A multi-station unmanned aerial vehicle track batch processing method in a complex background noise environment

Abstract

The application discloses a multi-station unmanned aerial vehicle track batch processing method under a complex background noise environment, which is operated based on a radar networking including multiple low-altitude monitoring radars and a data center. Firstly, a Cartesian coordinate system for measuring the measured airspace is established, and three-dimensional coordinates of all echo points detected by the radars are determined, then all echo points detected by the low-altitude monitoring radars within a period of time are superimposed into a blank frame according to original positions to generate a comprehensive echo frame. Then, the echo points belonging to adjacent two comprehensive echo frames are paired, the echo points possibly reflecting the real positions of targets are selected through a Hough transform voting mechanism, and corresponding observation vectors are formed according to the selected echo points. Then, the observation vectors are input into an LSTM network for analysis and screening, it is judged which is a track feature and which is noise, the noise is removed, the track feature is reserved and output, and finally, a final track result is obtained through an FCN.

Description

Technical Field

[0001] This invention belongs to the field of UAV trajectory initiation technology, specifically relating to a method for batch processing of multi-station UAV trajectories in complex background noise environments. Background Technology

[0002] With the widespread use of drones in logistics delivery, aerial photography, and other fields, low-altitude monitoring and control have become increasingly important. The initial track reflects the path a drone takes from takeoff to entering its predetermined route and is the primary step in low-altitude monitoring. The determination of the track relies on methods for processing the positions of targets appearing in chronological order. Efficient and accurate initial track processing is crucial for continuous monitoring of target trajectories. Currently, drone monitoring is mostly based on traditional single-station radar monitoring. Initial track processing methods mainly include the logical initialization method, which uses logical judgment to filter valid target points and establish an initial track; the multi-hypothesis correlation method, which uses one-sided multi-hypothesis correlation of radar point data to form temporary tracks and gradually verify their effectiveness; and the Hough transform method, which uses Hough transform to extract straight / curved features from radar or visual data. However, drones are usually in complex environments during takeoff. Buildings, trees, and ground object echoes from complex terrain can become noise in radar echoes, severely affecting the radar's continuous monitoring of drones. In addition, due to the high mobility of UAVs, the traditional trajectory processing methods mentioned above result in long trajectory processing cycles and low detection probabilities when performing initial batch processing of UAV trajectories.

[0003] With the widespread adoption of low-altitude monitoring radars and the development of the internet, radar networks composed of multiple radar stations and data centers can achieve multi-angle monitoring of targets by operating synchronously across multiple stations and uploading monitoring data to the data center in real time. This effectively avoids incomplete monitoring data caused by distance or azimuth limitations of single stations, and the potential information contained in multi-station monitoring data can also reduce noise, i.e., the impact of ground object echoes on track processing. However, monitoring data obtained by multi-station radars scanning the airspace at different times and wavelengths still needs to be processed by an effective data processing method to derive the track, and only then can radar networks fully realize the aforementioned advantages. Currently, exploration in the field of UAV track monitoring is still in its early stages, and research on track initial batch processing methods based on multi-station radars has not yet yielded significant results. Summary of the Invention

[0004] The purpose of this invention is to provide a method for batch processing of multi-station UAV tracks in complex background noise environments. This method is based on echo frame data measured by multi-station radar. It uses Hough transform to filter out echo points in the echo frames that may reflect the true location of the target. Based on the potential track information contained in these echo points, a target feature dataset is constructed. Then, a Long Short-Term Memory (LSTM) network is used to filter out data in the dataset that can truly reflect the target track based on correlation characteristics. Finally, a Fully Convolutional Network (FCN) is used to complete the track fitting.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for batch processing of multi-station UAV flight paths in complex background noise environments, based on a radar network, wherein the radar network includes multiple low-altitude monitoring radars and a data center. The multiple low-altitude monitoring radars are located in different locations but monitor the same airspace, and all low-altitude monitoring radars are connected to the data center. The method specifically includes the following steps:

[0007] Step 1. Establish a Cartesian coordinate system to measure the entire measured airspace. Multiple low-altitude monitoring radars continuously monitor the airspace simultaneously and transmit the measured echo frame data to the data center in real time. Each radar corresponds to a unique number. After receiving the echo frame data, the data center determines and records the azimuth coordinates of all echo points appearing in the echo frame according to the Cartesian coordinate system. Each echo point also corresponds to a unique number.

[0008] Step 2. The data center periodically superimposes the echo points that appear in all echo frames received within a certain period of time into a blank frame. The new frame generated after superposition is called a composite echo frame. The period of time is a cycle, and the cycle length is equal to the interval between the generation of two adjacent composite echo frames.

[0009] Step 3. Use the Hough transform voting mechanism to filter all echo points belonging to two adjacent composite echo frames, select echo points that may reflect the true position of the target, and obtain the potential track information of the echo point based on the coordinates corresponding to the selected echo point. Then, construct the observation vector corresponding to the echo point. The observation vector includes the position information and potential track information of the corresponding echo point.

[0010] Step 4. Further analyze the observation vectors, select the observation vectors that can reflect the true trajectory, and use the trajectory decision maker to fit the echo points corresponding to the selected observation vectors to obtain the initial trajectory.

[0011] In existing technologies, track initiation processing methods are mainly based on echo frame data obtained from single-station radar. However, single-station radar is susceptible to interference from complex terrain conditions, making reliable and stable monitoring of low-altitude targets such as UAVs difficult. Radar networks containing multi-station radars can detect the same airspace from different directions. Analyzing the echo frames received by multi-station radars can significantly increase the probability of determining the target's true location, while minimizing the impact of buildings, terrain obstructions, or terrain echoes on target detection. However, the data center of a radar network requires an effective processing method for multi-station radar echo frame data to fully realize its advantages. The radar network described in this invention can utilize radars with different electromagnetic wave wavelengths and scanning periods because the method processes tracks through the location of echo points. Regardless of the type of radar, as long as the echo frame contains echo points representing the target's true location, it will contribute to the final processed track.

[0012] Further optimization is needed. The radar network consists of m radars, where m ≥ 2. The echo frame received by the i-th radar in Step 1 is denoted as Rad_img. i , 1≤i≤m; for echo frame Rad_img i All echo points appearing in the echo frame are numbered, and the echo frame is named Rad_img. i All echo points appearing in the frame are represented as Rad_img i ={dot R i_1 ,…,dot R i_P |t s}, where P is the number of echo points in the echo frame, dot R i_1 and dot R i_P These are the first and Pth echo points detected by the i-th radar, respectively, t s Let {dot} be the time when the i-th radar receives the echo frame; the azimuth coordinates of any echo point in the echo frame are represented as {dot} R i_j |t s}={pos x ,pos y ,pos z}, where j is the echo point in the echo frame Rad_img i The index in the sequence, 1≤j≤P, pos x pos y pos zThese are the x-axis, y-axis, and z-axis coordinates of the point in the Cartesian coordinate system. The coordinates of the echo points are added to a set belonging to that point, and then the echo points contained in each echo frame are added to a set belonging to that frame. The echo points in each set are numbered sequentially, which facilitates subsequent data processing.

[0013] Further optimization involves the following method for superimposing the combined echo frames in Step 2: first, find t... s All echo frames within the same period are taken, and each echo point appearing in these echo frames is added to a blank frame according to its original coordinates. The resulting frame is the composite echo frame; the composite echo frame is denoted as Fusion_img. t The combined echo frame and all echo points within it are represented as Fusion_img t ={dot F t_1 ,…,dot F t_N |T}, where T is the period length, t is the ordinal number corresponding to the period, and N is the number of echo points in the synthesized echo frame. The echo points in the synthesized echo frame are denoted as dot. F t_k It also records the source information of the echo points, where k is the echo point in the Fusion_img composite echo frame. t In the sequence number, 1≤k≤N, the source information includes the Rad_img where the echo point is located. i The corresponding t s The purpose of this step is to categorize echo points into different batches to prepare for subsequent screening. Recording the source information of the echo points prepares for the analysis of potential track information contained in the subsequent echo points.

[0014] Further optimization involves the echo points in Step 3 first undergoing distance threshold filtering, followed by Hough transform voting mechanism filtering. The specific steps for distance threshold filtering are as follows:

[0015] Step 3.1. Set the distance threshold dis_ threshold The current composite echo frame is Fusion_img t The previous synthesized echo frame is Fusion_img t-1 Fusion_img t The mid-echo point is denoted as dot. F t_k Fusion_img t-1 The mid-echo point is denoted as dot. F t-1_k ;

[0016] Step 3.2. Retrieve Fusion_imgt Middle one echo point dot F t_k Calculate dot F t_k With Fusion_img t-1 Dot of each echo point F t-1_k Distance between The calculation formula is:

[0017]

[0018] Where pos x_t pos y_t pos z_t dot F t_k In a Cartesian coordinate system, the corresponding x-axis, y-axis, and z-axis coordinates, pos x_t-1 pos y_t-1 pos z_t-1 dot F t-1_k The corresponding x-axis, y-axis, and z-axis coordinates in the Cartesian coordinate system;

[0019] If dot F t_k dot F t-1_k The distance between two points is less than a distance threshold, which satisfies... If the condition is met, then the dot that meets the condition will be... F t_k dot F t-1_k Include the collection of points (Collection_Dot) t middle;

[0020] When Fusion_img t-1 All echo points in the middle are completed with the current dot. F t_k After calculating the distance, select another dot. F t_k Repeat the above process until Fusion_img t All echo points have been calculated.

[0021] The distance threshold filtering is used to initially identify echo points in the composite echo frame that are related to the flight path. The target monitored by this invention is a drone. Based on the maximum flight speed, the maximum distance the drone may travel within the composite echo frame period can be determined. If an echo point is far from all echo points in adjacent echo frames, the likelihood that the echo point represents a drone can be ruled out. Distance threshold dis_threshold These are empirical parameters, determined by the operator based on the target situation and the specific monitoring environment.

[0022] Further optimization involves the following steps in Step 3, where echo points are selected using the Hough transform voting mechanism:

[0023] Step 3.3. The specific steps for filtering echo points using the Hough transform voting mechanism are as follows:

[0024] Step 3.3.1. Assume there is a point M in the Cartesian coordinate system. Through point M, we can obtain an equation of the following form.

[0025] ρ = pos x_m cosθcosφ+pos y_m sinθcosφ+pos z_m sinφ (2)

[0026] Where pos x_m pos y_m pos z_m Let be the x-axis, y-axis, and z-axis coordinates of point M, respectively. Let l be the projection of the line OM connecting point M and the origin onto the xy-plane. Let θ be the angle between l and the positive x-axis. Let φ be the angle between the line OM and the xy-plane. Let ρ be the length of the line OM. In formula (2), pos... x_m pos y_m pos z_m By replacing the independent variables x, y, and z, we can obtain the equation of the line in the Cartesian coordinate system:

[0027] ρ=xcosθcosφ+ysinθcosφ+zsinφ (3)

[0028] This line also represents a point in Hough space with coordinates (ρ, θ, φ);

[0029] Step 3.3.2. Vote on the points in the Hough space to determine the line that passes through the most echo points, where all echo points belong to Collection_Dot. t The specific steps for voting are as follows:

[0030] a. Divide the range [-90°, 90°] corresponding to θ into multiple equal parts using the first type of dividing point, where the first type of dividing point is the angle value of θ. Divide the range [0, 180°] corresponding to φ into multiple equal parts using the second type of dividing point, where the second type of dividing point is the angle value of φ. Then combine the first type of dividing point and the second type of dividing point in pairs, and substitute each combination into formula (3) to generate a corresponding straight line formula. Then, use Collection_Dot tThe coordinates of the echo points are substituted one by one into the formula of each generated straight line to calculate the corresponding ρ value; then the frequency of all ρ values ​​is counted to find the ρ value with the most repetitions, as well as θ and φ in the formula of the straight line that calculates the ρ value. The points in Hough space are determined by finding the corresponding ρ, θ, and φ values.

[0031] b. Find multiple consecutive point sets Collection_Dot t After passing through the line with the most echo points, there will be multiple points distributed in the Hough space. Determine the range of these point distributions, divide ρ into multiple line segments of equal length within this range, and divide θ and φ into multiple angles of equal size. Discretize ρ, θ, and φ to divide the space within this range into a grid state. Then count the number of points in each grid and find the grid with the most points, which is recorded as the peak grid. Determine the equation of the line corresponding to the points in the peak grid and find the echo points on these line equations.

[0032] Unlike existing technologies that directly derive tracks through Hough transform voting, the purpose of Hough transform voting in this invention is to filter echo points, and the Hough transform voting in this invention involves two voting processes. After obtaining the specific coordinates of a certain number of echo points, using the Hough transform voting mechanism to filter out a straight line to derive the target's track for that stage is a common track processing method. However, existing track processing methods transform the two-dimensional coordinates of the echo points, failing to reflect the target's pitch motion. This invention processes three-dimensional coordinates, more realistically reflecting the target's trajectory. The Hough transform voting mechanism is based on statistical principles: multiple points with defined coordinates within a certain range can determine multiple straight lines. The voting mechanism counts which straight line has the most points. Because the target's trajectory is continuous, the trajectory over short distances can be approximated as a straight line; therefore, the straight line passing through more points is closer to the target's true trajectory. For this invention, each point set (Collection_Dot)... t Each Hough transform vote yields a single straight line representing the trajectory. After generating multiple lines, further processing is required, which involves another round of voting for selection. This selection process transforms these lines into points in Hough space, then divides the entire Hough space into a grid and counts the number of points in each grid. The grid division method is also empirical. This selection method essentially quantifies the overlap of lines. Many points located in the same grid indicate a high degree of overlap between the lines corresponding to these points, suggesting that the points corresponding to the actual trajectory are likely also within this grid, further narrowing down the range of the actual trajectory.

[0033] Further optimization is needed; the specific process for constructing the observation vector in Step 3 is as follows:

[0034] Step 3.4. Assemble the echo points identified in Step 3.3 and each echo point dot. F t_k The corresponding observation vector, wherein the elements of the observation vector include dot F t_k Coordinates in a Cartesian coordinate system, dot F t_k The Hough space coordinates corresponding to the line, and the target's position in the dot. F t_k The speed, heading angle, and pitch angle at the target are the UAV for which initial trajectory information needs to be obtained; the speed, heading angle, and pitch angle are obtained through dot... F t_k And with dot F t_k dot on the same straight line F t-1_k The Cartesian coordinates of the two points are calculated to obtain the dot. F t_k The velocity at that point is expressed as The heading angle is expressed as Where ξ is the heading offset, determined by the specific heading reference direction and the target heading, and the pitch angle is expressed as... echo dot F t_k The observation vector at point is denoted as Input_dot F t_k ={pos x_t ,pos y_t ,pos z_t ,ρ k ,θ k ,φ k V k avg ,Yaw k Pitch k}, where Δt is dot F t_k Located in Rad_img i Corresponding to t s With dot F t-1_k Located in Rad_img i Corresponding to t s The difference, ρ k θ k φ k dot F t_k The Hough space coordinates corresponding to the equation of the line.

[0035] The purpose of constructing observation vectors is to uncover the potential trajectory information contained in these selected echo points for further filtering. The observation vectors include multiple elements, including position and motion information, to provide various bases for further filtering and increase the accuracy of the final trajectory.

[0036] Further optimization involves using an LSTM network to filter observation vectors that reflect the true trajectory in Step 4. The trajectory decision maker is an FCN. The specific steps for deriving the initial trajectory are as follows:

[0037] Step 4.1. Input the observation vectors into the LSTM network. The LSTM network learns from the elements contained in the observation vectors to gain the ability to identify the temporal features of the true trajectory. It analyzes and filters all input observation vectors, defining the observation vectors that can reflect the true trajectory as trajectory features and retaining them, while defining observation vectors other than trajectory features as noise and eliminating them. Then, the trajectory features are used as the output result, denoted as Output_dot. F t_k ;

[0038] Step 4.2. Convert Output_dot F t_k Then, the input is processed by the FCN, using the FCN's deconvolution function in conjunction with Output_dot. F t_k The location and track information contained in Output_dot F t_k Corresponding dot F t_k and dot F t_k dot on the same straight line F t-1_k The fit is converted into a track line, and the initial batch processing result of the track is obtained.

[0039] LSTM networks are a type of neural network that uses a gate structure, including a forget gate, a candidate gate, an input gate, and an output gate. The specific computational process is shown in the following formula:

[0040] f t =min(σ(W) f [h t-1 ,x t ]+b f ), alpha)

[0041] O t =σ(W o [h t-1 ,x t ]+b o )

[0042] i t =σ(W i [h t-1 ,x t ]+b i ),

[0043] C` t =tanh(W c [h t-1 ,x t ]+b h )

[0044] C t =f t *C t-1 +i t *C` t

[0045] h t =O t *tanh(C t )

[0046] Where f t C` t i t O t These are the forget gate, candidate gate, input gate, and output gate, respectively. f W c W i W o These are the forget gate weight matrix, candidate gate weight matrix, input gate weight matrix, and output gate weight matrix, respectively. f b i b o These are the biases for the forget gate, input gate, and output gate, respectively. t It is the input value, h t with h t-1 These represent the current output value and the previous output value, respectively. σ is the sigmoid activation function, alpha is a hyperparameter used to limit the output range of the forget gate and avoid extreme gradient fluctuations, tanh is the activation function that restricts the output value range to (-1, 1), and C... t The memory cell state is used to control the output value h. t .

[0047] From the formula above, we can see that W f W c W i W o The calculation of the four weight matrices involves not only the input values ​​but also the previous output values, therefore f t C` t it O t The output of each of the four gates is affected by the output of the previous gate. In other words, W f W c W i W o The output of the four weight matrices is corrected each time; this is the process by which the LSTM network learns the temporal features of the real trajectory through the observation vectors. The current memory cell state C... t It consists of two parts, one part is passed through the forgetting gate f t The data that doesn't need to be retained from the previous memory cell state is filtered out, and the remaining part is the currently retained data. The two parts are linearly superimposed to form C. t Regarding track processing, the forget gate f t It can be used to filter out observation vectors containing noise, input gate i t This is used to retain the selected flight path features for continued iteration. Because UAVs exhibit continuous motion within a certain timeframe, while noise introduced by ground echoes and other sources lacks motion characteristics, motion characteristics can serve as a crucial basis for noise identification. Furthermore, comparing the positional and motion information contained in multiple observation vectors can also serve as another basis, allowing for multiple identification to increase accuracy. Using tanh as the activation function controls the impact of the previous calculation's output value on the subsequent calculation. t A negative value indicates that output is not only not allowed but also suppressed; a positive value indicates reasonable output. This occurs after all loop layers or loop calculations have been performed a predetermined number of times, and h... t When Y is positive, let t =h t Y t This is the final output result after all observation vectors have been analyzed and filtered by the LSTM network.

[0048] The deconvolution function of FCN is often used in image processing. While preserving pixel spatial information in the input image, the deconvolution layer of FCN can restore low-resolution feature maps to high-resolution images. The echo points corresponding to the track features obtained by the LSTM network can be considered as pixels in a low-resolution feature map, while the positional and motion information contained in the track features can be considered as the spatial information of the pixels. Ensuring that the track features reflect the real track, the deconvolution function of FCN can fit the echo points corresponding to the track features into a track path that is as close as possible to the real track, completing the initial batch processing of the track.

[0049] The beneficial effects of the method of the present invention are as follows:

[0050] 1. The method of the present invention realizes batch processing of the initial stage trajectory of small low-altitude targets such as UAVs through monitoring data of radar network. Compared with the existing trajectory initial batch processing method, the method of the present invention has a better processing effect on the initial batch processing of UAV trajectory.

[0051] 2. The radar network used in the method of the present invention can include radars that emit electromagnetic waves with different wavelengths and radars with different scanning periods, which is more flexible in terms of organization and scheduling compared with traditional radar networks;

[0052] 3. The coordinates corresponding to the echo points are all three-dimensional coordinates. The observation vector contains not only the position information of the corresponding echo point, but also the motion information. The position information includes Cartesian coordinates and the Hough space coordinates corresponding to the line it is on, which makes the selection of echo points and the track processing results more accurate.

[0053] 4. The combined use of LSTM networks and FCN can achieve better results in the initial batch processing of tracks compared to other data processing methods. Attached Figure Description

[0054] Figure 1 A schematic diagram of the overall process of the trajectory initiation batch processing method of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below through specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1:

[0057] In this embodiment, the radar network includes two low-altitude monitoring radars and a data center. The two low-altitude monitoring radars are located in different positions but their monitoring airspace overlaps. All low-altitude monitoring radars are connected to the data center. The track initiation batch processing method specifically includes the following steps:

[0058] Step 1. Establish a Cartesian coordinate system to measure the entire measured airspace, with upward as the positive z-axis, westward as the positive y-axis, and northward as the positive x-axis. The origin of the coordinate system is located at sea level, and the coordinate unit is meters (m). Two low-altitude monitoring radars simultaneously and continuously monitor the airspace, transmitting the measured echo frame data to the data center in real time. The two radars are designated as Radar 1 and Radar 2, with Radar 1 having a scan cycle of 2 seconds and Radar 2 having a scan cycle of 3 seconds. 18 seconds after the radar network is activated, one echo point appears in the echo frame received by Radar 1, and two echo points appear in the echo frame received by Radar 2. After determining the coordinates of these three echo points, the echo frame and the corresponding echo point coordinates are labeled Rad_img1={dot R 1_1 |18},{dot R 1_1 |18}={19,24,50}; Rad_img2={dot R 2_1 dot R 2_2 |18},{dot R 2_1 |18}={94,80,7},{dot R 2_2 |18}={30,110,10}, where the three numbers within the curly braces after the echo point are the x-axis, y-axis, and z-axis coordinates of that point. During the period from 16.5 to 24 seconds, in addition to the three echo points mentioned above, five other echo points appear, belonging to the following four echo frames: Rad_img1={dot R 1_1 dot R 1_2 |20},{dot R 1_1 |20}={13,20,50},{dot R 1_2 |20}={544,20,48}; Rad_img2={dot R 2_1 |21},{dot R 2_1 |21}={100,210,70}; Rad_img1={dot R 1_1 dot R 1_2 |22},{dot R 1_1 |22}={7,16,50},{dot R 1_2 |22}={20,39,57}.

[0059] Step 2. Within the time period of 16.5–24s, the echo frame superposition period length T = 2.5s, with the first period being 16.5–19s, and the second and third periods following the same pattern. The three combined echo frames and the echo points contained within the frames are represented as: Fusion_img1 = {dot F 1_1 ,dot F 1_2 ,dot F 1_3 |2.5}, dot F 1_1 ={19,24,50|18}, dot F 1_2 ={94,80,7|18}, dot F 1_3 ={30,110,10|18}; Fusion_img2={dot F 2_1 ,dot F 2_2 ,dot F 2_3 |2.5}, dot F 2_1 ={13,20,50|20}, dot F 2_2 ={544,20,48|20}, dot F 2_3 ={100,210,70|21}; Fusion_img3={dot F 3_1 ,dot F 3_2 |2.5}, dot F 3_1 ={7,16,50|22}, dot F 3_2 ={20,39,57|22}. The coordinates of the point are before the vertical line within the curly braces after the echo point, and the source information for the point is after the vertical line. The source information here refers only to the time when the echo point was detected by the radar.

[0060] Step 3. The specific steps for selecting echo points and constructing observation vectors are as follows:

[0061] Step 3.1. Set the distance threshold dis_ threshold =50m.

[0062] Step 3.2. Use Fusion_img2 as the current synthesized echo frame, and extract the echo points (dot) from Fusion_img2. F2_1 The dot is calculated using formula (1). F 2_1 The respective distances between the three echo points in Fusion_img1 Replace with dot after calculation F 2_2 Then switch to dot F 2_3 All calculation results are listed in Table 1.

[0063] Table 1. Distances between echo points in Fusion_img1 and Fusion_img2

[0064]

[0065] As can be seen from Table 1, only dot F 2_1 and dot F 1_1 The distance between them is less than the distance threshold dis_ threshold The above two points will be included in the point set Collection_Dot2.

[0066] Using Fusion_img3 as the current synthesized echo frame, extract the echo points (dot) from Fusion_img3. F 3_1 The dot is calculated using formula (1). F 3_1 The respective distances between the three echo points in Fusion_img2 Replace with dot after calculation F 3_2 All calculation results are listed in Table 2.

[0067] Table 2 Distances between echo points in Fusion_img2 and Fusion_img3

[0068]

[0069] Table 2 shows that dot F 2_1 With dot F 3_1 dot F 3_2 The distance between them is less than the distance threshold dis_ threshold All three points mentioned above will be included in the point set Collection_Dot3.

[0070] Step 3.3. The straight line passing through the two echo points in Collection_Dot2 is determined by voting on the points in the Hough space. In this embodiment, the range of [-90°, 90°] corresponding to θ in the Hough space is divided into 18000 parts, and the range of [0, 180°] corresponding to φ is also divided into 18000 parts. At this time, there are 18001 values ​​in θ: -90°, -89.99°, -89.98°, ..., 89.99°, 90°, and there are also 18001 values ​​in φ: 0, 0.01°, 0.02°, ..., 179.99°, 180°. Combining the values ​​in θ and φ in pairs can generate 324,036,001 combinations. Substituting these into formula (3) can form 324,036,001 straight line equations. Then, the coordinates (19,24,50) and (13,20,50) of the two echo points in Collection_Dot2 are substituted into the above 324,036,001 linear equations to calculate the corresponding ρ values. Because this step involves too many calculations, the resulting ρ values ​​are not listed individually. Statistics show that ρ = 13.15 appears only twice in all results. When ρ = 13.15, the corresponding θ and φ are θ = -56.31° and φ = 25.66°, respectively. Therefore, the coordinates of the point in Hough space are (13.15, -56.31°, 25.66°), and the corresponding linear equation is 30x - 45y + 26z - 789 = 0, meaning this line passes through the two echo points in Collection_Dot2.

[0071] Then, the same method was used to determine the straight line passing through two echo points in Collection_Dot3. Although there are three echo points in Collection_Dot3, the three points are not collinear. Therefore, voting on the points in Collection_Dot3 yields three straight line equations. The straight line passing through the points (13,20,50) and (7,16,50) corresponds to the Hough space coordinates (13.15,-56.31°,25.66°). The straight line passing through the points (13,20,50) and (20,39,57) corresponds to the Hough space coordinates (21.33,-20.22°,19.07°). The straight line passing through the points (7,16,50) and (20,39,57) corresponds to the Hough space coordinates (3.62,-52.63°,14.84°).

[0072] At this point, four points appear in the Hough space. Dividing the Hough space range of ρ values ​​[3,22] into a length of 1, θ into a range of [0,-60°] into a range of 1°, and φ into a range of [10°,30°] into a range of 1°, the statistical results show that the grid with the ρ values ​​[13,14], θ values ​​[-57°,-56°], and φ values ​​[25°,26°] contains two points, the most numerous, therefore this grid is the peak grid. The two points in the peak grid have coordinates (13.15,-56.31°,25.66°), corresponding to the three echo points of the line equation 30x-45y+26z-789=0, with coordinates (19,24,50), (13,20,50), and (7,16,50), i.e., the dot. F 1_1 dot F 2_1 dot F 3_1 It lies on that straight line.

[0073] Step 3.4. The three selected echo points belong to three different composite echo frames, which can be used to construct two observation vectors. Using dot... F 1_1 With dot F 2_1 The velocity calculated from the coordinates of the two points is V. 1 avg ={-3,-2,0}, in this embodiment the heading reference direction is due north, the target has a velocity in the negative x-axis direction, therefore ξ = 180°, and the heading angle is Yaw. 1 =213.69°, pitch angle is 0. dot F 2_1 The corresponding observation vector is established as Input_dot F 2_1 ={13,20,50,13.15,-56.31°,25.66°,-3,-2,0,213.69°,0}. dot F 2_1 With dot F 3_1 The velocity, heading angle, and pitch angle calculated from the coordinates of the two points are all equal to those described above; therefore, the observation vector is established as Input_dot. F 3_1 ={7,16,50,13.15,-56.31°,25.66°,-3,-2,0,213.69°,0}.

[0074] Step 4. The specific steps for further analyzing the observation vectors and deriving the trajectory are as follows:

[0075] Step 4.1. Input the observation vectors into a single-layer LSTM network. The LSTM network's forget gate weight matrix contains three weight matrices: bewith, cross, and rel. bewith is the weight for assigning a point to a track, used to measure the probability that an echo point belongs to a track or noise; cross is the weight for intersecting tracks, used to measure whether the intersection of echo point lines is a genuine cross track or noise interference; rel is the weight for inter-point correlation, used to measure the probability that different echo points belong to the same track. The LSTM network performs temporal feature analysis and filtering on the two input observation vectors, setting the number of iterations to 15. The analysis results show that both observation vectors can reflect the true track. These two observation vectors are defined as track features and retained through the input gate. The track features are then used as the output, denoted as Output_dot. F 2_1 With Output_dot F 3_1 .

[0076] Step 4.2. Convert Output_dot F 2_1 With Output_dot F 3_1 Then, the input is processed by the FCN, using the FCN's deconvolution function in conjunction with Output_dot. F 2_1 With Output_dot F 3_1 The location information and track information contained in the dot F 1_1 dot F 2_1 dot F 3_1 The three points are fitted to form a track line, and the initial batch processing result of the track is the straight line 30x-45y+26z-789=0. At this point, the initial batch processing of the track using the method of this invention is complete. The overall process of the method of this invention is as follows: Figure 1 As shown.

[0077] Example 2:

[0078] This embodiment is based on Embodiment 1. In this embodiment, in Step 3.3, the Hough space ρ value range [0, 24] is divided into sections with a length of 12, θ in the range [0, -60°] is divided into sections with an angle of 60°, and φ in the range [0, 30°] is also divided into sections with an angle of 15°. The statistical results show that there are three points in the grid with ρ value [12, 24], θ value [0, -60°], and φ value [15°, 30°], which is the most numerous. Therefore, this grid is a peak grid. The coordinates of two points in the peak grid are (13.15, -56.31°, 25.66°), and the coordinates of another point are (21.33, -20.22°, 19.07°). These points correspond to the line equation 88.69x - 32.66y + 32.67z - 2133 = 0, and the two echo points with coordinates (13, 20, 50) and (20, 39, 57), i.e., the dot matrix. F 2_1 dot F 3_2 It lies on this straight line. At this point, in Step 3.4, three observation vectors can be constructed, namely Input_dot F 2_1 ={13,20,50,13.15,-56.31°,25.66°,-3,-2,0,213.69°,0}、Input_dot F 3_1 ={7,16,50,13.15,-56.31°,25.66°,-3,-2,0,213.69°,0}、Input_dot F 3_2 ={20,39,57,-21.33,-20.22°,19.07°,3.5,8.5,3.5,69.78°,19.07°}.

[0079] Inputting the three observation vectors into the LSTM network, it was found that the lines connecting the points intersect at point (13, 20, 50). Cross analysis of the cross track weights showed that this was not a cross track. Further analysis using the weights assigned to the points as tracks (bewith) and the correlation weights (rel) between the points yielded the dot... F 1_1 dot F 2_1 dot F 3_1 The three points are located on the same straight line, and their temporal distribution corresponds to the track information, while the dot F 3_2 With dot F 1_1 dot F 2_1The two points cannot form a reasonable trajectory, and the track information does not correspond. Therefore, Input_dot F 3_2 ={20,39,57,-21.33,-20.22°,19.07°,3.5,8.5,3.5,69.78°,19.07°} is defined as noise and eliminated through a forgetting gate. The final track result is the same as in Example 1. Other parts in this example are the same as those in Example 1.

Claims

1. A method for batch processing of multi-station UAV flight paths under complex background noise environment, characterized in that, Operating based on a radar network, the radar network includes multiple low-altitude monitoring radars and a data center. The multiple low-altitude monitoring radars are located in different locations but monitor the same airspace. All low-altitude monitoring radars are connected to the data center. The method specifically includes the following steps: Step 1. Establish a Cartesian coordinate system to measure the entire measured airspace. Multiple low-altitude monitoring radars continuously monitor the airspace simultaneously and transmit the measured echo frame data to the data center in real time. Each radar corresponds to a unique number. After receiving the echo frame data, the data center determines and records the azimuth coordinates of all echo points appearing in the echo frame according to the Cartesian coordinate system. Each echo point also corresponds to a unique number. Step 2. The data center periodically superimposes the echo points that appear in all echo frames received within a certain period of time into a blank frame. The new frame generated after superposition is called a composite echo frame. The period of time is a cycle, and the cycle length is equal to the interval between the generation of two adjacent composite echo frames. Step 3. Use the Hough transform voting mechanism to filter all echo points belonging to two adjacent composite echo frames, select echo points that may reflect the true position of the target, and obtain the potential track information of the echo point based on the coordinates corresponding to the selected echo point. Then, construct the observation vector corresponding to the echo point. The observation vector includes the position information and potential track information of the corresponding echo point. Step 4. Further analyze the observation vectors, select the observation vectors that can reflect the true trajectory, and use the trajectory decision maker to fit the echo points corresponding to the selected observation vectors to obtain the initial trajectory.

2. The method for batch processing of multi-station UAV tracks in a complex background noise environment as described in claim 1, characterized in that: The radar network consists of m radars, where m ≥ 2. The echo frame received by the i-th radar in Step 1 is denoted as Rad_img. i , 1≤i≤m; for echo frame Rad_img i All echo points appearing in the echo frame are numbered, and the echo frame is named Rad_img. i All echo points appearing in the frame are represented as Rad_img i ={dot R i_1 ,…,dot R i_P |t s }, where P is the number of echo points in the echo frame, dot R i_1 and dot R i_P These are the first and Pth echo points detected by the i-th radar, respectively, t s Let {dot} be the time when the i-th radar receives the echo frame; the azimuth coordinates of any echo point in the echo frame are represented as {dot} R i_j |t s }={pos x ,pos y ,pos z }, where j is the echo point in the echo frame Rad_img i The index in the sequence, 1≤j≤P, pos x pos y pos z These are the x-axis, y-axis, and z-axis coordinates of the point in the Cartesian coordinate system, respectively.

3. The method for batch processing of multi-station UAV tracks in a complex background noise environment as described in claim 2, characterized in that, The method for superimposing the combined echo frames in Step 2 is as follows: first find t s All echo frames within the same period are taken, and each echo point appearing in these echo frames is added to a blank frame according to its original coordinates. The resulting frame is the composite echo frame; the composite echo frame is denoted as Fusion_img. t The combined echo frame and all echo points within it are represented as Fusion_img t ={dot F t_1 ,…,dot F t_N |T}, where T is the period length, t is the ordinal number corresponding to the period, and N is the number of echo points in the synthesized echo frame. The echo points in the synthesized echo frame are denoted as dot. F t_k It also records the source information of the echo points, where k is the echo point in the Fusion_img composite echo frame. t In the sequence number, 1≤k≤N, the source information includes the Rad_img where the echo point is located. i The corresponding t s .

4. The method for batch processing of multi-station UAV tracks in a complex background noise environment as described in claim 3, characterized in that, In Step 3, the echo points need to be filtered by a distance threshold first, and then filtered by the Hough transform voting mechanism. The specific steps of the distance threshold filtering are as follows: Step 3.

1. Set the distance threshold dis_ threshold The current composite echo frame is Fusion_img t The previous synthesized echo frame is Fusion_img t-1 Fusion_img t The mid-echo point is denoted as dot. F t_k Fusion_img t-1 The mid-echo point is denoted as dot. F t-1_k ; Step 3.

2. Retrieve Fusion_img t Middle one echo point dot F t_k Calculate dot F t_k With Fusion_img t-1 Dot of each echo point F t-1_k Distance between The calculation formula is: Where pos x_t pos y_t pos z_t dot F t_k In a Cartesian coordinate system, the corresponding x-axis, y-axis, and z-axis coordinates, pos x_t-1 pos y_t-1 pos z_t-1 dot F t-1_k The corresponding x-axis, y-axis, and z-axis coordinates in the Cartesian coordinate system; If dot F t_k dot F t-1_k The distance between two points is less than a distance threshold, which satisfies... If the condition is met, then the dot that meets the condition will be... F t_k dot F t-1_k Include the collection of points. t middle; When Fusion_img t-1 All echo points in the middle are completed with the current dot. F t_k After calculating the distance, select another dot. F t_k Repeat the above process until Fusion_img t All echo points have been calculated.

5. The method for batch processing of multi-station UAV tracks in a complex background noise environment as described in claim 4, characterized in that, Step 3, which uses the Hough transform voting mechanism to filter echo points, includes the following steps: Step 3.

3. The specific steps for filtering echo points using the Hough transform voting mechanism are as follows: Step 3.3.

1. Assume there is a point M in the Cartesian coordinate system. Through point M, we can obtain an equation of the following form. p = pos x_m cosθcosφ+pos y_m sinθcosφ+pos z_m sinφ (2) Where pos x_m pos y_m pos z_m Let be the x-axis, y-axis, and z-axis coordinates of point M, respectively. Let l be the projection of the line OM connecting point M and the origin onto the xy-plane. Let θ be the angle between l and the positive x-axis. Let φ be the angle between the line OM and the xy-plane. Let ρ be the length of the line OM. In formula (2), pos... x_m pos y_m pos z_m By replacing the independent variables x, y, and z, we can obtain the equation of the line in the Cartesian coordinate system: ρ=xcosθcosφ+ysinθcosφ+zsinφ (3) This line also represents a point in Hough space with coordinates (ρ, θ, φ); Step 3.3.

2. Vote on the points in the Hough space to determine the line that passes through the most echo points, where all echo points belong to Collection_Dot. t The specific steps for voting are as follows: a. Divide the range [-90°, 90°] corresponding to θ into multiple equal parts using the first type of dividing point, where the first type of dividing point is the angle value of θ. Divide the range [0, 180°] corresponding to φ into multiple equal parts using the second type of dividing point, where the second type of dividing point is the angle value of φ. Then combine the first type of dividing point and the second type of dividing point in pairs, and substitute each combination into formula (3) to generate a corresponding straight line formula. Then, use Collection_Dot t The coordinates of the echo points are substituted one by one into the formula of each generated straight line to calculate the corresponding ρ value; then the frequency of all ρ values ​​is counted to find the ρ value with the most repetitions, as well as θ and φ in the formula of the straight line that calculates the ρ value. The points in Hough space are determined by finding the corresponding ρ, θ, and φ values. b. Finding multiple consecutive point sets Collection_Dot t After passing through the line with the most echo points, there will be multiple points distributed in the Hough space. Determine the range of these point distributions, divide ρ into multiple line segments of equal length within this range, and divide θ and φ into multiple angles of equal size. Discretize ρ, θ, and φ to divide the space within this range into a grid state. Then count the number of points in each grid and find the grid with the most points, which is recorded as the peak grid. Determine the equation of the line corresponding to the points in the peak grid and find the echo points on these line equations.

6. The method for batch processing of multi-station UAV tracks in a complex background noise environment as described in claim 5, characterized in that, The specific process of constructing the observation vector in Step 3 is as follows: Step 3.

4. Assemble the echo points identified in Step 3.3 and each echo point dot. F t_k The corresponding observation vector, wherein the elements of the observation vector include dot F t_k Coordinates in a Cartesian coordinate system, dot F t_k The Hough space coordinates corresponding to the line, and the target's position in the dot. F t_k The speed, heading angle, and pitch angle at the target are the UAV for which initial trajectory information needs to be obtained; the speed, heading angle, and pitch angle are obtained through dot... F t_k And with dot F t_k dot on the same straight line F t-1_k The Cartesian coordinates of the two points are calculated to obtain the dot. F t_k The velocity at that point is expressed as The heading angle is expressed as Where ξ is the heading offset, determined by the specific heading reference direction and the target heading, and the pitch angle is expressed as... echo dot F t_k The observation vector at point is denoted as Input_dot F t_k ={pos x_t ,pos y_t ,pos z_t ,ρ k ,θ k ,φ k V k avg ,Yaw k Pitch k }, where Δt is dot F t_k Located in Rad_img i Corresponding to t s With dot F t-1_k Located in Rad_img i Corresponding to t s The difference, ρ k θ k φ k dot F t_k The Hough space coordinates corresponding to the equation of the line.

7. The method for batch processing of multi-station UAV tracks in a complex background noise environment as described in claim 6, characterized in that, In Step 4, the observation vectors that can reflect the true trajectory are filtered through an LSTM network. The trajectory decision maker is an FCN. The specific steps for obtaining the initial trajectory are as follows: Step 4.

1. Input the observation vectors into the LSTM network. The LSTM network learns from the elements contained in the observation vectors to gain the ability to identify the temporal features of the true trajectory. It analyzes and filters all input observation vectors, defining the observation vectors that can reflect the true trajectory as trajectory features and retaining them, while defining observation vectors other than trajectory features as noise and eliminating them. Then, the trajectory features are used as the output result, denoted as Output_dot. F t_k ; Step 4.

2. Convert Output_dot F t_k Then, the input is processed by the FCN, using the FCN's deconvolution function in conjunction with Output_dot. F t_k The location and track information contained in Output_dot F t_k Corresponding dot F t_k and dot F t_k dot on the same straight line F t-1_k The fit is converted into a track line, and the initial batch processing result of the track is obtained.

Images