Distributed three-dimensional extended target tracking method based on feature selection

Abstract

The invention discloses a distributed three-dimensional extended target tracking method based on feature selection, and the method comprises the following steps: obtaining the original measurement of each sensor at a target k moment, carrying out the preprocessing of the original measurement, and obtaining the preprocessed original measurement, the preprocessed original measurement is substituted into a GIW-PHD filtering algorithm to obtain estimation state information of a detection target of each sensor, the estimation state information of the detection target is preprocessed to obtain a target state of each sensor, and the target state is selected for feature extraction and fusion to obtain target state information at the moment k. According to the method, feature selection and fusion can be carried out on the target state of the sensor while the data filtering operation time is shortened, and the position information of the three-dimensional expansion target and the estimation precision of the three-dimensional expansion information are improved through feature splicing and fusion.

Description

Technical Field

[0001] The invention belongs to the technical field of target tracking, and in particular relates to a distributed three-dimensional extended target tracking method based on feature selection. Background Art

[0002] Three-dimensional target tracking is widely used in fields such as autonomous driving and robot navigation. Common three-dimensional target tracking methods include vision-based methods, model-based methods, multi-sensor based methods, etc.; with the increase of sensor resolution, the shape and size of the three-dimensional target can no longer be ignored, and it becomes a three-dimensional extended target with size and shape.

[0003] There are two types of three-dimensional extended target tracking methods based on multiple sensors: centralized tracking method and distributed method. Among them, centralized tracking methods, such as the paper "GIW-PHD Tracking Algorithm for Extended Targets Based on Multi-Sensor Information Fusion", ensure the accuracy of state information estimation of three-dimensional extended targets by fusing the measurement data of multiple sensors; however, since the centralized fusion algorithm has more measurement data for three-dimensional extended targets, the transmission data is larger, and its robustness and scalability are not strong.

[0004] Distributed tracking methods, such as the patent document "A Distributed Multi-Sensor Fusion Tracking Method and System" with publication number CN116028890A, obtain the probability posterior density through a local extended label multi-Bernoulli filter, fuse the local posterior probability density of multiple sensors based on the relative weight matrix, and update the target state after obtaining the fused probability posterior density, and finally determine the distributed multi-sensor fusion tracking result; this method first performs filtering and then fuses, which reduces the size of the transmitted data and can improve the robustness and scalability of the network without the need for a specific fusion center. However, it does not take into account that the tracking of three-dimensional extended targets under multiple sensors has different irradiation surfaces for the target and the obtained measurement data is different. Therefore, this method is not accurate enough in estimating the state tracking of three-dimensional extended targets. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention proposes a distributed three-dimensional extended target tracking method based on feature selection.

[0006] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0007] The present invention is a distributed three-dimensional extended target tracking method based on feature selection, comprising the following steps:

[0008] Measurement data acquisition step: obtaining the original measurement data of each sensor detecting the target at time k;

[0009] Measurement data preprocessing step: preprocessing the original measurement data of each sensor to obtain the original measurement data after preprocessing;

[0010] Estimated state information acquisition step: the original measurement data of each sensor after preprocessing is respectively brought into the GIW-PHD filtering algorithm to obtain the estimated state information of the target detected by each sensor;

[0011] Estimated state information preprocessing step: preprocessing the estimated state information of the target detected by each sensor to obtain the target state of each sensor;

[0012] Target state information acquisition step: Select the target state of each sensor for feature extraction and fusion to obtain the fused target state information at time k.

[0013] Furthermore, the original measurement data of each sensor consists of a target measurement source and Gaussian white noise.

[0014] Furthermore, the measurement data preprocessing step includes:

[0015] Initializing the original measurement data of each sensor respectively to obtain the original measurement data of each sensor after initialization;

[0016] The original measurement data of each sensor after initialization is projected onto the qz plane corresponding to each sensor to obtain the original measurement data of each sensor after preprocessing.

[0017] Furthermore, the measurement data preprocessing step further includes:

[0018] Project the original measurement data of each sensor after initialization onto the xy plane to obtain the initial measurement projection of each sensor;

[0019] Using the initial measurement projections of each sensor, respectively determine the average of the initial measurement projections of each sensor and the yaw angle of each sensor when detecting the target;

[0020] The straight line q corresponding to each sensor is determined according to the average value of the initial measurement projection of each sensor and the yaw angle when each sensor detects the target, and the qz plane corresponding to each sensor is determined according to the straight line q and the z-axis corresponding to each sensor.

[0021] Furthermore, the yaw angle of each sensor when detecting the target is determined by using the initial measurement projection of each sensor, including:

[0022] Centralizing the initial measurement projections of each sensor to obtain the centralized initial measurement projections of each sensor;

[0023] The covariance matrix of the initial measurement projection after centering of each sensor is calculated respectively. The angle between the eigenvector corresponding to the covariance matrix and the x-axis is the yaw angle of each sensor when detecting the target.

[0024] Furthermore, the step of obtaining estimated state information includes:

[0025] The state and covariance of the target at time k-1 are brought into the GIW-PHD filtering algorithm for prediction to obtain the predicted GIW component;

[0026] The original measurement data after preprocessing of each sensor are respectively brought into the GIW-PHD filtering algorithm to update the predicted GIW component to obtain the updated GIW component of each sensor;

[0027] According to the set pruning threshold, the merging threshold between components and the maximum number of GIW components, the updated GIW components of each sensor are pruned and merged to obtain the new GIW components of each sensor;

[0028] The GIW components with weights greater than 0.5 in the new GIW components of each sensor are selected to obtain the estimated state information of the target detected by each sensor.

[0029] Furthermore, the estimated state information preprocessing step includes:

[0030] According to the estimated state information of the target detected by each sensor, the ellipse estimated state of the extended target corresponding to each sensor is obtained respectively;

[0031] The ellipse estimation state of the extended target corresponding to each sensor is respectively combined with the yaw angle when each sensor detects the target to obtain the target state of each sensor.

[0032] Furthermore, the target state information acquisition step includes:

[0033] The target states of the two sensors with the largest long axis difference among the target states of each sensor are selected for feature extraction and fusion to obtain the extended state information of the fused target;

[0034] The extended state information of the fused target is used to fuse the target center position and obtain the fused target state information at time k.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention estimates the yaw angle by acquiring the measurement data of each sensor used to detect the target, and at the same time centers the measurement data, constructs a new two-dimensional coordinate system through the yaw angle, the centralized measurement data and the z-axis, and projects the measurement data to the two-dimensional coordinate system. By reducing the dimension of the original measurement data of the sensor, the time for data calculation in the GIW-PHD filter is greatly shortened.

[0037] 2. The present invention obtains estimated state information of the target detected by the sensor through filtering and screening, obtains the target state of each sensor in combination with the yaw angle when each sensor detects the target, selects the target state of the two groups of sensors with the largest long-axis difference for feature selection and fusion, and utilizes feature splicing and fusion to improve the position information of the three-dimensional extended target and the estimation accuracy of the three-dimensional extended information.

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is an overall flow chart of the method of the present invention.

[0040] Figure 2 It is a comparison diagram of the target center position tracking results in the x, y, and z directions of the method of the present invention and the actual target trajectory.

[0041] Figure 3 It is a comparison diagram of the root mean square error of the target position obtained by the method of the present invention and other radars after 100 Turkaro tests.

[0042] Figure 4 It is a comparison diagram of the root mean square error after 100 Turkaro tests of the target expansion state obtained by the method of the present invention and other radars.

[0043] Figure 5 It is a comparison diagram between the target expansion state estimated by the method of the present invention and the actual target expansion state at a certain moment. DETAILED DESCRIPTION

[0044] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the scheme according to the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0045] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0046] 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 are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0047] like Figure 1 As shown, this embodiment provides a distributed three-dimensional extended target tracking method based on feature selection, which specifically includes the following steps:

[0048] Step 1: Obtain the original measurement data of each sensor that detects the target at time k, where the original measurement data of the sensor consists of the target measurement source and Gaussian white noise.

[0049] The original measurements of each sensor of the target at time k are:

[0050] z k =ζ k +v

[0051] Among them, z k is the original measurement data of the sensor, v is the zero-mean Gaussian white noise, ζ k is the target measurement source at time k, and the target measurement source ζ can be expressed as:

[0052] ζ=s·A·U 1 ·U 2 ·U 3

[0053] A=[a·sinβcosγ,b·sinβsinγ,c·cosβ]

[0054]

[0055] Where A is the standard parametric equation vector of the ellipsoid, and β∈(0,π), γ∈(0,2π); U 1 , U 2 , U 3 are the rotation moments corresponding to the x, y, and z axes respectively; w 1 is the roll angle, w 2 is the pitch angle, w 3 is the yaw angle, a is the major axis, b is the minor axis, and c is the median axis; s∈(0,1) is the scaling factor; β, γ and s are random variables.

[0056] Step 2: preprocess the original measurement data of the sensor to obtain the original measurement data after preprocessing.

[0057] Step 2.1: Initialize the original measurement of each sensor to obtain the original measurement data of each sensor after initialization. Since each sensor detects one side of the target, only part of the measurement in the target measurement model is obtained. Taking the lth sensor as an example, the original measurement data detected by the lth sensor at time k after initialization is:

[0058]

[0059] in, represents the i-th measurement produced by the l-th sensor, x i represents the coordinate of the ith measurement generated by the lth sensor on the x-axis, y i shows the coordinate of the ith measurement generated by the lth sensor on the y-axis, z i : represents the coordinate of the i-th measurement generated by the l-th sensor on the z-axis, and i is the measurement number generated by the l-th sensor at time k.

[0060] Step 2.2: Project the original measurement data of the first sensor after initialization onto the xy plane to obtain the initial measurement projection of the first sensor. Specifically, project the original measurement data after initialization onto the xy plane to obtain the initial measurement projection of the first sensor. In The measurement is projected into the xy plane and is expressed as:

[0061]

[0062] in, is the initial measurement projection of the lth sensor, is the i-th measurement produced by the l-th sensor projected onto the xy-axis plane.

[0063] Step 2.3: Use the initial measurement projection of the lth sensor to determine the average of the initial measurement projection of the lth sensor and the yaw angle when the lth sensor detects the target.

[0064] Step 2.3.1: Centralize the initial measurement projection of the lth sensor to obtain the initial measurement projection of the lth sensor after centralization:

[0065]

[0066] in, is the initial measurement projection of the lth sensor The mean of .

[0067] Step 2.3.2: Calculate the covariance matrix of the first measurement projection after the lth sensor is centered Covariance matrix The corresponding eigenvector is r z,max , a 2×1 dimensional vector r z,max The angle with the x-axis is the yaw angle when the lth sensor detects the target

[0068] Step 2.4: Determine the straight line q according to the mean value of the initial measurement projection of the lth sensor and the yaw angle of the lth sensor when detecting the target. The straight line q and the z-axis determine the qz plane. Project the original measurement data of the lth sensor after initialization onto the qz plane to obtain the original measurement data of the lth sensor after preprocessing. Specifically:

[0069] Will pass The angle with the x-axis is The straight line q and the z-axis construct a new plane qz and measure In The measurement projection qz is put into the plane, and the original measurement data after preprocessing of the lth sensor can be expressed as:

[0070]

[0071] Among them, q i represents the coordinate of the original measurement data on the q axis after preprocessing, z i Represents the coordinate of the original measurement data on the z-axis after preprocessing, The calculation formula of parameter θ is: Among them, x i represents the coordinate of the ith measurement generated by the lth sensor on the x-axis, y i shows the coordinate of the ith measurement generated by the lth sensor on the y-axis.

[0072] Step 3: Submit the preprocessed raw measurement data into the GIW-PHD filtering algorithm to obtain the estimated state information of the sensor detecting the target.

[0073] Specifically, the original measurement data of the lth sensor after preprocessing in step 2, that is, the measurement data at time k in the lth sensor Substitute the GIW-PHD filtering algorithm to obtain the estimated state information of the target detected by the lth sensor

[0074] Step 3.1: Substitute the state and covariance of the target at time k-1 into the GIW-PHD filtering algorithm for prediction to obtain the predicted GIW component.

[0075] According to the state and covariance of the target at time k-1, the predicted PHD can be expressed as:

[0076] D k|k-1 (ξ k )=D S,k|k-1 (ξ k-1 )+γ k (ξ k )

[0077] Among them, D S,k|k-1 (ξ k-1 ) is the predicted GIW-PHD of the survival target at time k, γ k (ξ k ) is the GIW-PHD of the new target at time k, ξ k =(x k ,X k ) represents the state of the target, x k is the motion state vector of the target’s center of mass, X k The extended state of the target.

[0078] The predicted GIW–PHD for the survival target is:

[0079]

[0080] Among them, p S,k represents the target survival probability, represents the weight component of the i-th GIW component at time k-1, N(x;m,P) represents the mean m, the covariance matrix is ​​P Gaussian distribution, IW(X;v,V) represents the inverse Wishart distribution with v inverse Wishart degrees of freedom and V inverse Wishart scaling matrix, and the parameters in the formula are expressed as:

[0081]

[0082] Among them, F k-1 is the state transfer matrix, Q k-1 is the process noise, T s represents the sampling time, τ is the time attenuation factor, d is the target expansion dimension, They are the motion state, covariance matrix, inverse Wishart degrees of freedom, and inverse Wishart scaling matrix of the surviving target respectively.

[0083] The predicted GIW–PHD for the new target set is:

[0084]

[0085] in, They are the motion state, covariance matrix, inverse Wishart degrees of freedom, and inverse Wishart scaling matrix of the new target respectively.

[0086] Step 3.2: The original measurement data after preprocessing of the lth sensor in step 2 is brought into the GIW-PHD filtering algorithm to update the predicted GIW component to obtain the updated GIW component of the lth sensor. The updated GIW-PHD is expressed as:

[0087]

[0088] in, It is expressed as the missed detection part GIW-PHD, is to update ξ using the measurement subset W k GIW-PHD after, p∠Z k It means that p will measure the set Z k Partition into non-empty unit set W; It can be expressed as:

[0089]

[0090] in, ξ k The update formula is expressed as:

[0091]

[0092] ξ k =ξ k|k-1

[0093] in, ξ k are respectively represented as the weight and GIW component parameter of the jth GIW component, γ(j) is the measurement rate of the jth component, and p D is the probability of detecting the measurement.

[0094] It can be expressed as:

[0095]

[0096] In the formula, the update parameter is expressed as:

[0097]

[0098] in, are the motion state, covariance matrix, and inverse Wishart scaling matrix of the jth GIW component newly completed using the measurement subset W; are the mean and measurement covariance matrix of the measurement subset W, respectively. They are respectively the innovation factor, updated gain matrix, central measurement innovation, and innovation matrix obtained by updating the j-th GIW component using the measurement subset W.

[0099] The likelihood function is:

[0100]

[0101] Where |V| represents the determinant of the matrix, |W| represents the number of measurements in the measurement subset W, is the likelihood of the j-th GIW component after being updated using the measurement subset W.

[0102] The updated weights are:

[0103]

[0104] Among them, β FA,k is the number of clutter generated in the scanning unit survival space, represents the jth weight updated by the measurement subset W, is the likelihood of the jth GIW component after the measurement subset W is updated, p D Indicates the probability of the target being detected:

[0105]

[0106] Among them, δ i,j is the Kronecker delta function, w p It is expressed as the partition weight.

[0107] Step 3.3: According to the set pruning threshold Tr, the merging threshold U between components and the maximum number J of GIW components max The updated GIW component of the lth sensor is trimmed and merged to obtain the new GIW component of the lth sensor;

[0108] Specifically, the pruning threshold is set to Tr, the merging threshold between components is set to U, and the maximum number of GIW components is set to J. max ; Discard the GIW components whose weights are less than the pruning threshold Tr, and merge the GIW components whose Euclidean distances are less than U. If the number of GIW components is greater than J max , select the final GIW component before J max The item with the largest weight is pruned and merged in the domain to obtain the new GIW component of the lth sensor after merging.

[0109] Step 3.4: Select the weight of the new GIW component of the lth sensor The GIW component of the sensor is used to obtain the estimated state information of the target detected by the lth sensor. It is expressed as:

[0110]

[0111] in, They represent the estimated value of the motion state and the estimated value of the extended information of the jth target at time k, respectively. J represents the total number of estimated targets. is the inverse Wishart degrees of freedom of the j-th objective, and d is the objective expansion dimension.

[0112] Step 4: preprocess the estimated state information of the target detected by the lth sensor to obtain the target state of the lth sensor.

[0113] Step 4.1: Based on the estimated state information of the target detected by the first sensor Get the ellipse estimated state of the extended target corresponding to the lth sensor at time k.

[0114] Specifically, by formula X k =R k diag([a 2 ,b 2 ])R k T , where a, b are the major and minor axes, R k is the rotation matrix, and the ellipse estimation state of the extended target at time k is obtained in, is the major axis of the elliptical shape of the jth target, is the minor axis of the elliptical shape of the jth target, is the angle between the jth target and the coordinate axis q.

[0115] Step 4.2: Expand the target ellipse estimation state by k time The yaw angle when the lth sensor detects the target in step 2 Correspondingly, in this embodiment, taking the lth sensor detecting the jth target as an example, the target state of the lth sensor is obtained for:

[0116]

[0117] in, is the major axis of the elliptical shape of the jth target in the lth sensor, is the minor axis of the elliptical shape of the jth target in the lth sensor, is the angle between the jth target in the lth sensor and the coordinate axis q, is the yaw angle of the jth target in the lth sensor, is the position information of the jth target in the lth sensor, is the position information of the jth target in the lth sensor on the coordinate axis q, is the position information of the jth target in the lth sensor on the coordinate axis z, and the position information is in the qz plane.

[0118] By Location information can be Converted into the xyz three-dimensional coordinate system, that is in, is the position information of the jth target in the lth sensor on the coordinate axis x, is the position information of the jth target in the lth sensor on the coordinate axis y, is the position information of the jth target in the lth sensor on the coordinate axis z.

[0119] Step 5: Select the target state of the sensor for feature extraction and fusion to obtain the target state information at time k.

[0120] Step 5.1. Perform the above steps 1 to 4 in parallel at the same time to obtain the target state of all sensors that detect the target. In this embodiment, each sensor obtains the state of the j-th target respectively, and selects the target states of the two sensors 1 and 2 with the largest long axis difference among each sensor for feature extraction and fusion.

[0121] Sensor 1 is represented as the longer sensor of the two sensors, and its target state is represented as:

[0122]

[0123] in, is the major axis of the jth extended target in sensor 1, is the minor axis of the extended target j in sensor 1, To expand the angle between target j and the xy plane in sensor 1, To expand the yaw angle of target j in sensor 1, is the location information of extended target j in sensor 1.

[0124] Sensor 2 is represented as the shorter long-axis sensor of the two sensors, and its target state is represented as:

[0125]

[0126] in, is the major axis of the extended target j in sensor 2, is the minor axis of the extended target j in sensor 2, To expand the angle between target j and the xy plane in sensor 2, is the angle between the yz planes of the extended target j in the two sensors, is the location information of extended target j in sensor 2.

[0127] Feature extraction is performed on sensor 1 and sensor 2. The long axis of the extended target j in sensor 1 Long axis as a target for fusion extension The major axis of the extended target j in sensor 2 Short axis as a target for fusion extension Fuse the short axes of the extended target j from the two sensors The central axis as the fusion target Expand target j in sensor 1 and the extended target j in sensor 2 As the fusion target pitch angle and roll angle Expand target j in sensor 1 As the yaw angle of the fused target, the extended state information of the fused target is obtained.

[0128] The extended state information formula of the fusion target is expressed as follows:

[0129] Long axis Short Axis Center axis Yaw angle Pitch Angle Roll Angle

[0130] Step 5.2: Use the state extension information of the fused target to fuse the target center position and obtain the fused target state information at time k.

[0131] Specifically, the location information of the extended target j in sensor 1 is and a yaw angle of 90° Construct a space line λ parallel to the xy plane 1 ,λ 1 The parametric equation is expressed as:

[0132]

[0133] The location information of the extended target j in sensor 2 and a yaw angle of 90° Construct a space line λ parallel to the xy plane 2 ,λ 2 The parametric equation is expressed as:

[0134]

[0135] The z-direction values ​​of the two spatial lines are merged into Combine two space lines λ1 , 2 The parameter expression of is solved, and the fusion target state information at time k is expressed as:

[0136] m k ={x k ,y k ,z k}

[0137] Among them, x k is the coordinate of the fusion target position information on the x-axis at time k, and y k is the coordinate of the fused target position information at time k on the y-axis, z k is the coordinate of the fused target position information on the z-axis at time k.

[0138] The present invention obtains sensor measurement data to estimate the yaw angle, centers the measurement data, constructs a new two-dimensional coordinate system using the yaw angle, the centralized measurement data and the z-axis, and projects the measurement data onto the two-dimensional coordinate system to reduce the dimension of the data, greatly shortening the time for data calculation in the GIW-PHD filter; at the same time, the estimated state information of the target detected by the sensor is obtained through GIW-PHD filtering, the estimated state information of the target is combined with the yaw angle when the sensor detects the target to obtain the target state of the sensor, the target states of the two groups of sensors with the largest long-axis difference are selected for feature selection and fusion, and the splicing and fusion of features are used to improve the position information of the three-dimensional extended target and the estimation accuracy of the three-dimensional extended information.

[0139] Figure 2 This is a comparison chart of the target center position tracking results in the x, y, and z directions and the real target trajectory of the method of the present invention. The target moves at a constant speed in the three directions. The circled line in the figure is the estimated result, and the solid line is the real trajectory of the target. It can be seen that this method can accurately estimate the center position of the three-dimensional extended target.

[0140] Figure 3 This is a comparison chart of the root mean square error of the target position obtained by the method of the present invention and other radars after 100 Turkaro tests, showing the root mean square error of the two radars' individual filtering positions and the root mean square error of the distributed fusion through the method of the present invention. It can be clearly seen that the method of the present invention greatly improves the accuracy of position estimation.

[0141] Figure 4 This is a comparison chart of the root mean square error of the target extended state obtained by the method of the present invention and other radars after 100 Turkaro tests, showing the root mean square error of the extended state filtered by two radars separately and the root mean square error of the extended state obtained by distributed fusion through the method of the present invention. It can be seen that the method of the present invention greatly improves the estimation accuracy of the extended state.

[0142] Figure 5 This is a comparison diagram of the target expansion state estimated by the method of the present invention and the real target expansion state at a certain moment, when the real target has a yaw angle of 45°, a roll angle of 0°, and a pitch angle of 0°. It can be seen that the method of the present invention can accurately estimate the target expansion state.

[0143] Finally, it should be noted that the above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A distributed three-dimensional extended target tracking method based on feature selection, characterized in that: The following steps are involved: Measurement data acquisition step: obtaining the original measurement data of each sensor detecting the target at time k; The measurement data preprocessing step: preprocessing the original measurement data of each sensor to obtain the original measurement data after preprocessing; Estimated state information acquisition step: the original measurement data of each sensor after preprocessing is respectively brought into the GIW-PHD filtering algorithm to obtain the estimated state information of the target detected by each sensor; Estimated state information preprocessing step: preprocessing the estimated state information of the target detected by each sensor to obtain the target state of each sensor; Target state information acquisition step: Select the target state of each sensor to perform feature extraction and fusion to obtain the fused target state information at time k.

2. The distributed three-dimensional extended target tracking method based on feature selection according to claim 1 is characterized in that: The original measurement data of each sensor consists of a target measurement source and Gaussian white noise.

3. The distributed three-dimensional extended target tracking method based on feature selection according to claim 1 is characterized in that: The measurement data preprocessing step includes: Initializing the original measurement data of each sensor to obtain the original measurement data of each sensor after initialization; The original measurement data of each sensor after initialization is projected onto the qz plane corresponding to each sensor to obtain the original measurement data of each sensor after preprocessing.

4. The distributed three-dimensional extended target tracking method based on feature selection according to claim 3 is characterized in that: The measurement data preprocessing step also includes: Projecting the original measurement data of each sensor after initialization onto the xy plane to obtain the initial measurement projection of each sensor; Using the initial measurement projections of the sensors, respectively determine the average of the initial measurement projections of the sensors and the yaw angles of the sensors when detecting the target; The straight line q corresponding to each sensor is determined according to the average value of the initial measurement projection of each sensor and the yaw angle when each sensor detects the target, and the qz plane corresponding to each sensor is determined according to the straight line q and the z-axis corresponding to each sensor.

5. The distributed three-dimensional extended target tracking method based on feature selection according to claim 4 is characterized in that: The yaw angles of the sensors when detecting the target are determined respectively by using the initial measurement projections of the sensors, including: Centralizing the initial measurement projections of the sensors to obtain the centralized initial measurement projections of the sensors; The covariance matrix of the initially measured projection after centering of each sensor is calculated respectively, and the angle between the eigenvector corresponding to the covariance matrix and the x-axis is the yaw angle when each sensor detects the target.

6. The distributed three-dimensional extended target tracking method based on feature selection according to claim 1 is characterized in that: The estimated state information acquisition step comprises: The state and covariance of the target at time k-1 are brought into the GIW-PHD filtering algorithm for prediction to obtain the predicted GIW component; The original measurement data after preprocessing of each sensor are respectively brought into the GIW-PHD filtering algorithm to update the predicted GIW component to obtain the updated GIW component of each sensor; According to the set pruning threshold Tr, the merging threshold U between components and the maximum number of GIW components J max The updated GIW components of each sensor are trimmed and merged to obtain new GIW components of each sensor; The GIW components with weights greater than 0.5 in the new GIW components of each sensor are selected to obtain estimated state information of the target detected by each sensor.

7. The distributed three-dimensional extended target tracking method based on feature selection according to claim 5 is characterized in that: The estimated state information preprocessing step comprises: According to the estimated state information of the target detected by each sensor, the ellipse estimated state of the extended target corresponding to each sensor is obtained respectively; The target state of each sensor is obtained by combining the ellipse estimation state of the extended target corresponding to each sensor with the yaw angle when each sensor detects the target.

8. The distributed three-dimensional extended target tracking method based on feature selection according to claim 1 is characterized in that: The target status information acquisition step comprises: Selecting the target states of two sensors with the largest long axis difference among the target states of the sensors for feature extraction and fusion, and obtaining extended state information of the fused target; The extended state information of the fusion target is used to fuse the target center position to obtain the fusion target state information at time k.

Images