Target identification method, meteorological radar equipment and storage medium

Abstract

The invention discloses a target recognition method, meteorological radar equipment and a storage medium, and the recognition method comprises the steps: carrying out the recognition of a to-be-recognized three-channel image through a target recognition model, and obtaining each target detection frame and the confidence coefficient of each target detection frame; wherein the three-channel image is obtained by mapping a distance-azimuth-spectral distance map output by a meteorological radar; filtering out the target detection frames of which the confidence is lower than a confidence threshold to obtain high-confidence target detection frames; and based on a target speed and a speed threshold, classifying the high-confidence target detection frames to obtain different types of target detection frames. According to the method, the deep learning model is introduced on the basis of the distance-azimuth-spectral distance map, and the target and the background are distinguished through the deep learning model, so that the influence of backgrounds such as meteorological echoes and various non-meteorological clutters on target recognition is avoided, the target recognition precision is improved, the false alarm rate or missing report rate is reduced, and the target recognition accuracy is improved. And the blank that the meteorological radar cannot identify the LSS target is filled.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar, and particularly relates to a target identification method, a weather radar device and a storage medium. BACKGROUND

[0002] With the rapid popularization of low-altitude, slow-speed and small-size (LSS) targets (such as consumer drones), the potential risks brought by the LSS targets in the fields of public safety and air traffic control are growing. Although there are special detection radars for LSS targets on the market, the independent deployment of the radars often means high construction and maintenance costs, and may face challenges in airspace resource allocation and multi-system coordination. Given the commonality of existing weather radars in hardware configuration and basic detection principles, exploring the identification capability of LSS targets without sacrificing the core weather observation performance becomes a new approach with high cost-effectiveness and multi-functional integration potential.

[0003] Although the LSS target echoes have significant RCS (Radar Cross-Section) and considerable signal-to-noise ratio, in the weather observation mode, the radar resolution volume of the weather radar is usually large, and is filled with strong weather echoes (such as rain, hail, and snow mixture) and various non-weather clutter (such as terrain, biology, wind turbines, and air debris). The traditional signal processing method based on threshold or simple rules is difficult to accurately, real-time and robustly separate the unique signals of the LSS target from the massive and variable background and accurately classify the LSS target in the dynamically changing complex background, which is prone to high false positive rate or false negative rate.

[0004] Many current mainstream LSS target observation radars identify and classify targets according to micro-Doppler characteristics in a single radial direction. To scan the entire airspace, multiple radials need to be scanned. The identification in the radial direction, whether through traditional adaptive threshold or deep learning method, needs to perform convolution operation, which has large amount of calculation. The identification in multiple radial directions will greatly affect the real-time performance. SUMMARY

[0005] The present application aims to provide a target identification method, a weather radar device and a storage medium to solve the problem that the weather radar cannot effectively identify and classify LSS targets when performing conventional weather observation tasks.

[0006] The present application solves the above technical problems by the following technical solutions: a target identification method, the identification method comprising:

[0007] call a target recognition model, the target recognition model is a trained deep learning model, training samples of the deep learning model are three-channel images obtained by mapping a range-azimuth-spectrum distance map, and the range-azimuth-spectrum distance map is a processing result of a weather radar echo signal;

[0008] identify the three-channel image to be identified by using the target recognition model to obtain each target detection frame and a confidence thereof;

[0009] filter out the target detection frame with a confidence lower than a confidence threshold to obtain a high-confidence target detection frame;

[0010] based on the target speed and the speed threshold, classify the high-confidence target detection frame to obtain different-class target detection frames.

[0011] Further, the identification method further includes a target tracking process, comprising:

[0012] based on the current identification and classification result and the radar echo signal processing result, calculate the target state parameter at the current observation time;

[0013] based on the target state parameter at the current observation time, calculate the target state correction prediction value at the next observation time by using the Kalman filtering algorithm, and update the state covariance matrix;

[0014] obtain the next identification and classification result and the radar echo signal processing result;

[0015] based on the next identification and classification result and the radar echo signal processing result, calculate the target state parameter at the next observation time;

[0016] match the target state correction prediction value at the next observation time and the target state parameter at the next observation time to realize target tracking.

[0017] Further, the target state parameter includes target relative distance to the radar, target relative azimuth angle to the radar, target relative elevation angle to the radar, target radial speed, target azimuth angle speed and target elevation angle speed; and the calculation formula of the target relative distance to the radar and the target relative azimuth angle to the radar is:

[0018] ;

[0019] ;

[0020] wherein, denotes the distance of the kth target relative to the radar at the nth observation time; denotes the distance resolution on the range-azimuth-spectrum distance map; ​represents the coordinate of the center of the detection frame of the kth target in the radar coordinate system in the nth identification and classification; represents the elevation angle of the kth target relative to the radar in the nth observation; represents the azimuth angle of the kth target relative to the radar in the nth observation; represents the azimuth angle resolution on the range-azimuth-spectrum range map.

[0021] Further, the target state correction prediction value in the next observation is calculated by using the Kalman filtering algorithm, including:

[0022] calculating the target state prediction value in the next observation according to the target state parameter in the current observation and the state transition matrix;

[0023] calculating the variance matrix in the current observation according to the target state parameter in the current observation, the state covariance matrix and the process noise covariance matrix;

[0024] calculating the Kalman gain according to the variance matrix in the current observation, the observation matrix and the observation noise covariance matrix;

[0025] calculating the target state correction prediction value in the next observation according to the target state prediction value in the next observation, the Kalman gain, the target state parameter in the current observation and the observation matrix.

[0026] Further, the update formula of the state covariance matrix is:

[0027] ;

[0028] wherein, represents the state covariance matrix of the kth target in the n+1th observation; represents the unit matrix; represents the Kalman gain of the kth target; represents the observation matrix; represents the variance matrix of the kth target.

[0029] Further, the target state correction prediction value in the next observation and the target state parameter in the next observation are matched, including:

[0030] calculating the difference between the target state correction prediction value in the next observation and the target state parameter in the next observation;

[0031] judging whether it is the same target according to the difference and the difference threshold, if yes, the matching is successful; if no, the matching fails.

[0032] Further, when the target state correction prediction value at the next observation time fails to match the target state parameter at the next observation time, it is determined whether the state parameter prediction correction value of the corresponding target has failed to match for a plurality of times in succession, and if so, the target is lost, and if not, the target state correction prediction value at the next observation time is taken as the target state parameter at the next observation time, and target tracking is performed.

[0033] Based on the same concept, the present application also provides a weather radar device comprising a memory, a processor and a computer program / instruction stored on the memory, wherein the processor executes the computer program / instruction to implement the target identification method as described above.

[0034] Based on the same concept, the present application also provides a computer readable storage medium having a computer program / instruction stored thereon, wherein the computer program / instruction is executed by a processor to implement the target identification method as described above.

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

[0036] The present application introduces a deep learning model based on the range-azimuth-spectrum distance map output by the weather radar, distinguishes targets and backgrounds through the deep learning model, avoids the influence of weather echoes and various non-weather clutter backgrounds on target identification, improves target identification accuracy, reduces false positive rate or false negative rate, and fills the gap of weather radar in identifying LSS targets.

[0037] The present application makes full use of the powerful detection capability and data basis of the deployed weather radar, and expands its compatible identification capability for LSS targets at low cost and high efficiency, significantly saves the construction and operation cost of infrastructure, and greatly improves the comprehensive utilization rate and multifunctionality of the existing weather radar.

[0038] Compared with the prior art, the present application directly uses the final output result of the weather radar for target identification, without the need for identification at each radial, greatly reduces the amount of calculation, and improves the real-time performance of identification. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only one embodiment of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0040] Figure 1 is a flow chart of the target identification method in the embodiment of the present application;

[0041] Figure 2 is a three-channel image to be identified in the embodiment of the present application;

[0042] Figure 3 is a recognition result schematic diagram in the embodiment of the present application;

[0043] Figure 4 is a classification result schematic diagram in the embodiment of the present application;

[0044] Figure 5 is a flowchart of steps C1 to C6 in the target identification and tracking process in the embodiment of the present application;

[0045] Figure 6 is a flowchart of steps C7 to C15 in the target identification and tracking process in the embodiment of the present application;

[0046] Figure 7 is a flowchart of steps C16 to C31 in the target identification and tracking process in the embodiment of the present application;

[0047] Figure 8 is a Kalman filter tracking curve of the unmanned aerial vehicle target in the embodiment of the present application;

[0048] Figure 9 is an actual unmanned aerial vehicle target tracking effect diagram in the embodiment of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0050] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and some embodiments may not be described again for the same or similar concepts or processes.

[0051] Embodiment One

[0052] In order to solve the problem that the weather radar cannot effectively identify and classify LSS targets when performing a conventional weather observation task, the present application provides a target identification method, which introduces a deep learning model on the basis of the range-azimuth-spectrum distance map output by the weather radar, distinguishes targets and backgrounds through the deep learning model, and realizes efficient and accurate identification of potential LSS targets.

[0053] As shown in Figure 1 The target identification method applied to the weather radar provided by the embodiment of the present application includes the following steps:

[0054] Step A1: calling a target recognition model.

[0055] The target recognition model is a trained deep learning model, such as Mask R-CNN, YOLO series model, etc. Considering the real-time performance, the embodiment selects the YOLO series model, such as YOLOv11 model. In order to train the deep learning model, a sample data set is first constructed, and each sample in the sample data set includes a three-channel image obtained by mapping the range-azimuth-spectrum range map and a label. The acquisition process of the three-channel image is as follows:

[0056] Wrap the metal film and the sounding instrument around the air floating target (such as a balloon) and release it, and the actual position of the air floating target can be determined through the sounding instrument; meanwhile, there are unmanned aerial vehicles and other LSS targets in the airspace; scan the airspace using a weather radar, and perform pulse compression, interference suppression, clutter suppression, and spectrum range estimation processing (such as reflectivity, velocity, spectrum width, etc.) on the received radar echo signal to obtain a range-azimuth-spectrum range map; map the range-azimuth-spectrum range map to a three-channel image and save it.

[0057] Since the result of spectrum range estimation is a single-channel value, the range-azimuth-spectrum range map is a grayscale image, and the features are not obvious. In order to improve the recognition accuracy, the range-azimuth-spectrum range map is mapped to a three-channel image, i.e. an RGB image. In the PPI view, the closer to the center, the smaller the area (the smaller the radius, the smaller the arc length, and the smaller the corresponding area), which affects the recognition result, so the PPI view is not used for target recognition.

[0058] Considering that the volume of unmanned aerial vehicles and air floating objects is much larger than that of precipitation particles and water droplets in the air, and the body material of unmanned aerial vehicles and metal film both have high reflectivity, the reflectivity spectrum parameter will be significantly higher than the weather echo, so the labeling and recognition are based on the range-azimuth-spectrum range map.

[0059] According to the air floating target position and the actual position of the unmanned aerial vehicle returned by the sounding instrument, each target is labeled in the range-azimuth-spectrum range map to obtain the label of the range-azimuth-spectrum range map. During the labeling process, not only the known targets are labeled, but also other suspected targets are labeled. Considering that the volume of LSS targets is relatively small compared to the airspace, the class features on the range-azimuth-spectrum range map are not obvious, so the targets are not classified during labeling, and are all classified as one class of targets.

[0060] The sample data set is divided into a training set and a validation set, the training set is used to train the deep learning model, and the validation set is used to verify the trained deep learning model, i.e. to obtain the target recognition model.

[0061] Step A2: identifying the three-channel image to be identified by using the target recognition model to obtain each target detection frame and its confidence.

[0062] In actual identification, a range-azimuth-spectrum map output by a weather radar is acquired, and the range-azimuth-spectrum map is mapped into a three-channel image, that is, a three-channel image to be identified is obtained, as shown in FIG. 2. Figure 2 The three-channel image to be identified is input into the target recognition model, and each target detection frame and its confidence are output, as shown in FIG. 3. Figure 3 Figure 3 The numbers in FIG. 3 represent the confidence of the target detection frame. The target recognition model not only outputs the target detection frame, but also outputs the confidence of the target detection frame, which indicates the possibility that the target is a real target.

[0063] Step A3: filtering out the target detection frame with a confidence lower than a confidence threshold to obtain a high-confidence target detection frame.

[0064] The target detection frame with a confidence lower than the confidence threshold is considered to be a false target, and is thus filtered out. In this embodiment, the confidence threshold is set to 0.5.

[0065] Step A4: classifying the high-confidence target detection frame based on a target speed and a speed threshold to obtain different-class target detection frames.

[0066] The target speed is determined on the range-azimuth-spectrum map, and the high-confidence target detection frame is divided into different classes according to the target speed and the set speed threshold, so as to realize target classification. Figure 3 In the speed spectrum of the three targets in FIG. 4, the deeper the color of the target detection frame, the higher the target speed. In this embodiment, the speed threshold is set to 3 m / s, and the high-confidence target detection frame is divided into a UAV target detection frame and a floating object target detection frame according to the target speed and the speed threshold, as shown in FIG. 5. Figure 4 Figure 4 The numbers in FIG. 5 represent the confidence of the target detection frame.

[0067] In the specific embodiments of the present application, based on the identification and classification results, the present application also performs target tracking, and the specific tracking process includes:

[0068] Step B1: calculating the target state parameter at the current observation time based on the current identification and classification results and the radar echo signal processing results.

[0069] ​​In this embodiment, the target state parameters include the target's distance from the radar (i.e., the horizontal distance between the target and the weather radar), the target's azimuth relative to the radar, the target's elevation angle relative to the radar, the target's radial velocity, the target's azimuth velocity, and the target's elevation velocity. The target's elevation angle relative to the radar is the elevation angle of the current beam direction; the target's radial velocity is determined based on the radar echo signal; the target's distance from the radar and the target's azimuth relative to the radar are determined based on the recognition and classification results. The specific calculation formula is:

[0070] (1)

[0071] (2)

[0072] (3)

[0073] (4)

[0074] in,( , ) represents the coordinates of the upper left corner of the detection box of the kth target in the radar coordinate system during the nth recognition and classification; ( , ) represents the coordinates of the lower right corner of the detection box of the k-th target in the radar coordinate system during the n-th recognition and classification; ( , ) represents the coordinates of the detection box center of the k-th target in the radar coordinate system during the n-th recognition and classification; Indicates the distance of the kth target relative to the radar at the nth observation; Indicates the range resolution on the range-azimuth-spectrum diagram; represents the elevation angle of the kth target relative to the radar during the nth observation; represents the azimuth of the kth target relative to the radar at the nth observation; Indicates the azimuth resolution on the range-azimuth-spectrum spectrum; k indicates the sequence number of the tracked target. In this embodiment, the range resolution The azimuth resolution is 30m It is 1.2°.

[0075] In order to reduce the coordinate conversion operation from polar coordinates to Cartesian coordinates, two-dimensional coordinates are used to represent the target position.

[0076] At time zero, the target azimuth velocity and target pitch angular velocity are 0. During the tracking process, the target azimuth velocity and target pitch angular velocity can be defined as:

[0077] (5)

[0078] (6)

[0079] wherein, denotes the azimuth angle velocity of the kth target at the nth observation; denotes the elevation angle velocity of the kth target at the nth observation; denotes the azimuth angle of the k+1th target relative to the radar at the nth observation; denotes the elevation angle of the kth target relative to the radar at the n+1th observation, i.e., the difference between the azimuth angles of the target relative to the radar at adjacent two observations as the angular velocity in the horizontal direction (i.e., the azimuth angle velocity), the difference between the elevation angles of the target relative to the radar at adjacent two observations as the angular velocity in the vertical direction (i.e., the elevation angle velocity).

[0080] Step B2: based on the target state parameters at the current observation, the target state correction prediction value at the next observation is calculated by using the Kalman filtering algorithm, and the state covariance matrix is updated.

[0081] the target state matrix at the current observation is obtained according to the target state parameters at the current observation :

[0082] (7)

[0083] wherein, denotes the target state parameters (or the target state matrix) of the kth target at the nth observation, and the superscript T denotes the matrix transpose.

[0084] In order to ensure the tracking performance, the Kalman filtering algorithm is adopted in the application, the Kalman filtering algorithm is a self-adaptive filtering, and supports establishing multiple tracking links, for example, when two targets are close to each other, two sets of tracking parameters are refreshed according to the positions of the two targets, i.e., two tracking links are established; when the target is far away, the tracking link of the wrong target will lose the target, and the correct tracking is ensured. Before target tracking by using the Kalman filtering algorithm, the tracking parameters are initialized, which specifically include the target state matrix, the state transition matrix, the observation matrix, the state covariance matrix, the process noise covariance matrix and the observation noise covariance matrix. The target state matrix is initialized according to the observed target state parameters, and the initialization of other matrices is specifically as follows:

[0085] The state transition matrix P is:

[0086] (8)

[0087] The observation matrix H is:

[0088] (9)

[0089] State Covariance Matrix is:

[0090] (10)

[0091] Process Noise Covariance Matrix is:

[0092] (11)

[0093] Observation Noise Covariance Matrix is:

[0094] (12)

[0095] wherein, denotes the time interval between two adjacent observations; , and respectively denote the state covariance coefficient, the process noise covariance coefficient and the observation noise covariance coefficient. The greater the state covariance coefficient, the greater the uncertainty of the state estimation; the greater the process noise covariance coefficient, the more likely the target state to appear nonlinear change in the derivation process; the greater the observation noise covariance coefficient, the greater the observation error, and the less reliable the observation value.

[0096] It should be noted that the state covariance matrix will be updated along with the tracking process, so the state covariance matrix is only initialized once at the time when the target is discovered (i.e. n = 0).

[0097] In the specific embodiments of the present application, the Kalman filtering algorithm is used to calculate the target state correction prediction value at the next observation, including:

[0098] Step B2.1: Calculate the target state prediction value at the next observation according to the target state parameter at the current observation and the state transition matrix, and the specific formula is:

[0099] (13)

[0100] wherein, denotes the state prediction value (or target state prediction matrix) of the kth target at the n+1th observation, which is a linear prediction result; denotes the target state value of the kth target at the nth observation.

[0101] Step B2.2: Calculate the covariance matrix at the current observation according to the target state parameter at the current observation, the state covariance matrix and the process noise covariance matrix, and the specific formula is:

[0102] (14)

[0103] wherein, denotes the prediction covariance matrix of the kth target at the nth observation.

[0104] Step B2.3: Calculate the Kalman gain according to the prediction covariance matrix at the current observation, the observation matrix and the observation noise covariance matrix, and the specific formula is:

[0105] (15)

[0106] wherein, denotes the Kalman gain of the kth target.

[0107] Step B2.4: Calculate the target state correction prediction value at the next observation according to the target state prediction value at the next observation, the Kalman gain, the target state parameter at the current observation and the observation matrix, and the specific formula is:

[0108] (16)

[0109] wherein, denotes the state correction prediction value (or target state correction prediction matrix) of the kth target at the n+1th observation, that is, the optimal target state prediction value. Based on the linear prediction result of the present application, a correction value calculated based on the Kalman gain is added to realize the correction of the linear prediction result, reduce the prediction deviation and improve the prediction accuracy.

[0110] The update formula of the state covariance matrix is:

[0111] (17)

[0112] wherein, denotes the state covariance matrix of the kth target at the n+1th observation; denotes the unit matrix.

[0113] Step B3: Obtain the next identification and classification result and the radar echo signal processing result.

[0114] According to steps A1 to A4, the next identification and classification result can be obtained, and the next identification and classification result is the detection frame of different class targets obtained by the next identification and classification; the radar echo signal obtained in the next identification and classification is processed, and the radar echo signal processing result is obtained, including the target speed.

[0115] Step B4: Calculate the target state parameter at the next observation based on the next identification and classification result and the radar echo signal processing result.

[0116] The target state parameter at the next observation time includes the target's distance relative to the radar, the target's azimuth angle relative to the radar, the target's elevation angle relative to the radar, the target's radial velocity, the target's azimuth angle velocity and the target's elevation angle velocity at the next observation time.

[0117] Step B5: Matching the target state correction prediction value at the next observation time and the target state parameter at the next observation time to realize target tracking.

[0118] In the specific embodiment of the present application, matching the target state correction prediction value at the next observation time and the target state parameter at the next observation time includes:

[0119] calculating the difference between the target state correction prediction value at the next observation time and the target state parameter at the next observation time; judging whether it is the same target according to the difference and the difference threshold value, if yes, the matching is successful; if no, the matching fails. The specific formula is:

[0120] (18)

[0121] wherein, represents taking the absolute value; represents the 0th, 1st and 2nd elements of the target state correction prediction matrix , i.e. the predicted target's distance relative to the radar, the target's azimuth angle relative to the radar and the target's elevation angle relative to the radar; represents the 0th, 1st and 2nd elements of the target state matrix , i.e. the observed target's distance relative to the radar, the target's azimuth angle relative to the radar and the target's elevation angle relative to the radar; represents the difference threshold value, which is a three-element matrix including the distance difference threshold value, the azimuth angle difference threshold value and the elevation angle difference threshold value. As long as one of the elements exceeds the corresponding difference threshold value, the matching fails and the two are not the same target.

[0122] If there are multiple targets satisfying formula (18), the target with the smallest is taken as the new observation state matrix to perform Kalman filtering. If there is no target satisfying formula (18), the is taken as the new observation state matrix (i.e. the optimal estimation of the target state at the current time before the last observation is taken as the observed target state value) to perform Kalman filtering. If the target state prediction and the observed target state still cannot satisfy formula (18) for three times in succession, the target is considered to be lost.

[0123] Taking the unmanned aerial vehicle target and the air floating object target as examples, Figures 5 to 7 a flow chart showing the identification and tracking process is shown, Figure 5comprising steps C1 to C6, Figure 6 comprising steps C7 to C15, Figure 7 comprising steps C16 to C31. As Figures 5 to 7 shown, the specific process of target recognition and tracking includes:

[0124] Step C1: Obtain meteorological radar echo signals;

[0125] Step C2: Process the meteorological radar echo signals to obtain a range-azimuth-spectrum distance map;

[0126] Step C3: Map the range-azimuth-spectrum distance map into a three-channel image;

[0127] Step C4: Input the three-channel image into a target recognition model to obtain a target recognition result;

[0128] Step C5: Determine whether a target has not been recognized, if so, end the current identification and tracking process; if not, proceed to step C6;

[0129] Step C6: Determine whether it is the first identification, if so, proceed to step C7; if not, proceed to step C15;

[0130] Step C7: Create an empty tracking queue and iterate through the recognized targets starting from k = 0;

[0131] Step C8: Determine whether the confidence of the kth target is greater than a confidence threshold, if so, proceed to step C9; if not, proceed to step C14;

[0132] Step C9: Determine whether the speed of the kth target is greater than a speed threshold, if so, proceed to step C10; if not, proceed to step C11;

[0133] Step C10: Label the kth target as a drone and proceed to step C12;

[0134] Step C11: Label the kth target as a floating object and proceed to step C12;

[0135] Step C12: Determine that it is a new target, convert the detection box of the kth target to a radar coordinate system, and initialize the tracking parameters of the kth target;

[0136] Step C13: Perform tracking prediction and update the tracking parameters, and add the state correction prediction value and tracking parameters to the tracking queue;

[0137] Step C14: Increment the target number k by 1 and proceed to step C15;

[0138] Step C15: judge whether the value of target number k exceeds the number of target detection frames, if yes, end the current identification and tracking process; if no, go to step C8;

[0139] Step C16: traverse the identified targets from k=0;

[0140] Step C17: judge whether the confidence of the kth target is greater than the confidence threshold, if yes, go to step C18; if no, go to step C26;

[0141] Step C18: judge whether the speed of the kth target is greater than the speed threshold, if yes, go to step 19; if no, go to step C20;

[0142] Step C19: mark the target as a UAV, and go to step C21;

[0143] Step C20: mark the target as a kiter, and go to step C21;

[0144] Step C21: convert the detection frame of the kth target to the radar coordinate system;

[0145] Step C22: use the state observation value of the kth target to modify the predicted value of all targets in the tracking queue, judge whether the difference is greater than the difference threshold, if yes, go to step C23; if no, go to step C24;

[0146] Step C23: determine that the observed target is a new target, initialize the tracking parameters of the target, perform tracking prediction, add the state modified predicted value and tracking parameters to the tracking queue, and go to step C26;

[0147] Step C24: select the target with the smallest difference between the state modified predicted value and the state observation value in the tracking queue for matching;

[0148] Step C25: use the state modified predicted value and the corresponding tracking parameters of the matched target to perform tracking prediction and update the tracking parameters; use the updated state modified predicted value and tracking parameters of the target to update the state modified predicted value and tracking parameters of the matched target in the tracking queue;

[0149] Step C26: increment the target number k by 1;

[0150] Step C27: judge whether the value of target number k exceeds the number of detection frames, if yes, go to step C28; if no, go to step C18;

[0151] Step C28: judge whether there is a target in the tracking queue that has not been matched with the current observation identification result, if yes, go to step C29; if no, end the current identification and tracking process;

[0152] Step C29: add 1 to the target mismatch number j;

[0153] Step C30: Determine whether the target mismatch number j is greater than or equal to 3. If so, proceed to step C31; if not, end the current identification and tracking process;

[0154] Step C31: Determine that the target is lost, remove the target from the tracking queue, and end the current identification and tracking process.

[0155] Figure 8 The Kalman filter tracking curve of the UAV target is shown. Figure 9 The actual UAV target tracking effect diagram is shown. Figure 9 The numbers in the column represent the drone number. Figure 8 and Figure 9 It can be seen that the tracking method of the present invention can accurately track the observation path of the target with high tracking accuracy.

[0156] The present invention achieves an average LSS target recognition accuracy exceeding 90% in complex meteorological and non-meteorological clutter environments. By utilizing range-azimuth-spectral mapping and multi-spectral moment feature fusion, the present invention can effectively capture and identify LSS targets even when their volume is too small relative to the radar resolution, resulting in low parameter discrimination with traditional radars. This maximizes the potential of traditional weather radar data.

[0157] The present invention maps the range-azimuth-spectrum map into a three-channel image and inputs it into a trained deep learning model for inference, avoiding the traditional high-computational signal processing or convolution operations in multiple radial directions, greatly improving the processing speed and real-time performance. The actual measurement time is basically the same as that of conventional meteorological processing.

[0158] The present invention creatively introduces Kalman filtering and a multi-tracking link establishment mechanism, which can achieve continuous and stable tracking of LSS targets even in complex scenarios with relatively low radar observation refresh rates or close target spacing (such as swarms of flying drones).

[0159] This invention, for the first time, enables a weather radar to identify and track LSS targets without changing its hardware configuration and weather observation mode, thus realizing the innovative application of "one device for multiple uses" and overcoming the inherent limitation of traditional weather radars that they cannot effectively detect LSS targets.

[0160] Example 2

[0161] An embodiment of the present invention further provides a weather radar device, which includes: a memory, a processor, and a computer program / instruction stored in the memory, and the processor executes the computer program / instruction to implement the target recognition method in the embodiment of the present invention.

[0162] Although not shown, the weather radar apparatus includes a processor that can perform various appropriate operations and processes in accordance with programs and / or data stored in a read only memory (ROM) or programs and / or data loaded from a storage section into a random access memory (RAM). The processor can be a multi-core processor or can include a plurality of processors. In some embodiments, the processor can include a general-purpose main processor and one or more special-purpose coprocessors, such as a central processing unit, a graphics processing unit (GPU), a neural processing unit (NPU), a digital signal processor (DSP), and the like. In the RAM, various programs and data required for apparatus operation are also stored. The processor, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0163] The above processor and memory are used in common to execute programs / instructions stored in the memory, which, when executed by a computer, can implement the methods, steps, or functions described in the above embodiments.

[0164] Although not shown, the present embodiments also provide a computer-readable storage medium having stored thereon computer programs / instructions that, when executed by a processor, implement the target identification method in the present embodiments.

[0165] The computer-readable storage medium includes permanent and non-permanent, removable and non-removable media, which can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0166] The above disclosure is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or modifications within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A target recognition method characterized by, The identification method comprises: calling a target identification model, the target identification model being a trained deep learning model, a training sample of the deep learning model being a three-channel image obtained by mapping a range-azimuth-spectrum distance map, the range-azimuth-spectrum distance map being a processing result of a weather radar echo signal; identifying the three-channel image to be identified by using the target identification model to obtain each target detection frame and a confidence thereof; filtering out the target detection frame with a confidence lower than a confidence threshold to obtain a high-confidence target detection frame; classifying the high-confidence target detection frame based on a target speed and a speed threshold to obtain different-class target detection frames.

2. The object recognition method of claim 1, wherein, The identification method further comprises a target tracking process, comprising: calculating a target state parameter at a current observation time based on a current identification and classification result and a radar echo signal processing result; calculating a target state correction prediction value at a next observation time by using a Kalman filtering algorithm based on the target state parameter at the current observation time and updating a state covariance matrix; obtaining a next identification and classification result and a radar echo signal processing result; calculating a target state parameter at the next observation time based on the next identification and classification result and the radar echo signal processing result; matching the target state correction prediction value at the next observation time and the target state parameter at the next observation time to realize target tracking.

3. The object recognition method of claim 2, wherein, The target state parameter comprises a target range relative to a radar, a target azimuth angle relative to the radar, a target elevation angle relative to the radar, a target radial speed, a target azimuth angle speed and a target elevation angle speed; and a calculation formula of the target range relative to the radar and the target azimuth angle relative to the radar is: ； ； in, Indicates the distance of the kth target relative to the radar at the nth observation; Indicates the range resolution on the range-azimuth-spectrum diagram; ( , ) represents the coordinates of the detection box center of the k-th target in the radar coordinate system during the n-th recognition and classification; represents the elevation angle of the kth target relative to the radar during the nth observation; represents the azimuth of the kth target relative to the radar at the nth observation; Indicates the azimuth resolution on the range-azimuth-spectrum diagram.

4. The object recognition method of claim 2, wherein, calculating the target state correction prediction value at the next observation time by using the Kalman filtering algorithm comprises: calculating a target state prediction value at the next observation time according to the target state parameter at the current observation time and a state transition matrix; calculating a prediction covariance matrix at the current observation time according to the target state parameter at the current observation time, a state covariance matrix and a process noise covariance matrix; calculating a Kalman gain according to the covariance matrix at the current observation time, an observation matrix and an observation noise covariance matrix; calculating the target state correction prediction value at the next observation time according to the target state prediction value at the next observation time, the Kalman gain, the target state parameter at the current observation time and the observation matrix.

5. The object recognition method of claim 4, wherein, An update formula of the state covariance matrix is: ； wherein, denotes the state covariance matrix of the kth target at the n+1th observation; denotes the identity matrix; denotes the Kalman gain of the kth target; denotes the observation matrix; denotes the variance matrix of the kth target.

6. The object recognition method of claim 2, wherein, matching the target state correction prediction value at the next observation time and the target state parameter at the next observation time comprises: calculating a difference value between the target state correction prediction value at the next observation time and the target state parameter at the next observation time; determining whether the difference value and a difference threshold value are of a same target, if yes, matching is successful, and if no, matching fails.

7. The object recognition method of claim 6, wherein, When the target state correction prediction value at the next observation time and the target state parameter at the next observation time fail to match, it is determined whether a state parameter prediction correction value of a corresponding target has failed to match for a plurality of times, if yes, the target is lost, and if no, the target state correction prediction value at the next observation time is taken as the target state parameter at the next observation time to realize target tracking.

8. A weather radar apparatus comprising a memory, a processor, and a computer program / instructions stored on the memory, wherein, The processor executes the computer program / instruction to implement the target identification method as claimed in any one of claims 1-7.

9. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instruction is executed by the processor to implement the target identification method as claimed in any one of claims 1-7.

Images