target state x at time k k The velocity in is given by the acceleration component, which is obtained by dividing the velocity difference between the target state at time k and time k-1 by the time difference, that is,

[0040] Step 2.5: Trace the network T The initial target history steady-state input in The initial state can be given according to the target admission method and the initial measurement point sequence {z O(1),I(1) ,z O(2),I(2) ,...,z O(m),I(m)}, a feasible setting method is to set In in other All zero vectors;

[0041] Step 3: Trace the network T It consists of a measurement feature extraction part, a state feature extraction part, a unified feature generation part, and a stable state estimation part. The measurement feature extraction part is designed with a multi-head self-attention mechanism, a feedforward network, and a residual connection. The input is a stable measurement within the wave gate, and the output is a measurement feature. The state feature extraction part is designed with a feedforward network and a residual connection. The input is the historical stable state of the target, and the output is the state feature. The unified feature generation part is designed with a multi-head attention mechanism, a feedforward network, and a residual connection. The input is the measurement feature and the state feature, and the output is a unified feature. The measurement feature is used to generate the K matrix and the V matrix in the multi-head attention mechanism, and the state feature is used to generate the Q matrix of the first multi-head attention mechanism. The stable state estimation is designed with a feedforward network, the input is a unified feature, and the output is the current stable state of the target.

[0042] Step 3.1: Trace the network TIt consists of a measurement feature extraction part, a state feature extraction part, a unified feature generation part, and a stable state estimation part. The measurement feature extraction part is designed with a multi-head self-attention mechanism, a feedforward network, and a residual connection. The input is a stable measurement within the wave gate, which contains information of multiple measurement points within the wave gate, and the output is a measurement feature. The state feature extraction part is designed with a feedforward network and a residual connection. The input is a historical stable state of the target with added time information, which is composed of the time difference of multiple time points in the past and the target stable state. The output is a state feature. The unified feature generation part is designed with a multi-head attention mechanism, a feedforward network, and a residual connection. The input is a measurement feature and a state feature, and the output is a unified feature. The measurement feature is used to generate the K matrix and the V matrix in the multi-head attention mechanism, and the state feature is used to generate the Q matrix of the first multi-head attention mechanism. The stable state estimation is designed with a feedforward network, the input is a unified feature, and the output is the current stable state of the target.

[0043] Step 3.2: In the process of tracking network construction, a typical multi-head attention mechanism calculation method is MultiHead(Q,K,V)=Concat(head1,…,head h )W O , where Q, K, and V represent the three input matrices of this layer, representing the query matrix, key matrix, and value matrix respectively. Concat(·) means that the matrices are concatenated along the first dimension, that is, different matrices are concatenated along the rows of the matrix. h represents the number of attention channels. head i Represents the result of the i-th attention channel, calculated by the attention formula Calculated, d K represents the column size of the K matrix, and all W matrices are neural network weights. Furthermore, a corresponding multi-head self-attention mechanism is MultiHead(X,X,X), that is, the Q, K, and V of the multi-head self-attention mechanism are the same, which are all the input X of the layer, that is, the multi-head self-attention mechanism has only one input matrix;

[0044] Step 3.3: During the tracking network construction process, a typical feed-forward network setting method is z O =FN(z I )=max(0,z I W1+b1)W2+b2, where FN(·) represents the feedforward network function, z O For its output, z I is its input, W and b are the parameters to be optimized, representing the network weight and bias term respectively, and max(·) is the maximum value function;

[0045] Step 3.4: In the process of tracking network construction, a typical residual connection is implemented by a cross-layer connection line and a summation layer, that is, z O =F(z I )+z I , where F(·) is a neural network layer, which can be a multi-head self-attention mechanism, a multi-head attention mechanism, or a feed-forward network, z I is the input of F(·), z O is the output of the residual connection, i.e. z I In addition to being processed by F(·), the result obtained by processing by F(·) is directly added to the result obtained by the summation layer through cross-layer connection lines;

[0046] Step 4: Trace the network T The loss function of is composed of the steady state estimation loss and the network weight penalty term, where the steady state estimation loss is represented by the regression loss function, which is calculated by the steady state estimation and the true steady state. The true steady state can be obtained by reverse conversion using the network output conversion method.

[0047] Step 4.1: Trace the network T The loss function of is composed of the steady state estimation loss and the network weight penalty term, where the steady state estimation loss is represented by the regression loss function, which is calculated by the steady state estimation and the true steady state. The true steady state can be obtained by reverse conversion using the network output conversion method.

[0048] Step 4.2: A reasonable way to express the tracking network loss function is in represents the estimated output of the steady state of the tracking network, y represents the corresponding true steady state, ||g||2 represents the L2 norm, that is, the sum of the squares of the elements of the matrix and then the square root, W represents the weight in the tracking network, and α is the regularization coefficient, which needs to be set;

[0049] Step 5: Collect the measurement data of the early warning detection equipment and the true state value of the target contained therein, and construct a tracking network training data set. Each sample includes three parts: the stable measurement input within the wave gate, the target historical stable state input, and the target stable state output. T The two inputs and one output of the gate correspond to each other, where the stable measurement input is based on the measurement data collected by the early warning detection device at a certain time k and the state true value of a certain target at the corresponding time. By calling the network input generation method f I It is calculated that the historical steady state input of the target is based on the state truth value of the corresponding target k at multiple moments before, and the network output conversion method f OThe target steady state output is obtained by reverse conversion based on the true value of the corresponding target state at time k, which is obtained by reverse conversion through the network output conversion method. The neural network optimization method is further used to optimize the tracking network f based on the tracking network training data set. T Train, optimize and debug until the tracking network performance meets the requirements, and finally output the trained tracking network f T ;

[0050] Step 5.1: Collect the measurement data of the early warning detection equipment and the true state value of the target contained therein, and construct a tracking network training data set. Each sample includes three parts: the stable measurement input within the wave gate, the target historical stable state input, and the target stable state output. T The two inputs and one output of the gate correspond to each other, where the stable measurement input is based on the measurement data collected by the early warning detection device at a certain time k and the state true value of a certain target at the corresponding time. By calling the network input generation method f I It is calculated that the historical steady state input of the target is based on the state truth value of the corresponding target k at multiple moments before, and the network output conversion method f O The target steady state output is obtained by reverse conversion based on the true value of the corresponding target state at time k, which is obtained by reverse conversion through the network output conversion method;

[0051] Step 5.2: Further adopt the neural network optimization method to optimize the tracking network f based on the tracking network training data set. T Perform training, optimization, and network debugging until the tracking network performance meets the requirements, and finally output the trained tracking network f T , the optimization methods that can be used include gradient descent, stochastic gradient descent, Momentum, RMSProp, Adam, but are not limited to these;

[0052] Step 5.3: When constructing the tracking network training data set, the diversity of samples should be ensured, that is, various performance influencing factors of early warning detection equipment and various types of targets to be tracked should be considered, and detection data of various types of tracking targets and the true value of the target status should be collected under various influencing factors;

[0053] Step 5.4: When training the tracking network, the sample retention test method, leave-one-out cross-validation method, and k-fold cross-validation method can be used to divide the tracking network training data set into a training set and a test set to ensure that the final trained tracking network has good generalization ability;

[0054] Step 6: Deploy an automatic target tracking method in the early warning detection device, receive the measurement data generated by real-time detection, and adopt the target admission method f AStart the track and get the initial state information of the target. For each starting target, start the trained tracking network f under the limit of the number of tracking channels. T , real-time tracking is performed until the tracking is cancelled or the target is lost;

[0055] Step 6.1: Deploy an automatic target tracking method in the early warning detection device, receive the measurement data generated by real-time detection, and adopt the target acquisition method f A Start the track and get the initial state information of the target. For each starting target, start the trained tracking network f under the limit of the number of tracking channels. T , real-time tracking is performed until the tracking is cancelled or the target is lost;

[0056] Step 6.2: When the target admission is automatic, the target admission method continuously performs target initiation and starts the trained tracking network f T , until the maximum tracking channel limit is reached. At this time, the automatic target tracking method will give a prompt, and the operator can delete the target being tracked, release the tracking channel for the new target to track, or ignore the new starting target and keep tracking the current target. According to IMO regulations, a simple target tracking device should have at least 20 to 30 tracking channels, and a complete device should have at least 40 tracking channels;

[0057] Step 6.3: When the operator cancels the target tracking or when the target is lost, the automatic target tracking method deletes the tracking network and releases the tracking channel. One reasonable target loss judgment principle is that there is no measurement at the gate for m consecutive detection cycles, that is, f is used. I The calculated gate-internal stable measurement input is an all-zero matrix, and m is generally set to 3;

[0058] The embodiment of the present invention also provides an automatic target tracking device, such as Figure 3 As shown, the device comprises:

[0059] The target admission module is used to obtain the initial state of the target to be tracked, and then start the tracking network to continuously track the target. It can be started manually, that is, the target to be tracked is started at the operator's choice, or it can automatically start tracking the target in the entire detection area or in a limited area. One tracking network can track one target. For multiple targets started, the target admission module can start the corresponding number of tracking networks to track multiple targets;

[0060] Tracking network module, the main body of target tracking, consists of measurement feature extraction part, state feature extraction part, unified feature generation part, and steady state estimation part. It is started by the target acquisition module and automatically processes the measurement obtained by the detection device according to the target initial state input by the target acquisition module to achieve continuous tracking of the target;

[0061] An auxiliary computing module, including a network input generation method and a network output conversion method, is used to assist the tracking network in generating the required input and converting the output during the construction of the tracking network training data set and during the use of the tracking network;

[0062] The network training module includes two parts: tracking network training data set construction and tracking network training optimization, which are used to train, optimize and debug the tracking network.

[0063] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0064] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. An automatic target tracking method, characterized in that: The following steps are involved: Step 1: Automatic target tracking method by target admission method f A , Tracking Network T , network input generation method f I and the network output conversion method f O Joint composition, where the target admission method f A Using the measurement information at multiple times, the initial state information of the tracking target is obtained through the track initiation, and the tracking network is started in different tracking channels for the target to be tracked. T Then, the current measurement information and the target historical state information are used to generate the method f through network input I , get the tracking network f T Input, processed by the network, get the network output, and then pass the network output conversion method f O Forward conversion, to obtain the real-time status information of the tracking target, and then to achieve continuous tracking of the target, the measurement information at different times at least includes time and position information, and the status information of the tracking target at least includes time, position and speed information; Step 2: Trace the network T There are two inputs, one is the stable measurement within the wave gate, and the other is generated by the network input method f I The other is the target historical steady state, which is obtained by accumulating the steady state output of the tracking network with time information. The output of the tracking network is the target steady state, which is converted by the network output conversion method f O Forward conversion, finally get the current stable state of the target, where stable means not including time accumulation, stable measurement and stable state means not including pure time and pure position information, network input generation method f I According to the measurement gate set, the stable measurement in the output gate is calculated by using the stable state output and target state of the tracking network at the previous moment, and the network output conversion method f O It includes forward conversion and reverse conversion. The forward conversion calculates the current target state according to the current target stable state and the target state at the previous moment. The reverse conversion calculates the current stable state of the target according to the target state at the previous moment and the current target state. Step 3: Trace the network T It consists of a measurement feature extraction part, a state feature extraction part, a unified feature generation part, and a stable state estimation part. The measurement feature extraction part is designed with a multi-head self-attention mechanism, a feedforward network, and a residual connection. The input is a stable measurement within the wave gate, and the output is a measurement feature. The state feature extraction part is designed with a feedforward network and a residual connection. The input is the historical stable state of the target, and the output is the state feature. The unified feature generation part is designed with a multi-head attention mechanism, a feedforward network, and a residual connection. The input is the measurement feature and the state feature, and the output is a unified feature. The measurement feature is used to generate the K matrix and the V matrix in the multi-head attention mechanism, and the state feature is used to generate the Q matrix of the first multi-head attention mechanism. The stable state estimation is designed with a feedforward network, the input is a unified feature, and the output is the current stable state of the target. Step 4: Trace the network T The loss function of is composed of the steady state estimation loss and the network weight penalty term, where the steady state estimation loss is represented by the regression loss function, which is calculated by the steady state estimation and the true steady state. The true steady state is obtained by reverse conversion using the network output conversion method. Step 5: Collect the measurement data of the early warning detection equipment and the true state value of the target contained therein, and construct a tracking network training data set. Each sample includes three parts: the stable measurement input within the wave gate, the target historical stable state input, and the target stable state output. T The two inputs and one output of the gate correspond to each other, where the stable measurement input is based on the measurement data collected by the early warning detection device at a certain time k and the state true value of a certain target at the corresponding time. By calling the network input generation method f I It is calculated that the historical steady state input of the target is based on the state truth value of the corresponding target k at multiple moments before, and the network output conversion method f O The target steady state output is obtained by reverse conversion based on the true value of the corresponding target state at time k, which is obtained by reverse conversion through the network output conversion method. The neural network optimization method is used based on the tracking network training data set to optimize the tracking network f T Train, optimize and debug until the tracking network performance meets the requirements, and finally output the trained tracking network f T ; Step 6: Deploy an automatic target tracking method in the early warning detection device, receive the measurement data generated by real-time detection, and adopt the target admission method f A Start the track and get the initial state information of the target. For each starting target, start the trained tracking network f under the limit of the number of tracking channels. T , real-time tracking is performed until tracking is cancelled or the target is lost.

2. An automatic target tracking method as claimed in claim 1, characterized in that: In step 1, target admission method f A The target is started based on the track start rule. One track start rule is that within n detection cycles, if the measurement point sequence {z O(1),I(1) ,z O(2),I(2) ,...,z O(m),I(m) }, satisfying the constraints of formula (1), formula (2) and formula (3), then the measurement point sequence is taken as the initial track of the target to be tracked, and the target initial state of the output target in the O(m)th detection cycle is calculated according to formula (4): The measurement point subscripts O(i), I(i) indicate that the i-th measurement point in the sequence is the I(i)-th measurement obtained by the detection device in the O(i)-th detection cycle. At the same time, the detection cycle distribution range of m measurements cannot exceed n, that is, O(m)-O(1)≤n-1, n is set to 4, m is set to 3, and based on the track start rule, manual target admission is performed, that is, the first point z in the measurement point sequence is manually determined. O(1),I(1) , other points are obtained by searching or automatically entered, that is, in the whole area or in a limited area, a measurement point sequence that meets the target starting principle is searched, and the dynamic programming method is used to improve the search speed during the search process; Where V min ,V max ,a max ,ω max They are the minimum speed threshold, maximum speed threshold, maximum acceleration threshold and maximum angle speed threshold.

3. An automatic target tracking method as claimed in claim 2, characterized in that: In step 2, trace the network f T The input is measured steadily within the gate of the kth detection cycle. Set to It is composed of N gate-internal stationary measurements, which is an N×5 matrix. The internal stationary measurement of each gate is set to It is a 1×5 size matrix, and the target historical stable state input setting in The result is obtained by accumulating M steady-state outputs with added time information, which is a 1×5M matrix. The target steady-state output of the tracking network is set to They represent the X-direction velocity estimation, Y-direction velocity estimation, X-direction acceleration estimation, and Y-direction acceleration estimation, respectively, and are 1×4 matrices.

4. An automatic target tracking method as claimed in claim 3, characterized in that: In step 2, the network input generation method f I One of the measurement gate setting methods is that the distance between the measurement point position and the target predicted position at the corresponding moment is less than the set threshold, that is, Where g is the set threshold, G(i) represents the measurement sequence number corresponding to the i-th measurement that satisfies the gate constraint, and x k,G(i) ,y k,G(i) is the position of the measurement point, The predicted position of the target is x, which is the position of the target state at the previous moment. k-1 and speed And the acceleration contained in the steady-state output of the tracking network Calculated, that is Δt k,G(i) is the difference between the time of the measurement point and the target state time at the previous moment, and is represented by Δt k,G(i) =t k,G(i) -t k-1 Calculated, and then the stable measurement is calculated by formula (5) Finally, based on the stable measurement in N wave gates, If the number of measurement points that meet the requirements is less than N, a 1×5 zero matrix is ​​used as a stationary measurement for filling. If the number of measurement points that meet the requirements is greater than N, Δr k,G(i) Minimum N measurement points to calculate stable measurement; 5. An automatic target tracking method as claimed in claim 4, characterized in that: In step 2 the network output conversion method f O One of the forward transformations is the target state x at time k k Time t k By k-1 Add the average time difference measured in the wave gate to get, that is, Where N0 represents the number of non-zero stationary measurements, and the target state x at time k k The velocity component in is output by the target steady state The velocity component in is directly obtained, and the target state x at time k k The position component in is obtained by adding the previous position to the current velocity multiplied by time, that is, A reverse transformation to target steady-state output The velocity component in is directly determined by the target state x at time k k The velocity in is given by the acceleration component, which is obtained by dividing the velocity difference between the target state at time k and time k-1 by the time difference, that is, 6. An automatic target tracking method as claimed in claim 5, characterized in that: In step 2, trace the network f T The initial target history steady-state input in Initial state given by the target admission method and the initial measurement point sequence {z O(1),I(1) ,z O(2),I(2) ,...,z O(m),I(m) }Get, one setting method is to set In in ,other i=1:M-1 are all zero vectors.

7. An automatic target tracking method as claimed in claim 1, characterized in that: In step 3, a multi-head attention mechanism calculation method is MultiHead(Q,K,V)=Concat(head1,…,head h )W O , where Q, K, and V represent the three input matrices of this layer, representing the query matrix, key matrix, and value matrix respectively. Concat(·) means that the matrices are concatenated along the first dimension, that is, different matrices are concatenated along the rows of the matrix. h represents the number of attention channels. head i Represents the result of the i-th attention channel, calculated by the attention formula head i (Q,K,V)=Attention(QW i Q ,KW i K ,VW i V ) is calculated, d K represents the column size of the K matrix, and all W matrices are the weights of the neural network. Furthermore, a corresponding multi-head self-attention mechanism is MultiHead(X,X,X), that is, the Q, K, and V of the multi-head self-attention mechanism are the same, which are all the input X of this layer, that is, the multi-head self-attention mechanism has only one input matrix.

8. An automatic target tracking method as claimed in claim 1, characterized in that: In step 3, a feedforward network setting method is z O =FN(z I )=max(0,z I W1+b1)W2+b2, where FN(·) represents the feedforward network function, z O For its output, z I is its input, W and b are the network parameters to be optimized, representing the network weight and bias term respectively, and max(·) is the maximum value function.

9. An automatic target tracking method as claimed in claim 1, characterized in that: In step 4, a tracking network loss function is expressed as in represents the estimated output of the steady state of the tracking network, y represents the corresponding true steady state, ||·||2 represents the L2 norm, that is, the square root of the sum of the squares of the elements of the matrix, W represents the weight in the tracking network, and α is the regularization coefficient, which needs to be set.