Helicopter collision avoidance radar one-dimensional range image target feature extraction and identification method

Abstract

The application discloses a kind of helicopter anti-collision radar one-dimensional range image target feature extraction and identification method, comprising the following steps: S1, the echo signal received by radar is carried out FFT operation to obtain one-dimensional range image;S2, one-dimensional range image data is carried out two-stage threshold detection;S3, with the center of the target scattering point set track of being determined, target clustering is carried out;S4, to the target scattering point set after clustering, unknown target feature vector is formed by extracting feature parameter feature;S5, by training algorithm, establish expert knowledge base, according to expert knowledge base, unknown target feature vector is carried out decision tree identification decision, and target attribute recognition result is obtained;S6, power line confirmation is carried out using power line bragg scattering effect and tower line group characteristics, and target recognition result is output.The feature extracted by the application has good robustness, high correct identification rate, can effectively distinguish power line, tower, chimney, wind tower and other typical obstacles of helicopter, and has good clutter suppression performance.

Description

Technical Field

[0001] This invention relates to the field of helicopter collision avoidance technology, specifically to a method for extracting and identifying target features in a one-dimensional range image of a helicopter collision avoidance radar. Background Technology

[0002] Helicopters are highly maneuverable aerial platforms, widely used in various fields due to their advantages such as low altitude, low speed, and vertical takeoff and landing capabilities. When performing missions, helicopters often fly at low or very low altitudes, facing extremely complex ground environments. These environments include dense shrubs, forests, hills, high-voltage power grids, and man-made structures such as chimneys and wind towers, posing a serious threat to the safety of helicopters flying at low altitudes. One effective means of obstacle avoidance for helicopters is to use collision avoidance radar for early warning and avoidance. Millimeter-wave collision avoidance radar, due to its small size, light weight, narrow beam, low power consumption, and all-weather, all-time operation advantages, occupies an important position in helicopter obstacle detection equipment. Early collision avoidance radars utilized the Bragg scattering characteristics caused by the periodic structure of power lines for target detection. However, because power lines have a small RCS and weak echo signals, detection solely through the Bragg effect has a low probability of success and is greatly affected by ground features. Methods based on tower-line group characteristics and Hough transform line detection, building upon Bragg scattering effect detection, utilize the topology of inter-tower connections or Hough transform line detection to identify power lines. However, in practical applications, it has been found that due to the wide range of applications of helicopters, there are various obstacles in the area in front of them. In addition to power lines and power towers, man-made structures such as chimneys and wind towers not only pose obstacles and threats to the safety of helicopters, but conventional tower line groups and Hough transform line detection and recognition cannot adapt to these complex target environments.

[0003] In recent years, radar target recognition technology has been increasingly researched and applied in radar engineering as an emerging field. The most crucial aspect of radar target recognition is target feature selection and extraction; the ability of the extracted features to effectively distinguish various targets directly impacts the recognition outcome. Feature extraction and recognition using one-dimensional high-resolution range images is an important branch of radar target recognition. Currently, research on feature recognition based on high-resolution one-dimensional range images of radar mostly focuses on large target recognition applications such as ground vehicles and large ships at sea, with very little research on its application in helicopter-borne collision avoidance radar. This is mainly because typical targets of helicopter-borne collision avoidance radar, such as power towers, power lines, chimneys, and wind towers, are relatively small in size, and their radial one-dimensional range images are not large. After typical CFAR detection, the number of strong scattering points in the radial range direction is very limited, making it difficult to present a true one-dimensional range image of the target, thus making feature extraction extremely challenging.

[0004] However, analysis of a large amount of echo data from millimeter-wave radar illumination reveals that, due to the short wavelength, narrow beam, wide bandwidth, and high resolution of millimeter-wave radar, typical obstacle targets such as power towers, chimneys, wind towers, and power lines, under millimeter-wave illumination, contain multiple scattering points in the echo, forming a one-dimensional range profile with varying heights. Further analysis shows that, since radar echoes are the superposition of radar waves reflected from scattering points at different spatial locations of the target, differences in the target's spatial structure and materials determine its different scattering coefficients. The scattering points of several typical obstacle targets exhibit certain differences in both the energy domain and the structural spatial domain. Feature extraction and comparison of the one-dimensional range profiles of these typical targets can achieve the differentiation of different targets. However, as mentioned earlier, to avoid an increase in the false alarm rate of radar detection due to noise and ground clutter, the target detection threshold of collision avoidance radar is usually tens of decibels higher than the background noise and clutter. As the radar antenna scans, some weak scattering points in the target echo may be missed, leading to the loss of some one-dimensional range information of the target, which is detrimental to feature recognition. Analysis of typical target echo data at different radar distances and illumination angles reveals that while the strong scattering points of the target can be detected by the typical CFAR threshold, most of the weak scattering points, although below the detection threshold, are stronger than background clutter. If the information of the target's weak scattering points can be preserved while suppressing clutter, then collision avoidance radar can achieve target feature extraction and identification based on a one-dimensional range profile.

[0005] The target feature extraction and recognition method for one-dimensional range images of millimeter-wave collision avoidance radar employs a two-stage threshold detection approach. The first-stage threshold maximizes the suppression of ground clutter interference, ensuring that the false alarm probability is kept within a controllable range. The second-stage threshold detection obtains the weak scattering points of targets detected by the first-stage threshold. During target clustering, targets are associated using the scattering points obtained from the first-stage threshold as centers. Points detected by the second-stage threshold are either classified as weak scattering points or discarded based on their association with points detected by the first-stage threshold. This approach ensures both clutter suppression performance and preserves relatively rich one-dimensional range image information, laying the foundation for feature extraction. Using the target's one-dimensional range image, multi-dimensional feature parameters such as relative RCS, radial length, number of peak points, and radial amplitude distribution entropy are extracted for each target, forming an unknown feature vector. Then, a decision tree is used to identify and judge the unknown feature vector using an expert knowledge base trained with a large amount of measured data, yielding the target recognition result. Because the power line echo is very weak, in order to further suppress the interference of clutter on the performance of the radar system, after completing the classification and identification of four typical targets, the Bragg scattering characteristics and tower group characteristics are used again to further confirm the power line targets, thereby further improving the reliability of the system. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a method for extracting and identifying target features from a one-dimensional range image of a helicopter collision avoidance radar, thus solving the problem of helicopter collision avoidance.

[0007] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for extracting and identifying target features in a one-dimensional range image of a helicopter collision avoidance radar, comprising the following steps:

[0008] S1. Perform FFT calculation on the echo signal received by the radar to obtain a one-dimensional range profile;

[0009] S2. Perform two-level threshold detection on the one-dimensional range image data to obtain the set of definite target scattering points and candidate target scattering points;

[0010] S3. Cluster the target based on the trace of the target scattering point set. If the scattering point in the candidate target scattering point set can be successfully associated with the trace of the target scattering point set, it is judged as a weak scattering point of the target and added to the target scattering point set; otherwise, it is discarded as clutter.

[0011] S4. Extract the target relative RCS, target radial length, number of peak points, amplitude entropy, and amplitude fluctuation feature parameters from the clustered target scattering point set to form an unknown target feature vector;

[0012] S5. Establish an expert knowledge base through training algorithms, and make a decision tree recognition judgment on the feature vector of the unknown target based on the expert knowledge base to obtain the target attribute recognition result.

[0013] S6. Power line identification is performed using the Bragg scattering effect of power lines and the characteristics of tower line groups, and the target identification results are output.

[0014] Further: Step S1 specifically involves: In the frequency-modulated continuous wave radar, after the radar received signal and the local oscillator signal are mixed, the beat signal between the two is obtained. The digital signal after amplification and AD conversion of the radar received signal is subjected to FFT operation to obtain the corresponding signal power spectrum. Since the beat signal frequency of targets at different distances is different, the signal power spectrum obtained after FFT operation forms a one-dimensional range image {x(n), n=1、2….N} of the entire range coverage area, where n is the scattering point range cell number and N is the number of scattering point range cells.

[0015] Further: Step S2 specifically includes:

[0016] S21. The one-dimensional distance image data is sequentially fed into the sliding window detector. The center of the sliding window detector is the detection unit, the two sides of the detection unit are the protection units, and the outer side of the protection unit is the reference unit.

[0017] S22. When each detection unit is detected, the reference units on both sides are sorted by amplitude. After sorting, strong points are removed. The three largest values ​​and the points marked as targets in their respective regions are removed to eliminate the influence of strong targets or strong interference in the reference units on the detection threshold.

[0018] S23. For the reference units on both sides where strong points have been removed, calculate the average amplitude values ​​Y1 and Y2 on both sides respectively.

[0019] S24. Compare the average values ​​of the two sides and select the smaller value as the reference value for calculating the reference threshold;

[0020] S25. Calculate the detection threshold, including the primary detection threshold and the secondary detection threshold;

[0021] The formula for calculating the first-level detection threshold T1 is:

[0022] T1 = α·Min(Y1,Y2)

[0023] In the above formula, α is the first-level detection threshold coefficient;

[0024] The formula for calculating the secondary detection threshold T2 is:

[0025] T2 = 0.6·T1

[0026] S26. The amplitudes of the unit to be detected and the two-level detection thresholds are compared respectively, and the results are as follows:

[0027] {n∈A|x(n)>T1}

[0028] {n∈B|x(n)>T2}

[0029] In the above formula, n is the distance unit number of the scattering point, A is defined as the set of scattering points of the determined target, and B is the set of scattering points of the candidate target.

[0030] Further: Step S3 specifically includes:

[0031] S31. Taking any scattering point in set A as the center, traverse all other unassociated scattering points in set A. When the distance between two scattering points is less than 5 gates, they are clustered into one target. Points that are successfully associated are marked as associated.

[0032] S32. Using the newly associated scattering point as the center, traverse all other unassociated scattering points in set A and perform clustering and labeling until all scattering points belonging to this target in set A have been clustered. Then, construct a target set S from all the scattering points belonging to this target. k Define S k , k = 1, 2, ..., is the set of distance cell indices for the scattering point of the k-th target;

[0033] S33, with S k Using each scattering point as the center, traverse all unassociated scattering points in set B. When two scattering points are within 3 range gates, they are determined to be scattering points belonging to the k-th target, and the scattering point is marked as associated.

[0034] S34. Add the successfully associated scattering points from set B to S. k In this process, a set of scattering points from k targets is formed;

[0035] S35. Cluster the remaining unassociated scattering points in set A according to steps S31 to S34 to form a set of scattering points of other targets until all scattering points in A are associated, and obtain the clustered target scattering point set.

[0036] Further: Step S4 specifically includes:

[0037] S41. Assume that the clustered target scattering point set S contains p scattering points for the k-th target. The p scattering points of the k-th target are sorted according to their distance cell indices to obtain the sorted distance cell indices set {R_num}. k,i For each set i = 1, 2, ..., p, extract the amplitude of the scattering points corresponding to the set to obtain the amplitude set {amp}. k,i , i = 1, 2, ..., p}, where i is the scattering point index;

[0038] S42. Calculate the target cohesion distance R k The calculation formula is:

[0039]

[0040] S43. Extract the target's relative RCS, the calculation formula is as follows:

[0041]

[0042] In the above formula, RCS k For the target relative RCS;

[0043] S44, Extract radial length L k The calculation formula is:

[0044] L k =[max(R_num k,i )-min(R_num k,i )+1]·dR

[0045] In the above formula, dR is the distance cell length;

[0046] S45, Extracting the number of radial peaks PEAK k The calculation formula is:

[0047] PEAK k =length{i|amp k,i ≥amp k,i-j ,amp k,i ≥amp k,i+j ,i=1,2..p,j=1,2}

[0048] In the above formula, length{i|} ​​represents the number of i that satisfy the condition in parentheses, and j is the index;

[0049] S46. Extract the amplitude distribution entropy. entropy The calculation formula is:

[0050]

[0051]

[0052] In the above formula, amp′ k,i It is the ratio of the amplitude at the scattering center to the sum of the amplitudes;

[0053] S47. Extracting the normalized variance of amplitude distribution The calculation formula is:

[0054]

[0055] In the above formula, m x This represents the average amplitude of the target scattering point;

[0056] S48. Extract the main peak energy ratio C, the calculation formula is:

[0057]

[0058] In the above formula, amp k,j This represents the peak amplitude.

[0059] Furthermore, the expert knowledge base is established offline by using measured echo data of known typical obstacle targets obtained from various illumination angles and distances, and by using feature vectors obtained through detection, clustering, and feature extraction, and then training algorithms.

[0060] Further, in step S5, attribute recognition specifically involves: using a decision tree for classification. The decision tree includes a root node, a set of intermediate nodes, and some terminal nodes. The decision tree recognition process starts from the root node. Based on the corresponding feature attributes in the items to be classified obtained from feature extraction, a splitting predicate is marked according to the magnitude of the feature attribute values ​​and the decision threshold of the branches, and an output branch is selected. Different feature attributes are used sequentially, and the selection of output branches is performed from top to bottom until a leaf node is reached. The category stored in the leaf node is taken as the result of the decision tree classification.

[0061] Furthermore, the order of use of the characteristic attributes is as follows: target radial length, relative RCS, number of peak points, main peak energy ratio, dispersion coefficient, amplitude normalized variance, and amplitude distribution entropy.

[0062] Further: Step S6 specifically includes:

[0063] S61. Process and confirm the target data after attribute identification in multiple repeating cycles of radar scanning one week from left to right or from right to left, and determine whether there are power line targets among the targets detected in one week of scanning. If so, proceed to step S62; otherwise, proceed to step S69.

[0064] S62. Determine if the target number of power lines is greater than 2. If yes, proceed to step S63; otherwise, proceed to step S64.

[0065] S63. Determine whether the target azimuth of the electric power line satisfies the Bragg scattering angle relationship. If yes, proceed to step S69; otherwise, proceed to step S64.

[0066] S64. Determine whether there is a power tower among the targets detected by the weekly scan. If so, proceed to step S65; otherwise, proceed to step S68.

[0067] S65. Project the power lines and power towers onto the aircraft's geographic coordinate system, and then proceed to step S66.

[0068] S66. For the power towers projected onto the aircraft's geographic coordinate system, make inter-tower connections in the east-north plane, and then proceed to step S67.

[0069] S67. Determine whether the power line target is within the area of ​​the inter-tower connection. If so, the power line target is a confirmed target and proceed to step S69; otherwise, proceed to step S68.

[0070] S68. For targets that cannot be identified as power lines, mark them as suspected power line targets and proceed to step S69.

[0071] S69. Output the target recognition result.

[0072] Furthermore: Targets marked as suspected power lines will be confirmed by the radar display control terminal during weekly data checks. If the power line target fails to be identified as a confirmed power line target in three consecutive weeks of data in the same scanned airspace, it will be discarded as interference.

[0073] The beneficial effects of this invention are as follows:

[0074] 1. High probability of correct identification. Current methods for identifying power lines based solely on the Bragg scattering effect, tower-line group feature identification, or Hough transform line detection all employ only a single feature and focus solely on tower lines, resulting in a low probability of correct identification in complex environments. The millimeter-wave collision avoidance radar one-dimensional range image feature extraction and target identification method leverages the small wavelength, narrow beam, and high resolution of millimeter-wave radar. It obtains a set of one-dimensional range image scattering points for power lines, power towers, chimneys, and wind measurement towers through two-stage threshold detection. It extracts multi-dimensional feature parameters, including energy characteristic parameters (relative RCS), structural size characteristic parameters (radial length), and backscattering characteristics (number of peak points, amplitude entropy, etc.), forming a feature vector. The real-time feature vector is then used to identify target attributes through an expert knowledge base trained on a large amount of measured data. This method allows for the classification and identification of multiple target types, extracts rich and robust features, and achieves a high probability of correct identification even in complex environments.

[0075] 2. Strong clutter suppression capability. The target feature extraction and identification method for millimeter-wave collision avoidance radar in this invention employs a two-stage threshold detection process on the received echo range image. The first-stage threshold uses OS-CFAR threshold detection to suppress clutter and obtain definite target scattering points, ensuring that the false alarm probability during detection is controlled within the system's allowable range. The second-stage threshold uses 0.6 times the first-stage threshold to obtain weak scattering points of the definite target in the echo. During target clustering, the definite scattering points detected by the first-stage threshold are used as the center. If weak scattering points generated by the second-stage threshold can be successfully associated with the definite scattering points, they become one of the target scattering points; otherwise, they are discarded as clutter, effectively avoiding clutter interference. When using target identification feature values, large-area ground targets such as houses, continuous forests, and typical obstacles for helicopter collision avoidance radar are first distinguished by the target radial length feature, avoiding interference from large-area strong ground targets on target feature identification. Considering the small size of power lines and the potential for interference to affect their echo characteristics, after identifying power lines, towers, chimneys, and wind measurement towers, the Bragg scattering characteristics of power lines and the characteristics of tower-line groups are used again to confirm the power lines, resulting in a system with strong anti-clutter capability.

[0076] 3. The algorithm is simple, effective, and easy to implement in engineering. The current invention uses algorithms based on tower line group feature recognition or Hough transform line segment recognition to process and identify all detected data. This process is susceptible to clutter interference and involves a large computational load, making it difficult for miniaturized collision avoidance radars to meet the resource requirements of the algorithm. The millimeter-wave collision avoidance radar one-dimensional range image target feature extraction and recognition method identifies attributes through target detection and clustering, feature extraction, and classification. During detection and clustering, a large amount of clutter interference is suppressed. After extracting the one-dimensional range image features, continuous large targets on the ground (such as houses and forests) are first screened by target radial length, followed by typical obstacle recognition. This avoids interference from large-area targets on the recognition algorithm and reduces the computational load. After several types of typical targets are identified, Bragg scattering and group characteristics are used for power line confirmation. At this point, the data volume is small, the algorithm is simple and effective, the computational load is low, and it has strong adaptability for engineering applications. Attached Figure Description

[0077] Figure 1 This is a flowchart of the one-dimensional range image target feature extraction and recognition process of millimeter-wave collision avoidance radar.

[0078] Figure 2 yes Figure 1 Block diagram of the two-level threshold detection principle.

[0079] Figure 3 yes Figure 1 Flowchart of target feature extraction process.

[0080] Figure 4 yes Figure 1 Flowchart of the power line target confirmation process. Detailed Implementation

[0081] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0082] like Figure 1 As shown, a method for extracting and identifying target features in a one-dimensional range image of a helicopter collision avoidance radar includes the following steps:

[0083] S1. Perform FFT calculation on the echo signal received by the radar to obtain a one-dimensional range profile;

[0084] In frequency-modulated continuous wave radar, the radar received signal and the local oscillator signal are mixed to obtain their beat signal. The digital signal obtained after amplification and A / D conversion of the received signal is subjected to FFT operation to obtain the corresponding signal power spectrum. Since the beat signal frequency is different for targets at different distances, the signal power spectrum obtained after FFT operation forms a one-dimensional range profile {x(n), n=1、2….N} for the entire range coverage area.

[0085] S2. Perform two-level threshold detection on the one-dimensional range image data to obtain the set of definite target scattering points and candidate target scattering points;

[0086] The target detection employs a sliding window approach to perform two-level threshold detection on one-dimensional distance image data. An automatic filtering method is introduced during the detection process: each point identified as a target is marked. When forming the threshold, points already identified as targets are discarded as strong points and no longer participate in the threshold calculation. (See also...) Figure 2 The target detection steps are as follows:

[0087] Step 21: The one-dimensional range image data obtained after FFT calculation is sequentially fed into the sliding window detector. The center of the sliding window detector is the detection unit, and the two sides of the detection unit are the protection units. The product of the number of protection units M and the range resolution length is set to be about twice the length of the tower. The protection units do not participate in the threshold formation. The outer side of the protection units is the reference unit. Considering the influence of strong targets or strong interference to be eliminated during detection, the number of reference units N is selected to be 16 to 32.

[0088] Step 22: When detecting each detection unit, first sort the reference units on both sides according to their amplitude. After sorting, strong points are removed. Remove the three largest values ​​and the points marked as targets in their respective regions to eliminate the influence of strong targets or strong interference in the reference units on the detection threshold.

[0089] Step 23: For the reference cells on both sides where strong points have been removed, calculate the average amplitude values ​​Y1 and Y2 on both sides respectively.

[0090] Step 24: Compare the average values ​​of the two sides and select the smaller one, Z, as the reference value for calculating the reference threshold.

[0091] Step 25: Calculate the detection threshold: The formula for calculating the first-level detection threshold is as follows:

[0092] T1 = α·Min(Y1,Y2)

[0093] The formula for calculating the level 2 detection threshold is as follows:

[0094] T2 = 0.6·T1

[0095] Step 26: Compare the amplitudes of the unit to be detected and the two-level detection thresholds respectively, and output the results as follows:

[0096] {n∈A|x(n)>T1}

[0097] {n∈B|x(n)>T2}

[0098] Where n is the scattering point distance cell number, and A is defined as the set of definite target scattering points, and B is the set of candidate target scattering points.

[0099] S3. Cluster the target based on the trace of the target scattering point set. If the scattering point in the candidate target scattering point set can be successfully associated with the trace of the target scattering point set, it is judged as a weak scattering point of the target and added to the target scattering point set; otherwise, it is discarded as clutter.

[0100] Target clustering involves aggregating scattering points from sets A and B, grouping scattering points from the same target together and separating scattering points from different targets. To ensure clutter suppression performance during processing, the target clustering process uses a point in set A as the center. If a scattering point in set Y can be successfully associated with a scattering point in set X, it becomes one of the target's scattering points; otherwise, it is discarded as clutter. The specific target clustering process is as follows:

[0101] Step 31: First, take any scattering point in set A as the center, traverse all other unassociated scattering points in set A, and when the distance between two scattering points is less than 5 distance gates, they are clustered into one target. Points that are successfully associated are marked as associated.

[0102] Step 32: Using the newly associated scattering point as the center, traverse all other unassociated scattering points in set A and perform clustering and labeling until all scattering points belonging to this target in set A have been clustered. Then, construct a target set S from all scattering points belonging to this target. k Define S k (k=1,2…) is the set of distance cell indices for the scattering point of the k-th target.

[0103] Step 33: With S k Using each scattering point as the center, traverse all unassociated scattering points in set B. If two scattering points are within 3 range gates, they are determined to be scattering points belonging to the k-th target, and the scattering point is marked as associated.

[0104] Step 34: Add the successfully associated scattering points from B to S. k In this process, the set of scattering points of the k-th target is formed.

[0105] The remaining unassociated scattering points in set A are clustered in the same way as described above, forming scattering point sets for other targets, until all scattering points in A are associated.

[0106] S4. Extract the target relative RCS, target radial length, number of peak points, amplitude entropy, and amplitude fluctuation feature parameters from the clustered target scattering point set to form an unknown target feature vector;

[0107] Target feature extraction utilizes one-dimensional range profile data (including scattering point number and scattering point amplitude information) formed by scattering points of the same target in set S to extract multi-dimensional features of the target, including relative RCS, radial length, number of radial peak points, amplitude distribution entropy, amplitude distribution normalized variance, dispersion coefficient, and main peak energy ratio.

[0108] Assuming the k-th target contains p scattering points, see [reference] Figure 3 The steps for target feature extraction are as follows:

[0109] Step 41: Sort the p scattering points of the k-th target according to the distance cell index, and obtain the sorted set of distance cell indices {R_num}. k,i For each set i = 1, 2, ..., p, extract the amplitude of the scattering points corresponding to the set to obtain the amplitude set {amp}. k,i , i = 1, 2, ..., p}.

[0110] Step 42: Calculate the target cohesion distance, R. k The calculation formula is as follows:

[0111]

[0112] Step 43: Target relative RCS extraction. The formula for calculating the target relative RCS is as follows:

[0113]

[0114] Step 44: Radial length extraction. The formula for calculating the radial length is as follows:

[0115] L k =[max(R_num k,i )-min(R_num k,i )+1]·dR

[0116] Where dR is the distance length of the distance unit.

[0117] Step 45: Radial peak count extraction. The radial peak count is extracted using the five-neighborhood peak method, and the calculation formula is as follows:

[0118] PEAK k =length{i|amp k,i ≥amp k,i-j ,amp k,i ≥ampk,i+j ,i=1,2..p,j=1,2}

[0119] Where length{i|} ​​represents the number of i that satisfy the condition within the parentheses.

[0120] Step 46: Extracting amplitude distribution entropy. The calculation process for amplitude distribution entropy is as follows:

[0121] Step 1: Divide the amplitude of the scattering centers by the sum of their amplitudes to obtain:

[0122]

[0123] Step 2: Calculate the entropy of the amplitude distribution at the scattering center.

[0124]

[0125] Step 47: Extract the normalized variance of the amplitude distribution. The formula for calculating the normalized variance of the amplitude distribution is as follows:

[0126]

[0127] Where, m x This is the average amplitude of the target scattering point.

[0128] Step 48: Extraction of the dispersion coefficient. The formula for calculating the dispersion coefficient is as follows:

[0129]

[0130] Where amp_m represents the peak amplitude of the target scattering point, and R_m represents the distance cell number of the peak amplitude point.

[0131] Step 49: Extraction of main peak energy ratio, assuming amp k,j The formula for calculating the main peak energy ratio, representing the peak amplitude, is as follows:

[0132]

[0133] S5. Establish an expert knowledge base through training algorithms, and make a decision tree recognition judgment on the feature vector of the unknown target based on the expert knowledge base to obtain the target attribute recognition result.

[0134] Offline, the measured echo data of known typical obstacle targets (power lines, power towers, chimneys, and wind measurement towers) obtained from various illumination angles and distances are used to establish an expert knowledge base by using the feature vectors obtained through the same detection, clustering, and feature extraction processes described above, and then training algorithms.

[0135] Attribute recognition uses decision trees for classification. A decision tree is a tree structure consisting of a root node, a set of intermediate nodes, and several terminal nodes. It also includes branches, splitting attributes, and categories. Each terminal node represents one of four categories: chimney, iron tower, power line, and wind measurement tower. The decision tree recognition process starts from the root node. Based on the corresponding feature attributes of the item to be classified obtained from feature extraction, a splitting predicate is marked according to the magnitude of the feature attribute value and the decision threshold of the branch, and an output branch is selected. Different feature attributes are used sequentially, and the selection of output branches is performed from top to bottom until a leaf node is reached. The category stored in the leaf node is taken as the result of the decision tree classification. In the decision tree classification process, the feature usage order is as follows: target radial length, relative RCS, number of peak points, main peak energy ratio, dispersion coefficient, amplitude normalized variance, and amplitude distribution entropy.

[0136] S6. Power line identification is performed using the Bragg scattering effect of power lines and the characteristics of tower line groups, and the target identification results are output.

[0137] Power line confirmation is achieved by processing target data identified through attribute recognition within multiple repetitive cycles of a radar scan (either from left to right or right to left). (See [reference needed]). Figure 4 The power line verification process is as follows:

[0138] Step S61: Determine whether there are power line targets among the targets detected by the weekly scan. If so, proceed to step S62; otherwise, proceed to step S69.

[0139] Step S62: Determine if the target number of power lines is greater than 2. If yes, proceed to step S63; otherwise, proceed to step S64.

[0140] Step S63: Determine whether the target azimuth of the electric field line satisfies the Bragg scattering angle relationship. If yes, proceed to step S69; otherwise, proceed to step S64.

[0141] Step S64: Determine whether there is a power tower among the targets detected by the weekly scan. If yes, proceed to step S65; otherwise, proceed to step S68.

[0142] Step S65: Project the line and tower targets onto the aircraft's geographic coordinate system, and then proceed to step S66;

[0143] Step S66: For the power towers projected onto the aircraft's geographic coordinate system, make inter-tower connections in the east-north plane, and then proceed to step S67.

[0144] Step S67: Determine whether the power line target is within the neighborhood of the inter-tower connection. If so, the power line target is a confirmed target, and proceed to step S69; otherwise, proceed to step S68.

[0145] Step S68: Mark the target that could not be identified as a power line as a suspected power line target and proceed to step S69;

[0146] Step S69: Output the target recognition result.

[0147] Specifically, targets marked as suspected power lines will be confirmed by weekly data at the radar display control terminal. If the power line target fails to be identified as a confirmed power line target in three consecutive weeks of data in the same scanned airspace, it will be discarded as interference.

Claims

1. A method for feature extraction and recognition of a one-dimensional range profile target of a helicopter collision avoidance radar, characterized in that, The method comprises the following steps: S1, performing FFT operation on the echo signal received by the radar to obtain a one-dimensional range image; S2, performing two-stage threshold detection on the one-dimensional range image data to obtain a set of determined target scattering points and alternative target scattering points; S3, performing target clustering with the set of determined target scattering point tracks as the center, and if a scattering point in the set of alternative target scattering points can be successfully associated with a track in the set of determined target scattering points, the scattering point is determined to be a weak scattering point of the target, and is added to the set of target scattering points, otherwise, the scattering point is discarded as clutter; S4, extracting target relative RCS, target radial length, peak point number, amplitude entropy, and amplitude fluctuation characteristic parameter characteristics from the clustered target scattering point set to form an unknown target feature vector; S5, establishing an expert knowledge base through a training algorithm, and performing decision tree recognition and judgment on the unknown target feature vector according to the expert knowledge base to obtain a target attribute recognition result; S6, confirming the power line by using the power line Bragg scattering effect and the tower line group characteristics, and outputting the target recognition result; The step S3 specifically comprises: S31, taking any scattering point in the set A as the center, traversing all other unassociated scattering points in the set A, and clustering the two scattering points into a target when the distance between the two scattering points is less than 5 gates, and marking the associated points as associated; S32, centering on the newly associated scattering point, traversing all other unassociated scattering points in set A and clustering and marking, until all scattering points belonging to the target in set A are clustered, and all scattering points belonging to the target form a target set S k , define S k , k = 1, 2…, is the scattering point distance unit serial number set of the kth target S33, to S k Each scattering point in the center of the set B, all the unassociated scattering points, when two scattering points are less than 3 distance gate, judging as belonging to the kth target scattering point, and mark the scattering point has been associated; S34, add the successfully associated scattering points in set B to S k In which the scattering point set of k targets is formed; S35, clustering the remaining unassociated scattering points in the set A according to steps S31-S34 to form a scattering point set of other targets, until all scattering points in the set A are associated, and obtaining a clustered target scattering point set.

2. The helicopter collision avoidance radar one-dimensional range profile target feature extraction and identification method according to claim 1, characterized in that, The step S1 is specifically: in the frequency modulation continuous wave radar, after mixing the radar receiving signal and the local oscillator signal, the beat signal of the two is obtained, the digital signal of the radar receiving signal after amplification and AD conversion is subjected to FFT operation, the corresponding signal power spectrum is obtained, since the beat signal frequencies of the targets located at different distances are different, the signal power spectrum obtained after the FFT operation forms a one-dimensional range image {x(n), n=1, 2…N} of the entire range coverage area, n is the distance unit serial number of the scattering point, and N is the number of distance unit of the scattering point.

3. The method according to claim 1, characterized in that, The step S2 specifically comprises: S21, sequentially inputting the one-dimensional range image data into a sliding window detector, the center position of the sliding window detector being a detection unit, the two sides of the detection unit being protection units, and the outside of the protection units being reference units; S22, when each detection unit is detected, sorting the two sides of the reference units according to the amplitudes, performing strong point elimination after the sorting, and eliminating three large values in the respective regions and points that have been marked as targets to eliminate the influence of strong targets or strong interference in the reference units on the detection threshold; S23, calculating the amplitude average values Y1 and Y2 of the two sides of the reference units after the strong point elimination; S24, comparing the two amplitude average values, and selecting the smaller one as the reference value for reference threshold calculation; S25, calculating the detection threshold, including a first detection threshold and a second detection threshold; The calculation formula of the first detection threshold T1 is: In the above formula, is a first detection threshold coefficient; The calculation formula of the second detection threshold T2 is: S26, performing amplitude comparison on the detection unit and the two-stage detection threshold to obtain the following results: In the above formula, n is the distance between the scattering point and the unit number, and A is defined as the target scattering point set and B is the alternative target scattering point set.

4. The method according to claim 1, characterized in that, The step S4 specifically comprises: S41, assuming that the kth target of the clustered target scattering point set S contains p scattering points, the p scattering points of the kth target are sorted according to the scattering point distance unit serial number, and the sorted distance unit serial number set is obtained{ , i = 1, 2…p}, the scattering point amplitude corresponding to the set is extracted, and the amplitude set{ , i = 1, 2…p} is obtained, i is the scattering point serial number; S42, calculate the target condensation distance R k The calculation formula is: S43, extracting the target relative RCS, and the calculation formula is: In the above formulae, Target relative RCS; S44, extract the radial length L k The calculation formula is: In the formula, dR is the distance unit length; S45, extracting the number of radial peaks PEAK k The calculation formula is: In the above formula, represents the number of i satisfying the condition in the parentheses, and j is the serial number. S46, extract the amplitude distribution entropy The calculation formula is: In the above formula, is the ratio of the amplitude of the scattering centers to the total amplitude. S47, extracting amplitude distribution normalized variance The calculation formula is: In the above formula, is the mean value of the target scatter point amplitude; S48, extracting the main peak energy ratio C, and the calculation formula is: In the above formula, is the peak point amplitude.

5. The method according to claim 1, characterized in that, The expert knowledge base is established by training the feature vectors obtained by detection, clustering, and feature extraction from the measured echo data of the known typical obstacle targets obtained at multiple illumination angles and illumination distances.

6. The method according to claim 1, characterized in that, The attribute recognition in the step S5 is specifically: using a decision tree for classification recognition, the decision tree including a root node, a set of intermediate nodes and some terminal nodes, the decision tree recognition process starting from the root node of the decision tree, marking a split predicate according to the corresponding feature attribute of the to-be-classified item obtained by feature extraction, and selecting an output branch according to the size of the feature attribute value and the decision threshold of the branch; using different feature attributes in turn, performing selection of the output branch from top to bottom until reaching a leaf node, and storing the class of the leaf node as the result of the decision tree classification recognition.

7. The method according to claim 6, characterized in that, The use order of the feature attributes is: target radial length, relative RCS, peak point number, main peak energy ratio, dispersion coefficient, amplitude normalized variance and amplitude distribution entropy.

8. The method according to claim 1, characterized in that, The step S6 is specifically: S61, processing and confirming the target data after attribute recognition in a plurality of repeated periods of radar scanning from left to right or from right to left for one round, judging whether there is a power line target in the targets detected by one round scanning, if yes, entering step S62, otherwise entering step S69; S62, judging whether the number of power line targets is greater than 2, if yes, entering step S63, otherwise entering step S64; S63, judging whether the azimuth angle of the power line target satisfies the Bragg scattering angle relationship, if yes, entering step S69, otherwise entering step S64; S64, judging whether there is a power tower in the targets detected by one round scanning, if yes, entering step S65, otherwise entering step S68; S65, projecting the power line and power tower targets to the aircraft geographic coordinate system, and then entering step S66; S66, connecting the power towers in the east-north plane after projecting the power towers to the aircraft geographic coordinate system, and then entering step S67; S67, judging whether the power line target is in the field of the inter-tower connection line, if yes, the power line target is a determined target, entering step S69, otherwise entering step S68; S68, marking the target that cannot be determined as a power line as a suspected power line target, and entering step S69; S69, outputting the target recognition result.

9. The method according to claim 8, characterized in that, The target marked as a suspected power line will be confirmed in the radar display control terminal for one week of data, if the power line target cannot be determined as a determined power line target in the same scanning airspace for three consecutive weeks of data, it will be discarded as interference.

Images