A navigation radar-based water surface target detection and tracking method

Abstract

A surface target detection and tracking method based on navigation radar, relating to the field of environmental perception technology, addresses the problem of mismatches that easily occur when multiple targets are close together in radar echo images in existing technologies. This application first uses an image region connectivity algorithm to aggregate echoes with low fragmentation; secondly, it combines a map matching algorithm to accurately identify and remove echoes with high fragmentation, such as land and islands; thirdly, it uses the Kuhn-Munkres (KM) algorithm, which combines echo morphology and location, to match echoes belonging to the same target in two consecutive echo frames; finally, it uses the EMD algorithm with an ARIMA prediction model sequence to predict targets whose echoes disappear. This method improves the anti-interference capability and matching accuracy of navigation radar in target detection and tracking, thus solving the problem of mismatches easily occurring when multiple targets are close together in radar echo images.

Description

Technical Field

[0001] This invention relates to the field of environmental perception technology, specifically to a method for detecting and tracking water surface targets based on navigation radar. Background Technology

[0002] With the rapid development of artificial intelligence and related fields, the rational use of unmanned surface vessels (USVs) to replace manned surface ships can accomplish a series of complex and dangerous tasks in aquatic environments where human control is impossible. The prerequisite for USVs to perform missions is safe autonomous navigation capability. For USVs working in their respective environments and tasks, high-precision environmental perception is a crucial prerequisite for safe navigation. Currently, the main methods for USVs to detect the water surface environment include navigation radar, lidar, and visual cameras. LiDAR and visual cameras have relatively small detection ranges, while navigation radar can achieve detection at longer distances and its detection accuracy is less affected by weather and sea conditions; therefore, navigation radar is mostly used for water surface environmental perception. To ensure that USVs have good perception capabilities of their surroundings, further research is needed on the target detection and tracking performance of navigation radar.

[0003] In theory, navigation radar can accurately acquire the reflected echoes of targets, and unmanned surface vessels (USVs) can obtain information about their surrounding working environment by receiving radar echo data. However, in practical use, due to the complexity of the marine environment, the intensity and completeness of the echoes received by radar are unstable, easily leading to fragmented areas in the echo image. This necessitates echo aggregation before target detection. The DBSCAN algorithm, commonly used for clustering fragmented targets, has a computational complexity that increases with the amount of data, resulting in slow operation. Using convolutional neural networks for target detection requires establishing a corresponding dataset and places high demands on hardware. When navigation radar tracks a target, it needs to match echoes belonging to the same target in multiple consecutive frames of echo images. The KM algorithm, based on positional relationships, is commonly used, but mismatches are prone to occur when multiple targets are close together in the radar echo image. Summary of the Invention

[0004] The purpose of this invention is to address the problem of mismatches that easily occur when multiple targets are located close to each other in radar echo images in the prior art, and to propose a surface target detection and tracking method based on navigation radar.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] A method for detecting and tracking water surface targets based on navigation radar includes the following steps:

[0007] Step 1: Receive navigation radar and inertial navigation data, and solve the navigation radar and inertial navigation data to obtain radar echo data. Then, preprocess the radar echo data. The preprocessing includes image correction, coordinate transformation and filtering. The coordinate transformation is used to map the radar echo data from the polar coordinate system to the rectangular coordinate system.

[0008] Step 2: Use the Two-pass algorithm to process the preprocessed radar echo data to obtain a labeled radar echo image;

[0009] Step 3: Based on the range of the navigation radar, acquire a map covering that range, then convert the geographic coordinates of the map to a rectangular coordinate system using Mercator projection, then perform a translation operation to convert it to the screen coordinate system, and finally normalize the scale of the converted map and radar echo image.

[0010] Step 4: Use a distance matching algorithm based on adaptive threshold segmentation to match the normalized map and radar echo images, identify echoes belonging to land or islands, and determine the corresponding label, i.e., batch number, for echoes belonging to land or islands.

[0011] Step 5: Obtain the specific location and contour information of the echo corresponding to the target batch number, use the binary KM matching algorithm to match the echoes of the same target in two consecutive frames, and assign them the same batch number;

[0012] Step 6: Determine whether the newly appearing echo is a target or clutter. If it is clutter, do not assign a new batch number. If it is a target, assign a new batch number to the newly appearing target.

[0013] Step 7: Detect the distance between the target and the unmanned surface vessel (USV), and calculate the target's heading and speed based on the echo information of consecutive frames. Then, determine whether there are any glitches or interferences in the distance between the target and the USV, the target's heading, and the target's speed. If there are glitches or interferences, perform EMD algorithm decomposition to obtain the IMF (Information Function Factor), and remove the highest frequency IMF from the IMF. Then, reconstruct the signal from the remaining IMF to obtain the distance between the target and the USV, the target's heading, and the target's speed after removing glitches or interferences. Otherwise, determine whether the echo corresponding to a target that has been assigned a batch number and has appeared for more than 5 consecutive frames has disappeared. When the target disappears, use the ARIMA model to predict the position of the disappeared target. When the echo corresponding to the disappeared target appears in the next 3 frames, match the predicted position of the target with the position where the target appeared. If the matching result is consistent, assign the same batch number; otherwise, remove the disappeared target, thus completing the target detection and tracking.

[0014] Step 8: After removing the missing target, predict the position of all echoes in the next 3 consecutive frames, and then use the EMD algorithm to optimize the current distance, heading and speed of all targets from the unmanned surface vessel.

[0015] Furthermore, the conversion of geographic coordinates to a rectangular coordinate system is expressed as follows:

[0016]

[0017] Where (X,Y) are the latitude and longitude coordinates of a point P on the Earth's surface, (x,y) are the rectangular coordinates corresponding to (X,Y), K is the radius of the reference latitude circle, L0 is the longitude of the origin of the Earth's surface, and the latitude of the origin is 0.

[0018] Furthermore, the dimensions of the converted map are expressed as follows:

[0019]

[0020] Among them, P c P and R represent the dimensions of the radar image and map in the screen coordinate system, respectively. c R represents the actual distance in the radar image, and R represents the actual distance in the map.

[0021] Furthermore, the specific steps in step 4 of matching the normalized map and radar echo image using a distance matching algorithm based on adaptive threshold segmentation are as follows:

[0022] Step 4.1: Use the pixels of the unmanned surface vessel's location in the radar echo image and the normalized map as matching points to achieve preliminary matching between the radar echo image and the normalized map, and obtain a matching image.

[0023] Step 4.2: Binarize the radar echo image. The normalized map is then binarized using an adaptive threshold segmentation algorithm, which specifically includes:

[0024] Step 4.2.1: Find the extreme gray values ​​of the normalized map, i.e., the maximum gray value l. max (x,y) and minimum gray level l min (x,y), calculate the maximum gray level l of the normalized map. max (x,y) and minimum gray level l min The mean G1 of (x,y);

[0025]

[0026] Step 4.2.2: Using G1 as the initial threshold for the iterative method, and using the initial threshold as the boundary threshold, find the extreme gray values ​​of all pixels in the normalized map that are greater than G1 and the extreme gray values ​​of all pixels that are less than G1, and obtain the average value G of all extreme gray values ​​of pixels greater than G1. 11 And the average of the grayscale extreme values ​​of all pixels less than G1, G 12 Thus, the second threshold G2 is obtained;

[0027]

[0028] Step 4.2.3: Update the threshold to G2, and repeat step 4.2.2 until |G i -G i-1 When |<0.1, stop and select G. i Binarize the normalized map;

[0029] Step 4.3: Extract the contours of the radar echo image and map obtained from Step 4.2 using the Canny operator;

[0030] Step 4.4: Based on the image matching in Step 4.1, and combined with the contours of the radar echo image and the map, select the minimum bounding rectangles of the radar echo image and the map respectively, and use the larger of the two minimum bounding rectangles as the matching area of ​​the radar echo image and the matching area of ​​the map respectively.

[0031] Step 4.5: Calculate the Hausdorff distance between the matching region of the radar echo image and the matching region of the map, and select the matching point with the highest similarity for matching.

[0032] Furthermore, the specific steps of step 4.5 are as follows:

[0033] Assume that the pixels in the matching region of the radar echo image and the matching region of the map constitute sets A and B, respectively:

[0034] A = {a1, a2, ..., a} p B = {b1, b2, ..., b} q}

[0035] Among them, a p Let b represent the p-th pixel in set A. q This represents the q-th pixel in set B;

[0036] The Hausdorff distance between A and B is defined as:

[0037] H(A,B)=max(h(A,B),h(B,A))

[0038]

[0039]

[0040] Where i = 1, 2...p, j = 1, 2...q, h(B,A) represents the Hausdorff from B to A, and h(A,B) represents the Hausdorff from A to B.

[0041] Furthermore, in step 5, the specific steps of matching echoes belonging to the same target in two consecutive frames using the binary KM matching algorithm are as follows:

[0042] Step 5.1: Assume the specific coordinates of the echo from the previous frame are (x1, x2, x3, ..., x...). n The corresponding echo area is (s1, s2, s3, ..., s). n The specific coordinates of the echo in the next frame are (y1, y2, y3, ..., y2). n The corresponding echo area is (t1, t2, t3, ..., t). n Find the minimum sum of the Euclidean distances of the specific locations of the echoes in the two frames, sum1, and the minimum sum of the Euclidean distances of the echo areas, sum2, respectively.

[0043]

[0044] Step 5.2: Based on sum1 and sum2, and using the KM algorithm for matching, obtain the matching result.

[0045] Furthermore, the specific steps for performing EMD algorithm decomposition to obtain the IMF are as follows:

[0046] Step 7.2.1: First, acquire the signal sequence of information regarding the target's distance, heading, or speed from the unmanned surface vessel, represented as:

[0047]

[0048] Where r is the signal margin, IMF k (t) represents the k-th intrinsic mode component, and K represents the total number of intrinsic mode components;

[0049] Then, all the maxima and minima of s(t) are obtained. The maxima and minima are fitted with cubic spline fitting functions to obtain the upper and lower envelope curves of s(t), and the mean m(t) of the upper and lower envelope curves is obtained.

[0050] Step 7.2.2: Calculate the components judge Is it an IMF? If it is an IMF, then denote IMF as I. i (t), and perform step 7.2.3, if it is not the IMF, then let Then proceed with step 7.2.1 until the IMF is obtained:

[0051]

[0052] judge To be considered an IMF member, two conditions must be met:

[0053] (1) The difference between the number of zeros and the number of extreme points in the signal is no greater than 1;

[0054] (2) The mean of the upper and lower envelopes of the signal at any point is 0;

[0055] Step 7.2.3: I i Let r(t) be the i-th IMF component, and calculate the residual r(t):

[0056]

[0057] Step 7.2.4: Stop the decomposition when the residual r(t) is a constant or a monotonic function.

[0058] Furthermore, the specific steps in step 4.3 of extracting the radar echo image and map contour using the Canny operator are as follows:

[0059] The Canny operator is used to extract the contours of radar echo images and maps respectively. The extracted contours are then filtered based on the mean area of ​​the contours, with contours greater than the mean area being retained and those smaller than the mean area being discarded.

[0060] Furthermore, the filtering process in step 1 includes isolated noise point filtering and echo threshold filtering.

[0061] Furthermore, the image correction employs an image sliding window correction method.

[0062] The beneficial effects of this invention are:

[0063] This application first uses an image region connectivity algorithm to aggregate echoes with low fragmentation; secondly, it combines a map matching algorithm to accurately identify and remove echoes with high fragmentation, such as land and islands; thirdly, it uses the Kuhn-Munkres (KM) algorithm, which combines echo morphology and location, to match echoes belonging to the same target in two consecutive echo images; finally, it uses the EMD algorithm with an ARIMA prediction model sequence to predict targets whose echoes disappear, thereby improving the anti-interference capability and matching accuracy of navigation radar in the target detection and tracking process, and thus solving the problem of mismatch when there are multiple targets with small positions in the radar echo image.

[0064] Compared to echo aggregation methods achieved through clustering algorithms, this application has a shorter runtime, fewer algorithm parameters, and is simpler and more convenient to implement. Furthermore, compared to traditional target tracking algorithms, this application improves matching accuracy and anti-interference capabilities. Overall, this application addresses the issue of low target information accuracy obtained after unmanned surface vessels (USVs) are equipped with navigation radar for target detection and tracking. Attached Figure Description

[0065] Figure 1 Image before correction of radar echo image;

[0066] Figure 2 The image after correction of the radar echo image;

[0067] Figure 3 This is a schematic diagram of the filtered radar echo;

[0068] Figure 4 A schematic diagram of radar echoes for regional connectivity;

[0069] Figure 5 This is a schematic diagram of map scale normalization;

[0070] Figure 6 This is a schematic diagram of the adaptive iterative threshold segmentation method;

[0071] Figure 7 This is a schematic diagram of direct binarization of a map;

[0072] Figure 8 This is a schematic diagram of map adaptive threshold segmentation binarization;

[0073] Figure 9 This is a schematic diagram of the result after matching the echo image with the map.

[0074] Figure 10 Match the set to a bipartite graph;

[0075] Figure 11 This is a schematic diagram of the echo from the previous frame;

[0076] Figure 12 This is a schematic diagram of the echo in the current frame;

[0077] Figure 13 This is a trajectory diagram of the experimental scenario.

[0078] Figure 14 The trajectory diagram for experimental scenario two;

[0079] Figure 15 The experimental scenario features a three-trajectory diagram;

[0080] Figure 16 This is a four-trajectory diagram of the experimental scenario;

[0081] Figure 17 This is a comparison chart of detection distance information for Scene 1;

[0082] Figure 18 A comparison chart of distance information errors for scene 1;

[0083] Figure 19 A comparison chart of heading information for scenario one;

[0084] Figure 20A comparison chart of heading information errors for scenario one;

[0085] Figure 21 A comparison chart of flight speed information detected in Scene 1;

[0086] Figure 22 A comparison chart of flight speed information errors for Scene 1;

[0087] Figure 23 This is a comparison chart of detection distance information for Scene 2;

[0088] Figure 24 This is a comparison chart of distance information errors in Scene 2;

[0089] Figure 25 Comparison chart of heading information for scenario two;

[0090] Figure 26 This is a comparison chart of heading information errors for scenario two.

[0091] Figure 27 A comparison chart of flight speed information detected in Scenario 2;

[0092] Figure 28 This is a comparison chart of flight speed information errors in scenario two.

[0093] Figure 29 This is a comparison chart of detection distance information for Scene 3;

[0094] Figure 30 A comparison chart of distance information errors in three scenes;

[0095] Figure 31 A comparison chart of flight speed information detected in scenario three;

[0096] Figure 32 A comparison chart of flight speed information errors in three scenarios;

[0097] Figure 33 This is a comparison chart of detection distance information for Scene 4;

[0098] Figure 34 A comparison chart of distance information errors in scene four;

[0099] Figure 35 A comparison chart of heading information for scenario four;

[0100] Figure 36 A comparison chart of heading information errors for scenario four;

[0101] Figure 37 A comparison chart of flight speed information detected in scenario four;

[0102] Figure 38 This is a comparison chart of flight speed information errors in scenario four. Detailed Implementation

[0103] It should be noted that, where there is no conflict, the various embodiments disclosed in this application can be combined with each other.

[0104] Specific implementation method one: Refer to Figure 1 This embodiment describes a method for detecting and tracking water surface targets based on navigation radar, which includes the following steps:

[0105] Step 1: Receive navigation radar and inertial navigation data according to the UDP protocol of the navigation radar and inertial navigation data transmission on the unmanned surface vessel, and solve the navigation radar and inertial navigation data to obtain radar echo data. Then, preprocess the radar echo data. The preprocessing includes image correction, coordinate transformation and filtering. The coordinate transformation is used to map the radar echo data from the polar coordinate system to the rectangular coordinate system.

[0106] Step 2: Use the Two-pass algorithm to process the preprocessed radar echo data to obtain a labeled radar echo image;

[0107] Step 3: Based on the range of the navigation radar, acquire a map covering that range, then convert the geographic coordinates of the map to a rectangular coordinate system using Mercator projection, then perform a translation operation to convert it to the screen coordinate system, and finally normalize the scale of the converted map and radar echo image.

[0108] Step 4: Use a distance matching algorithm based on adaptive threshold segmentation to match the normalized map and radar echo images, identify echoes belonging to land or islands, and determine the corresponding label, i.e., batch number, for echoes belonging to land or islands.

[0109] Step 5: Obtain the specific location and contour information of the echo corresponding to the target batch number, use the binary KM matching algorithm to match the echoes of the same target in two consecutive frames, and assign them the same batch number;

[0110] Step 6: Determine whether the newly appearing echo is a target or clutter. If it is clutter, do not assign a new batch number. If it is a target, assign a new batch number to the newly appearing target.

[0111] Step 7: Detect the distance between the target and the unmanned surface vessel (USV), and calculate the target's heading and speed based on the echo information of consecutive frames. Then, determine whether there are any glitches or interferences in the distance between the target and the USV, the target's heading, and the target's speed. If there are glitches or interferences, perform EMD algorithm decomposition to obtain the IMF (Information Function Factor), and remove the highest frequency IMF from the IMF. Then, reconstruct the signal from the remaining IMF to obtain the distance between the target and the USV, the target's heading, and the target's speed after removing glitches or interferences. Otherwise, determine whether the echo corresponding to a target that has been assigned a batch number and has appeared for more than 5 consecutive frames has disappeared. When the target disappears, use the ARIMA model to predict the position of the disappeared target. When the echo corresponding to the disappeared target appears in the next 3 frames, match the predicted position of the target with the position where the target appeared. If the matching result is consistent, assign the same batch number; otherwise, remove the disappeared target, thus completing the target detection and tracking.

[0112] Step 8: After removing the missing target, predict the position of all echoes in the next 3 consecutive frames, and then use the EMD algorithm to optimize the current distance, heading and speed of all targets from the unmanned surface vessel.

[0113] Step 1: Receive data collected by navigation radar and inertial navigation according to the protocol, and perform image correction, coordinate transformation and filtering on the collected data;

[0114] In step one, based on the UDP protocol for transmitting navigation radar and inertial navigation data on the unmanned surface vessel (USV), a corresponding interface is designed for data reception. The received data is then processed to obtain the original data information. Due to changes in the USV's own motion state and the complexity of its environment, the radar echo data exhibits angular variations, necessitating radar image correction. This invention utilizes an image sliding window correction method, which ensures the continuity of the radar image, avoids blind spots or overlaps, and minimizes the impact of angular shifts. Using two consecutively acquired radar echo image data, combined with inertial navigation information, a north-direction radial pulse ordinal number is searched in each of the preceding and following radar echo images, and the middle portion is extracted to form a new complete radar image. Since the radial pulse number in the new image differs from that in the original image, affecting subsequent coordinate transformations, random interpolation or deletion of the radial data is required to maintain consistency. The radial ordinal numbers selected for interpolation and deletion should be evenly distributed across the parts of the USV with significant heading changes to minimize the cumulative effect of angular shifts caused by data volume variations. A comparison of the radar echoes before and after correction is shown below. Figure 1 and Figure 2 As shown.

[0115] After correcting the echo image, since the original echo image is data in a polar coordinate system, and the connected component algorithm requires data in a Cartesian coordinate system, the radar echo image needs to be pre-transformed. This invention employs a coordinate mapping method, directly mapping points in the polar coordinate system to their unique corresponding points in the Cartesian coordinate system. By iterating through each point, the mapping relationship is searched one by one, ultimately obtaining the mapping matrix between the coordinate systems. This algorithm requires a relatively long time for the first mapping, but can be reused subsequently for rapid coordinate system transformation.

[0116] After coordinate transformation, the echo data still contains some interference from land, ocean waves, and isolated noise generated during the transformation process, requiring filtering. Filtering of radar echo images mainly includes isolated noise filtering and echo threshold filtering. The erosion-dilation algorithm can filter out isolated noise in the echo image; statistical algorithms, by calculating the mean and standard deviation of all echoes using a moving kernel and determining a suitable threshold according to certain criteria, can filter out threshold noise in the echo image. The filtered radar echo image is shown below. Figure 3 As shown.

[0117] Step 2: Perform region connectivity operation on the preprocessed radar echo image to aggregate echoes with less fragmentation.

[0118] In step two, after preprocessing the original echo image, the resulting radar echo map initially displays the status of the water surface target. Due to the radar's own performance and environmental influences, different echoes may represent the same target. To determine the specific status of the water surface target, the echoes need to be aggregated. Connected components divide the echo map according to pixels, marking adjacent foreground regions with the same pixel values. Each individual connected component forms a labeled block. By acquiring the geometric parameters such as the contour, circumscribed rectangle, and centroid of the echoes in different blocks, the information of the corresponding target is obtained.

[0119] The two-pass method scans the image twice to identify and mark all connected regions, making it suitable for connectivity analysis of echo images. An 8-way neighborhood connectivity is achieved using a row-by-row scanning approach. During the first scan, each pixel P is traversed starting from the top left corner of the echo image. If the value of P is not 0, the values ​​of its 8 neighbors are checked.

[0120] The specific implementation process of two-pass scanning is as follows:

[0121] (1) If all values ​​in a domain are 0, then label them with a value (incrementing from 0);

[0122] (2) If there is a non-zero pixel value in the neighborhood, take the smallest value in the neighborhood and assign it to P. Use the disjoint set to establish a connection between the pixels in the neighborhood and assign them the same label value.

[0123] (3) During the scanning process, pixels within the same connected region may be assigned one or more different label values. Therefore, it is necessary to merge these pixels that belong to the same connected region but have different label values. The second-stage scanning is to set an appropriate threshold to group pixels with equivalent label values ​​that have an equal relationship into a connected region and assign them the same label value.

[0124] During the scanning process, foreground pixels (target) in the echo image are marked as 1, and background pixels (water surface) are marked as 0. The final echo image, after region connectivity and labeling, is as follows: Figure 4 As shown.

[0125] Step 3: Perform scale normalization on the echo image and map so that they can be applied to image matching operations;

[0126] In step three, the map itself is in a geodetic coordinate system, and the coordinates of various landmarks in the map data are marked with latitude and longitude. In the screen coordinate system, the coordinates of the landmarks need to be displayed as pixels, so the geographic coordinates need to be converted to a rectangular coordinate system through Mercator projection when the map is displayed. Considering the Earth as an ellipsoid with the origin latitude at 0 and the origin longitude at L0, assuming that the latitude and longitude coordinates of a point P on the sphere are (X,Y), its corresponding rectangular coordinates are (x,y), and the solution for (x,y) is shown in equation (1):

[0127]

[0128] Where K is the radius of the reference latitude circle, and the value of K is calculated as shown in equation (2):

[0129]

[0130] Converting the map to a Cartesian coordinate system requires translation and other operations to convert it to the screen coordinate system, so that the map and radar echo image can be displayed in the same coordinate system. The prerequisite for fusing radar images and maps in the same coordinate system is that the image scale and display direction are consistent. Both the echo image and the map are displayed facing due north, but the map displays a large area of ​​static sea information, while the echo image displays sea information within the radar scan range at a certain moment. The two display scales differ significantly when on the same screen. Therefore, it is necessary to normalize the size of the multi-scale images. By using the relationship between actual distance and pixel points, the map size corresponding to the current echo image size is calculated using equation (3):

[0131]

[0132] Where P c P and R represent the dimensions of the radar image and map in the screen coordinate system, respectively. c R represents the actual distance in the radar image, and R represents the actual distance on the map. The map after coordinate transformation and scale normalization is shown below. Figure 5 As shown.

[0133] Step 4: Use a distance matching algorithm based on adaptive threshold segmentation to match the map and echo images, identify echoes belonging to land and islands, and further determine the batch number of the target echo.

[0134] In step four, the matching of the echo image and the map is divided into two stages. The first stage uses information provided by the inertial navigation system in the echo image to obtain the current location of the unmanned surface vessel (USV). Its latitude and longitude correspond to a single pixel on the map. The corresponding pixel in the echo image and the map is used as the matching point to achieve a preliminary matching between the echo image and the map. However, in practice, there are calculation errors in establishing the relationship between latitude and longitude and the pixel; furthermore, the inertial navigation measurement process also introduces systematic errors. Therefore, the results obtained by directly matching using the USV's latitude and longitude information are not entirely satisfactory. The second stage uses a Hausdorff distance matching algorithm with adaptive threshold segmentation to correct the matching results of the first stage. The specific implementation steps are as follows:

[0135] (1) Binarization of Echo Images and Maps: Echo images can be directly binarized, but maps have rich colors. Directly expressing the grayscale value of the image through brightness values ​​results in little difference in grayscale values ​​in areas where land and sea meet, affecting feature extraction from the map. Therefore, a reasonable threshold needs to be selected to perform the map binarization operation. In the map, since the ocean is a large blue area, it can be used as a white background, while all obstacles on the sea surface can be used as a black foreground. Different maps require different thresholds for binarization. This invention proposes an adaptive threshold segmentation algorithm to satisfy map binarization. First, the extreme grayscale values ​​of the map are found, and the maximum grayscale value l of the map is calculated. max (x,y) and minimum gray level l min The mean G1 of (x,y) is shown in equation (4):

[0136]

[0137] l min (x, y) represents the grayscale value of each pixel. G1 is then used as the first coarse threshold in the iterative method. Next, using this first threshold as a boundary threshold, the extreme grayscale values ​​of all pixels in the image that are greater than or less than G1 are found, and their corresponding averages are calculated to obtain G. 11 and G 12 The second rough threshold G2 is obtained as shown in equation (5):

[0138]

[0139] Finally, the threshold is updated to G2. Then, the mean grayscale value of all pixels in the image that is greater than G2 and less than G2 is found to obtain the precise threshold G3. This threshold is updated to G3, and step three is executed again until |G i -G i-1 When |<&, the algorithm stops, and G is selected. i Binarize the map. The specific algorithm flow is as follows: Figure 6 As shown, Figure 5 The results of the direct binarization self-adaptive threshold segmentation algorithm are, for example... Figure 7 and Figure 8 As shown.

[0140] (2) Extracting contours from echo images and maps: Based on the binary image, the edge information of the image is extracted as the image feature information. The Canny edge detection operator is selected to extract image features. However, since there is a lot of contour information in the echo image and the map, and many echo contours in the echo image are different from the target contours in the real environment, directly extracting the contours of the two images for matching will increase the computation time and reduce the registration rate. Therefore, it is necessary to set a threshold for the contour size, filter all the obtained contours according to the threshold, and delete the contours with contour values ​​less than the threshold, so that the processed image has obviously similar large contours.

[0141] (3) Matching the echo image with the map: To further improve the running speed and ensure real-time performance, region filtering is added on the basis of contour filtering. The matching region is selected as the neighborhood of the reference pixel point when the initial matching is performed, and the specific region threshold is determined according to the proportion of contour pixel values ​​in the whole image. After determining the matching region, it is assumed that the pixels in the matching regions of the map and echo image after contour extraction form sets A and B respectively, as shown in Equation (6):

[0142] A = {a1, a2, ..., a} p B = {b1, b2, ..., b} q} (6)

[0143] The minimum Hausdorff distance between A and B is defined as shown in equation (7):

[0144]

[0145] By calculating the Hausdorff distance to the matching area, the optimal matching state can be obtained. Registration is then performed based on the calculated matching points to correctly correct the initial matching result. After matching, echoes belonging to land and islands in the echo image are removed from the map, retaining only echoes of water targets. This reduces resource waste in subsequent work and improves detection performance. The final map matching result is as follows: Figure 9 As shown.

[0146] Step 5: Obtain the specific location and contour information of the echo corresponding to the target batch number, use the binary KM matching algorithm to match the echoes of the same target in two consecutive frames, and assign them the same batch number.

[0147] In step five, once the radar acquires a new echo image, all targets are refreshed and cannot be linked to the previous frame. Associating two consecutive echo images is crucial for obtaining the target's speed and heading. Labels corresponding to echoes from land, islands, etc., are removed, leaving only the label values ​​for water targets. The echo center coordinates are extracted as the current target location. Target matching can be achieved using the target positions in two consecutive frames combined with the KM algorithm. However, when multiple targets are close together, mismatches are prone to occur. Therefore, echo morphology is also considered as a matching factor, and setting appropriate weights can improve matching accuracy to some extent.

[0148] Assume the position coordinates of the water surface target detected in the previous frame are (x1, x2, x3, ..., x...). n The corresponding target echo area is (s1, s2, s3, ..., s). n The next frame detects the position coordinates of the target on the water surface (y1, y2, y3, ..., y2). n Find the minimum values ​​sum1 and sum2 of the sum of the Euclidean distances of the targets in the two frames, as shown in equation (8).

[0149]

[0150] Based on two matching factors, the concepts of feasible top-level labels and equal subgraphs are introduced to solve for the optimal matching of bipartite graphs. The point sets of two consecutive radar echo images constitute two sets (u, v) of the bipartite graph, as shown below. Figure 10 As shown. In the bipartite graph formed by the sets, the left side is the set u(u1,u2,…,u…). i The right side represents the set v(v1, v2, ..., v). j ), corresponding to each group of left and right connected u i and v j The weight of the connector is w ij The algorithm transforms the problem of finding the maximum weighted matching into the problem of finding a perfect matching by assigning a vertex label to each vertex. During the algorithm's execution, for any connecting line, there exists u i+v j ≥w ij The matching of two sets is successful. The specific steps for matching two sets are as follows:

[0151] First, take the w in the diagram. ij Invert the values ​​to construct the initial adjacency matrix. Transform the adjacency matrix by subtracting the minimum value of each row (column), so that each row (column) has at least one 0.

[0152] Secondly, a trial matching loop is performed to find the optimal solution:

[0153] (1) Construct the minimum number of straight lines that cover all zero elements in the adjacent matrix. The lines can only be in a single row or column and cannot span different rows or columns. If any element is covered by two lines, mark it with x*.

[0154] (2) Determine whether the loop needs to be terminated. The condition is whether the number of lines drawn in the first step is equal to the number of vertices in each set. If they are equal, terminate the loop; otherwise, continue the operation.

[0155] (3) Transform the elements in the adjacency matrix to obtain more zero elements: First, denote all elements not covered by the line as Φ, find the smallest element in Φ and denote it as k; then subtract k from all elements in Φ and add k to all elements marked with x*.

[0156] Finally, the transformed adjacency matrix is ​​output. Starting from a row or column with only one 0 element, this 0 element is marked as 0*. This ensures that every vertex in this row or column has another point for matching. Then, other elements in the row and column containing 0* are ignored. For cases where multiple 0s appear in the same column or row, the zero element in the row or column with the largest second-smallest value is matched first. After repeated selection, if 0* appears in different rows and columns, it means the maximum matching has been completed, and the vertex corresponding to 0* is the matching result. The echo matching results of two consecutive frames obtained by the KM matching algorithm are as follows: Figure 11 and Figure 12 As shown.

[0157] Step 6: Determine whether the newly appearing echo is a target or clutter, and assign a new batch number to the newly appearing target;

[0158] In step six, when the wind and waves are large, the surface environment is more variable, and the radar may detect waves as target echoes, requiring timely removal of clutter; when the true target echo is obliterated (like... Figure 11 and Figure 12(Echo 25 in the image disappears in the second frame). Target information should be retained promptly, and its possible location predicted. Since ocean waves are transient and unstable, this invention eliminates targets by calculating the number of consecutive frames the same target appears in the radar echo map. When a newly appearing target echo is displayed continuously for more than 3 frames, it is considered a new target; all others are treated as clutter and eliminated. When a target echo that has appeared for 5 or more frames suddenly disappears, its target information is retained, the target location in the next frame is predicted, and effective matching is performed when it reappears.

[0159] Step 7: Calculate the target's heading and speed based on the echo information of consecutive frames, use the EMD algorithm with added prediction model to predict the position of the target when the echo disappears, remove interference signals in the detection results, and optimize the target distance, heading and speed information output each time.

[0160] In step seven, the information obtained from detecting the target will be subject to various types of noise interference, and the acquired information needs to be optimized to improve the accuracy of detection and tracking. This invention proposes an EMD algorithm with an ARIMA prediction model to improve the performance of target tracking.

[0161] The EMD algorithm decomposes the data signal based on the time-scale characteristics of the target information measured by the navigation radar. It does not require any pre-defined basis functions and has strong adaptability. Assuming the input signal is s(t), after EDM decomposition, it will form multiple intrinsic mode functions (IMFs) and a residual component, as shown in equation (9):

[0162]

[0163] Each IMF contains specific frequency components of the original signal. The components decomposed first contain the high-frequency components of s(t), and the frequency decreases as the number of decompositions increases. Decomposition stops when the margin r is constant or a monotonic function. The decomposed IMF should meet two conditions: (1) the difference between the number of zeros and the number of extreme points in the signal cannot be greater than 1; (2) the mean of the upper and lower envelopes of any point in the signal is 0. The process of decomposing to obtain the IMF is as follows:

[0164] (1) Obtain all the maximum and minimum values ​​of s(t), and obtain the upper and lower envelope curves of s(t) by fitting the cubic spline fitting function respectively, and calculate the envelope mean m(t);

[0165] (2) Calculate the components Determine if it is an IMF; if it is, proceed to (3); otherwise, proceed to (4). Repeat operation (1) until the IMF is obtained, denoted as I. i (t):

[0166]

[0167] (3) I at this time i (t) is the i-th IMF component, and the residual r(t) is obtained as shown in equation (11):

[0168]

[0169] (4) Stop decomposing s(t) when r(t) satisfies the condition of residual balance.

[0170] In the actual decomposition process, the mean of the envelope of the IMF is generally not 0. When equation (12) holds, it is considered to meet the condition, λ = 0.25.

[0171]

[0172] Although the EMD algorithm can remove some noise interference during the detection process, it still has the following problems:

[0173] (1) When waves or ships suddenly approach the target, spikes will appear in the detection signal.

[0174] (2) EMD exhibits endpoint effects. The decomposition process requires finding extreme points, but the boundary points of the signal are uncertain, easily leading to errors in the decomposition at these points. As the number of decompositions increases, these errors accumulate, affecting the overall signal decomposition.

[0175] (3) When the target suddenly disappears, there will be a missing time series in the original signal. It is necessary to predict the target information in time to ensure the accuracy of EMD decomposition.

[0176] To address the above problems, this invention proposes an AM-EMD algorithm by adding an ARIMA model to the EMD algorithm.

[0177] To establish an ARIMA model for the target information sequence detected by radar, the difference order d needs to be determined first. When the original sequence is differised by d orders, it will become a stationary random sequence, denoted as S(n). Assuming ε(n) is a white noise sequence, a hybrid model of p-order autoregression and q-order moving average can be established as shown in equation (13):

[0178] s(n)=α1s(n-1)+α2s(n-2)+…+α p s(np)+ε(n)-β1ε(n-1)-β2ε(n-2)-…-β q ε(nq)(13)

[0179] Wherein, coefficient α i and βj The resulting polynomials are denoted as equation (14):

[0180] α(u)=1-α1u-α2u 2 -…-α p u p β(u)=1-β1u-β2u 2 -…-β p u p (14)

[0181] At this point, the roots of α(u) and β(u) are both outside the unit circle, and s(n) can be further expressed as shown in equation (15):

[0182]

[0183] Where the coefficient ψ j Satisfying equation (16):

[0184]

[0185] It can be seen that in the process of predicting the sequence value at a certain future time, this model needs to adjust the parameter α. i and β j Estimate the influence of existing data on the time series value. A prerequisite for parameter estimation is obtaining the values ​​of the orders p and q. The least information (AIC) criterion is used to determine the model order and parameters.

[0186] The AIC criterion is defined as shown in equation (17):

[0187] AIC=-2ln(L)+2k (17)

[0188] Where L is the likelihood function of the model, and k is the number of independent parameters. The AIC criterion function consists of two terms: the first term reflects the goodness of fit of the model, decreasing as the order increases; the second term indicates the number of model parameters, increasing as the order increases. When Equation 17 has a minimum value, the corresponding orders p and q are the ideal orders. After obtaining the parameters d, p, and q, the parameter α can be... i and β j Least squares estimation is performed to fit the ARIMA model, which is then used to predict the target location in the echo image of the new frame. The steps of the AM-EMD algorithm are as follows:

[0189] (1) Determine whether there are glitches in the detected signal sequence. If there are, proceed to step (2); otherwise, proceed to step (3).

[0190] (2) Determine the interval where the spurs are located, decompose it separately using the EMD algorithm, remove the highest frequency IMF, and obtain a new sequence to replace the original sequence.

[0191] (3) Establish an ARIMA model for the signal sequence and determine the parameters for establishing the model;

[0192] (4) Predict the location of the target that disappears at a certain moment and the location of the target at the last moment to obtain a complete signal sequence;

[0193] (5) Perform EMD decomposition on the signal sequence obtained in (4) and recombine the IMF components to obtain the final signal sequence.

[0194] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0195] This invention is based on the navigation radar mounted on an unmanned surface vessel. Through quantitative analysis of experimental data, the target detection and tracking algorithm of this invention was compared with the detection and tracking algorithm based on DBSCAN clustering combined with Hungarian + Kalman filtering.

[0196] like Figure 11 and Figure 12 As shown, the experiment was conducted on an unmanned surface vessel (USV) experimental platform, which mainly consisted of two USVs, A and B, a navigation radar, and an inertial navigation system. The experimental water area was Zhanghe Reservoir in Jingmen City. USV A was the test vessel, and USV B was the target vessel. The navigation radar was a YR18M series Yuanzhou radar with a ranging range of 50m to 12nm, a horizontal beamwidth of 5.2°, a vertical beamwidth of 25°, a transmission frequency of 9.3–9.4GHz, and an installation height of 5m above the water surface. During hardware installation, it was difficult to ensure that the radar's forward direction was perfectly aligned with the bow of the moving platform, resulting in a certain azimuth error. This azimuth error caused the error between the target position detected by the navigation radar and its actual position to increase with distance, leading to excessive deviation in the target detection information and an inability to accurately perceive the water surface environment. Therefore, static calibration was needed to correct the system error before the experiment.

[0197] I. Test Experiment Scenario of this Invention

[0198] After calibration, target detection and tracking tests were conducted. The relative motion scenarios of the two boats were diverse. To demonstrate the effectiveness of the algorithm presented in this paper under various motion environments, four typical scenarios were designed to compare the algorithm with commonly used detection and tracking algorithms:

[0199] Scenario 1: Boat A remains stationary. Boat B moves in a circle with Boat A's location as the center, a radius of approximately 500 meters, and a speed of approximately 12 knots. Figure 13As shown in the image. This scenario primarily tests the performance of the algorithm presented in this paper on target detection and tracking under continuously changing angles in a stationary state.

[0200] Scenario 2: Boat A is moving at a constant speed of about 8 knots, in a low-speed state, while boat B is moving at... Figure 14 The trajectory shown depicts a uniform motion at approximately 10 knots. Initially, the two boats are 730 meters apart, and at the final moment, they are 320 meters apart. This scenario primarily tests the detection and tracking performance of the algorithm presented in this paper on targets with large-angle changes under uniform motion conditions.

[0201] Scenario 3: Boat A first moves in a curved path and then in a straight line, accelerating from 12 knots to 24 knots; Boat B remains stationary. Figure 15 As shown, the two boats were initially 420m apart and eventually 1100m apart. This scenario primarily tests the detection and tracking performance of the algorithm presented in this paper on stationary targets under variable-speed motion conditions.

[0202] Scenario 4: Boat A moves in a straight line, accelerating uniformly from 8 knots to 16 knots; Boat B moves in a straight line at a constant speed of 8 knots. The two boats move towards each other. Figure 16 As shown, the two boats were initially 1000m apart and ended 300m apart. This scenario primarily tests the detection and tracking performance of targets moving in opposite directions under varying speeds.

[0203] II. Quantitative Verification Analysis of the Invention

[0204] For ease of description, during quantitative analysis and verification, scheme A is defined as a detection and tracking algorithm based on DBSCAN clustering combined with Hungarian + Kalman filtering, and scheme B is the target detection and tracking algorithm of this invention. The speed, heading, and distance information obtained from detecting and tracking vessel B in the above four scenarios are compared. The comparison results for scenarios one, two, three, and four are as follows: Figures 17-38 As shown.

[0205] according to Figures 17-22It can be seen that during the circular motion of boat B around boat A, the actual distance between the two boats could not be guaranteed to be a fixed value. At 14:09:12, due to internal noise in the radar hardware, the echo pattern underwent significant deformation, and the distance between boat B and boat A also changed abruptly. Algorithm A produced a large distance detection error, with a maximum error of 34.45m, while the distance error of algorithm B increased slightly, with a maximum error of 12.85m. The heading detection of both algorithms A and B also changed abruptly at this time, resulting in an increase in heading error for both. The maximum error of algorithm A reached 26.30°, and the maximum error of algorithm B increased to 14.25°. Throughout the process, boat B's speed remained relatively stable, decelerating by two knots around 14:07:09. At this time, algorithm A failed to stably detect the change in speed, resulting in a maximum speed error of 2.04m / s. The overall error of algorithm B remained relatively stable, with a maximum of 0.84m / s. Overall, algorithm A's detection results for the target's distance, heading, and speed fluctuated significantly, while algorithm B's detection results were generally smoother.

[0206] according to Figures 23-28 It can be seen that when boat A is moving at a constant speed in a straight line, the detection results of algorithms A and B are similar, and their overall detection performance is not significantly different. However, in this scenario, boat B's trajectory involves several sharp turns, and under these circumstances, algorithm B's detection performance is significantly better than algorithm A's. Between 14:26:31 and 14:28:49, boat B makes a 180° turn and then begins an "S"-shaped curve motion. The frequent turns cause algorithm A to fail to detect boat B's current motion state in a timely manner. Although the distance error is relatively stable, the heading error shows multiple abrupt changes, deviating significantly from the true heading. The speed error also fluctuates greatly, exhibiting two peaks, resulting in poor overall performance. In this scenario, Algorithm A achieved a maximum distance error of 12.67m, a maximum heading error of 22.89°, and a maximum speed error of 1.38m / s. In contrast, Algorithm B was more stable overall, with a slightly reduced distance error and a maximum error of 9.42m. Algorithm B showed a clear advantage in heading and speed detection errors, with a maximum heading error of 15.67° and a maximum heading error of 0.37m / s.

[0207] according to Figures 29-32It can be seen that during the period from 14:52:28 to 14:54:25, when boat B was stationary, boat A was moving at a constant speed along a curved path. During this time, the distance detection results of both methods A and B were relatively stable, with the detection error of algorithm A being slightly larger than that of algorithm B. At 14:54:00, other targets appeared within 100m of boat B, causing algorithm A to misjudge the echo and resulting in a sudden change in the detected speed information. This led to a maximum speed error of 0.80m / s for algorithm A, while algorithm B successfully eliminated this interference signal. Between 14:54:26 and 15:00:07, vessel A underwent variable-speed linear motion. When vessel A maneuvered, the detection performance of algorithm A was significantly affected by the vessel's movement, resulting in a slight decrease in accuracy and an overall increase in the error of the measured distance information, with a maximum error of 23.71m. However, the speed error remained relatively stable. Algorithm B was less affected, and its distance error remained relatively stable, with a maximum error of 11.79m. Since algorithm B automatically classified targets with speeds less than 0.5m / s as stationary targets, its speed remained at 0.

[0208] according to Figures 33-38 It can be seen that during the process of boats A and B moving towards each other, the distance between them first shortens and then increases. Between 11:07:21 and 11:07:49, boats A and B gradually approach each other. As the distance decreases, the distance detection accuracy of both algorithms A and B shows an increasing trend. However, due to the variable speed movement of boat A, algorithm A is significantly affected by this, resulting in greater overall disturbance, especially in the course error, which is very unstable. Between 11:07:49 and 11:07:55, boats A and B meet. At this time, because the two boats are close together, the radar resolution cannot reach a sufficiently high accuracy, so the distance errors of both algorithms A and B increase, with algorithm A showing a more pronounced trend, and its speed error also slightly increasing. Algorithm B maintains both course and speed within a small and stable error range, demonstrating a clear advantage in accuracy. Between 11:07:55 and 11:08:12, vessels A and B moved in opposite directions, with the distance gradually increasing. Algorithm A's detection results still showed interference, with unstable distance error. The heading error also gradually increased because Algorithm A failed to update vessel B's motion status in time after the two vessels switched from moving towards each other to moving in opposite directions. The speed error remained relatively stable. Algorithm B maintained stable detection results for distance, heading, and speed, with a good error trend. In this scenario, Algorithm A's maximum distance error reached 23.96m, its maximum heading error reached 19.6°, and its maximum speed error reached 0.67m / s. Algorithm B's maximum distance error was 13.88m, its maximum heading error was 9.05°, and its maximum speed error was 0.46m / s.

[0209] The RMSE of algorithms A and B in the four scenarios above was calculated, and the accuracy of target detection and tracking of the two algorithms was further compared. The results are shown in Table 1. It can be seen that algorithm B outperforms algorithm A in all scenarios, with an improvement of at least 50% in distance accuracy, at least 28% in heading accuracy, and at least 35% in speed accuracy.

[0210] Table 1 Comparison of accuracy of test results in four scenarios

[0211]

[0212] The running time of the two algorithms for processing each frame of image is compared, and the results are shown in Table 2. The comparison shows that Algorithm B is slightly more efficient than Algorithm A. This is because the first frame of image involves an initialization process, which takes a relatively long time for both algorithms. From the second frame onwards, the image processing speed remains relatively stable. Algorithm B's running speed is about 20% faster than Algorithm A.

[0213] Table 2 Comparison of Algorithm Running Time

[0214]

[0215] By comparing the root mean square error of the detection results and the algorithm running speed, it can be proved that the surface target detection and tracking algorithm of the unmanned surface vessel equipped with navigation radar of the present invention has improved accuracy and effectively controlled the algorithm time compared with the traditional target detection and tracking algorithm.

[0216] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solution of the present invention and should not be used to limit the scope of protection. Any modifications made in accordance with the claims and specification of the present invention that are only partial should still fall within the protection scope of the present invention.

Claims

1. A method for detecting and tracking water surface targets based on navigation radar, characterized in that, Includes the following steps: Step 1: Receive navigation radar and inertial navigation data, and solve the navigation radar and inertial navigation data to obtain radar echo data. Then, preprocess the radar echo data. The preprocessing includes image correction, coordinate transformation and filtering. The coordinate transformation is used to map the radar echo data from the polar coordinate system to the rectangular coordinate system. Step 2: Use the Two-pass algorithm to process the preprocessed radar echo data to obtain a labeled radar echo image; Step 3: Based on the range of the navigation radar, acquire a map covering that range, then convert the geographic coordinates of the map to a rectangular coordinate system using Mercator projection, then perform a translation operation to convert it to the screen coordinate system, and finally normalize the scale of the converted map and radar echo image. Step 4: Use a distance matching algorithm based on adaptive threshold segmentation to match the normalized map and radar echo images, identify echoes belonging to land or islands, and determine the corresponding label, i.e., batch number, for echoes belonging to land or islands. Step 5: Obtain the specific location and contour information of the echo corresponding to the target batch number, use the binary KM matching algorithm to match the echoes of the same target in two consecutive frames, and assign them the same batch number; Step 6: Determine whether the newly appearing echo is a target or clutter. If it is clutter, do not assign a new batch number. If it is a target, assign a new batch number to the newly appearing target. Step 7: Detect the distance between the target and the unmanned surface vessel (USV), and calculate the target's heading and speed based on the echo information of consecutive frames. Then, determine whether there are any glitches or interferences in the distance between the target and the USV, the target's heading, and the target's speed. If there are glitches or interferences, perform EMD algorithm decomposition to obtain the IMF (Information Function Factor), and remove the highest frequency IMF from the IMF. Then, reconstruct the signal from the remaining IMF to obtain the distance between the target and the USV, the target's heading, and the target's speed after removing glitches or interferences. Otherwise, determine whether the echo corresponding to a target that has been assigned a batch number and has appeared for more than 5 consecutive frames has disappeared. When the target disappears, use the ARIMA model to predict the position of the disappeared target. When the echo corresponding to the disappeared target appears in the next 3 frames, match the predicted position of the target with the position where the target appeared. If the matching result is consistent, assign the same batch number; otherwise, remove the disappeared target, thus completing the target detection and tracking. Step 8: After removing the missing target, predict the position of all echoes in the next 3 consecutive frames, and then use the EMD algorithm to optimize the current distance, heading and speed of all targets from the unmanned surface vessel.

2. The method for detecting and tracking water surface targets based on navigation radar according to claim 1, characterized in that... The conversion of geographic coordinates to a rectangular coordinate system is expressed as follows: Where (X, Y) are the latitude and longitude coordinates of a point P on the Earth's surface, (x, y) are the rectangular coordinates corresponding to (X, Y), K is the radius of the reference latitude circle, L0 is the longitude of the origin of the Earth's surface, and the latitude of the origin is 0.

3. The method for detecting and tracking water surface targets based on navigation radar according to claim 2, characterized in that... The dimensions of the converted map are expressed as follows: Among them, P c P and R represent the dimensions of the radar image and map in the screen coordinate system, respectively. c R represents the actual distance in the radar image, and R represents the actual distance in the map.

4. The method for detecting and tracking water surface targets based on navigation radar according to claim 3, characterized in that... The specific steps for matching the normalized map and radar echo image using the distance matching algorithm based on adaptive threshold segmentation in step 4 are as follows: Step 4.1: Use the pixels of the unmanned surface vessel's location in the radar echo image and the normalized map as matching points to achieve preliminary matching between the radar echo image and the normalized map, and obtain a matching image. Step 4.2: Binarize the radar echo image. The normalized map is then binarized using an adaptive threshold segmentation algorithm, which specifically includes: Step 4.2.1: Find the extreme gray values ​​of the normalized map, i.e., the maximum gray value l. max (x, y) and minimum gray level l min (x, y), calculate the maximum gray level l of the normalized map. max (x, y) and minimum gray level l min The mean G1 of (x, y); Step 4.2.2: Using G1 as the initial threshold for the iterative method, and using the initial threshold as the boundary threshold, find the extreme gray values ​​of all pixels in the normalized map that are greater than G1 and the extreme gray values ​​of all pixels that are less than G1, and obtain the average value G of all extreme gray values ​​of pixels greater than G1. 11 And the average of the grayscale extreme values ​​of all pixels less than G1, G 12 Thus, the second threshold G2 is obtained; Step 4.2.3: Update the threshold to G2, and repeat step 4.2.2 until |G i -G i-1 When |<0.1, stop and select G. i Binarize the normalized map; Step 4.3: Extract the contours of the radar echo image and map obtained from Step 4.2 using the Canny operator; Step 4.4: Based on the image matching in Step 4.1, and combined with the contours of the radar echo image and the map, select the minimum bounding rectangles of the radar echo image and the map respectively, and use the larger of the two minimum bounding rectangles as the matching area of ​​the radar echo image and the matching area of ​​the map respectively. Step 4.5: Calculate the Hausdorff distance between the matching region in the radar echo image and the matching region in the map, and select the matching point with the highest similarity for matching.

5. The method for detecting and tracking water surface targets based on navigation radar according to claim 4, characterized in that... The specific steps of step 4.5 are as follows: Assume that the pixels in the matching region of the radar echo image and the matching region of the map constitute sets A and B, respectively: in, This represents the p-th pixel in set A. This represents the q-th pixel in set B; The Hausdorff distance between A and B is defined as: Where i = 1, 2...p, j = 1, 2...q, Hausdorff represents the route from B to A. Represents the Hausdorff from A to B.

6. The method for detecting and tracking water surface targets based on navigation radar according to claim 5, characterized in that: In step 5, the specific steps for matching echoes belonging to the same target in two consecutive frames using the binary KM matching algorithm are as follows: Step 5.1: Assume the specific location coordinates of the echo from the previous frame are... The corresponding echo area is The specific location coordinates of the echo in the next frame are: The corresponding echo area is Find the minimum sum of the Euclidean distances of the specific locations of the echoes in the two frames, sum1, and the minimum sum of the Euclidean distances of the echo areas, sum2, respectively. Step 5.2: Based on sum1 and sum2, and using the KM algorithm for matching, obtain the matching result.

7. The method for detecting and tracking water surface targets based on navigation radar according to claim 6, characterized in that: The specific steps for performing EMD algorithm decomposition to obtain IMF are as follows: Step 7.2.1: First, acquire the signal sequence of information regarding the target's distance, heading, or speed from the unmanned surface vessel, represented as: Where r is the signal margin. Let K be the k-th intrinsic mode component, and K be the total number of intrinsic mode components. Then, all the maxima and minima of s(t) are obtained. The maxima and minima are fitted with cubic spline fitting functions to obtain the upper and lower envelope curves of s(t), and the mean m(t) of the upper and lower envelope curves is obtained. Step 7.2.2: Calculate the components ,judge Is it an IMF? If it is an IMF, then record it as IMF. And proceed to step 7.2.3, if it is not the IMF, then let =s(t), and execute step 7.2.1 until the IMF is obtained: judge To be considered an IMF member, two conditions must be met: (1) The difference between the number of zeros and the number of extreme points in the signal is no greater than 1; (2) The mean of the upper and lower envelopes of the signal at any point is 0; Step 7.2.3: For the i-th IMF component, calculate the residual r(t): ； Step 7.2.4: Stop the decomposition when the residual r(t) is a constant or a monotonic function.

8. A method for detecting and tracking water surface targets based on navigation radar according to claim 7, characterized in that... The specific steps for extracting the radar echo image and map contour using the Canny operator in step 4.3 are as follows: The Canny operator is used to extract the contours of radar echo images and maps respectively. The extracted contours are then filtered based on the mean area of ​​the contours, with contours greater than the mean area being retained and those smaller than the mean area being discarded.

9. A method for detecting and tracking water surface targets based on navigation radar according to claim 1, characterized in that... The filtering process in step 1 includes isolated noise point filtering and echo threshold filtering.

10. A method for detecting and tracking water surface targets based on navigation radar according to claim 1, characterized in that... The image correction method uses an image sliding window correction technique.

Images