A method for detecting small targets on the sea surface based on combined convex hull
By extracting features such as relative weighted average amplitude, relative variation coefficient and whitening peak height ratio in marine radar, and combining clustering and convex hull algorithms, a non-convex judgment area is formed, which solves the low signal-to-mism ratio in sea surface small target detection. and false alarm issues improve detection performance and efficiency.
Patent Information
- Application Number
- CN202210301533.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-25
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-03-25
AI Technical Summary
In high-resolution marine radar, the difficulty in detecting small targets on the sea surface such as ice floes, boats and aircraft wreckage is that low signal-to-miss ratio leads to low detection probability and many false alarm points. It is difficult for the existing technology to extract effective features and obtain controllable false alarms. Non-convex judgment area.
The three-dimensional feature space is constructed by using a sea surface small target feature detection method based on a combined convex hull. By extracting three features of relative weighted average amplitude, relative variation coefficient and whitening peak height ratio, a three-dimensional feature space is constructed, and non-convexity is formed through clustering algorithms and convex hull algorithms. Judgment area to achieve robust detection of small targets on the sea surface.
Improve the probability of small target detection on sea surface under low signal-to-miscellaneous ratio conditions, provides new ideas for non-convex judgment design, reduces the computational complexity in the time frequency domain, and obtains higher detection performance and smaller false alarms. Rate.
Smart Images

Figure CN114594463B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of radar signal processing, and in particular to a method for detecting features of small sea targets based on a combined convex hull. Background Art
[0002] For high-resolution ocean radars, it is a long-term and difficult task to detect small targets on the sea surface, such as floating ice, small boats, frogmen, and aircraft wreckage, at a small angle of grazing. The main reason is that small targets often move slowly and have a very small signal-to-clutter ratio (SCR). In this case, detection of sea radars often results in low detection probability and many false alarm points. Therefore, small target detection under high-resolution sea clutter background has always been a hot topic and difficulty for scholars at home and abroad.
[0003] At present, feature-based detection has been considered as an effective method, which compresses the difference between sea clutter and target echo into features and then detects them in feature space. As the number of joint features increases, the performance of the detector is improved to a certain extent. However, more joint features bring difficulties to the design of the decision area and also require more time. Considering the performance improvement and computational cost, the combination of three features is a compromise. Therefore, the difficulty of the detector lies mainly in extracting effective features and obtaining the three-dimensional (3D) false alarm controllable decision area. For the first difficulty, many methods have been developed to extract features in different domains under long-term observation. From the perspective of time domain, some scholars have proposed relative average amplitude (RAA) to describe the difference in power between clutter and target. Some scholars have proposed fractal Hurst index to describe the fractal characteristics of sea clutter time series. From the perspective of frequency domain, some scholars use relative doppler peak height (RDPH) and relative vector entropy (RVE) to reflect geometric features. These two features plus RAA form a typical three-feature-based detector, opening up the feature detection framework. From the perspective of the time-frequency domain, some scholars have extracted three time-frequency features from the intensity and geometric properties of the time-frequency image. Generally, the two-dimensional time-frequency domain can obtain more information than the one-dimensional time-frequency domain, but the two-dimensional time-frequency is obviously computationally complex. For the second difficulty, the fast convex hull algorithm is the most typical 3D single classifier algorithm, which can obtain a decision area where false alarms can be controlled. Subsequently, some scholars proposed a one-class support vector machine (OSVM) algorithm that can obtain a non-convex decision area and thus improve performance. However, in order to control the false alarm rate, the classifier needs to search a large number of parameters and the false alarm control accuracy is limited. Therefore, how to design a non-convex algorithm with accurate and controllable false alarms is the focus of research. Summary of the invention
[0004] Purpose of the invention: In order to overcome the defects existing in the background technology, the present invention discloses a feature detection method for small sea targets based on combined convex hull. The method can make full use of the characteristics of time domain and frequency domain through three newly extracted features, and the combined convex hull algorithm integrates the characteristics of the original convex hull to accurately control false alarms, and at the same time obtains a non-convex judgment area, thereby realizing robust detection of small sea targets under different detection environments.
[0005] Technical solution: The method for detecting small sea surface targets based on combined convex hull disclosed in the present invention comprises the following steps:
[0006] S1. Assume that the radar receives N consecutive pulses in a range unit. These N pulses constitute an observation vector z = [z(1), z(2), ..., z(N)] T , called the unit to be detected CUT; at the same time, the observation vector z of P reference units around CUT is obtained p ,p-1,2,...,P, construct the target detection problem;
[0007] S2, the CUT observation vector z in S1 and the reference unit observation vector z p Convert to the frequency domain and calculate the Doppler amplitude spectrum; use the spectrum vector of the reference unit to whiten the spectrum vector of the CUT to obtain the whitened spectrum;
[0008] S3, in the time domain, the observation vector in S1 is processed in segments, the average amplitude value of each pulse echo is set as the weighting factor, and the relative weighted average amplitude is calculated;
[0009] In the frequency domain, the relative coefficient of variation and the whitening peak height ratio are extracted for the spectrum in S2 and the whitened spectrum respectively to measure the geometric difference of the spectrum;
[0010] S4, construct a three-dimensional feature space using the three features extracted in S3 to form a feature vector ξ as a test statistic;
[0011] S5. Under the condition of a given false alarm rate, the sea clutter training samples are first divided into K clusters using a clustering algorithm, and then the convex hull algorithm is performed on each cluster. Finally, the K convex hulls are combined to form the final non-convex decision region Ω.
[0012] S6. Calculate the test statistic ξ and the position of the decision area Ω to determine whether there is a target in the observation vector z:
[0013] If ξ∈Ω, it means that the observation vector z contains the target echo;
[0014] if This indicates that the observation vector z does not contain the target echo.
[0015] Among them, S1 classifies the target detection problem into the following binary hypothesis testing problem:
[0016]
[0017] Where c represents the sea clutter vector, s represents the target vector, and c p Indicates that the pth unit is the sea clutter vector, the H0 hypothesis indicates that the observation vector contains only sea clutter, and the H1 hypothesis indicates that the observation vector contains the target;
[0018] The essence of detection is to determine which category the CUT observation vector belongs to. The detection problem is regarded as a two-classification problem, where H0 is assumed to be the first category and H1 is assumed to be the second category.
[0019] Furthermore, for the CUT observation vector z in S1, the calculated Doppler amplitude spectrum is:
[0020]
[0021] Where f represents the frequency variable, f r is the pulse repetition frequency, exp(.) represents the exponential function;
[0022] The spectrum of CUT is whitened, and the whitened spectrum WS is calculated according to the following formula:
[0023]
[0024] Among them, Z p (f) represents the Doppler amplitude spectrum of the pth reference unit.
[0025] Further, it includes extracting time domain features:
[0026] Based on the assumption that sea clutter and target short time series remain unchanged, the observation vector in S1 is first divided into Q segments
[0027]
[0028] Where L represents the number of pulses per segment, satisfying L×Q=N, and the weighted average amplitude WAA is defined as:
[0029]
[0030] Among them, ω q represents the weighting factor of the qth segment, and the relative weighted average amplitude RWAA is defined as:
[0031]
[0032] Extract frequency domain features:
[0033] In order to measure the geometric differences between the sea clutter and the target echo spectrum, the relative coefficient of variation RCV is defined as:
[0034]
[0035] Among them, CV is the coefficient of variation. CV is used to describe the difference in spectrum fluctuations and is calculated according to the following formula:
[0036]
[0037] For the whitened spectrum in S2, the whitened peak height ratio WPHR is defined as:
[0038]
[0039] where #Θ represents the operator for the number of elements in a set Θ, and the set Θ is given by:
[0040] Θ=[-f r / 2,f max -Δ]∪[f max +Δ,f r / 2]
[0041]
[0042] Among them, f max is the frequency value with the maximum whitening spectrum, Δ is f max Surrounding protected area.
[0043] Furthermore, the three features extracted from S3, RWAA, RCV, and WPHR, are recorded as ξ1, ξ2, and ξ3 respectively; each feature is taken as a dimension to construct a three-dimensional feature space, then the CUT observation vector in S1 is converted into a 3D feature vector:
[0044]
[0045] This characteristic quantity is used as a test statistic for discrimination.
[0046] Furthermore, S5 specifically includes the following steps:
[0047] S5.1 radar obtains a large number of sea clutter samples. According to S2 to S4, a 3D training sample set is constructed. It is assumed that according to the original convex hull algorithm, a convex judgment area under which false alarms can be controlled has been obtained, and the area contains A samples;
[0048] S5.2 In sample set A, randomly select K samples as cluster centroids, denoted as The remaining samples are divided into K categories according to the criterion of the closest Euclidean distance to the cluster centroid.
[0049]
[0050] S5.3 In each cluster, calculate the new cluster centroid
[0051]
[0052] Among them, N k is the number of samples in the kth cluster, and all samples are redistributed by the above formula;
[0053] S5.4 Repeat step S5.3 until all samples in A have no cluster transformation and the set A is completely divided into K clusters, that is, A = A1∪A2∪...∪A K ;
[0054] S5.5 In the kth category, the convex hull is obtained by G k The polyhedron composed of triangular faces encloses N k samples, and wrap the K clusters with the convex hull algorithm, so that the K convex hulls obtain the combined convex hull, which is the final non-convex decision area Ω
[0055] Ω={CH(A1),CH(A2),...,CH(A K )}
[0056] Among them, CH(A) represents the convex hull formed for the sample set of A.
[0057] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows: three new features are proposed, which fully utilize the differences between clutter and targets in the time domain and frequency domain. At the same time, a large amount of computational cost in the time and frequency domain is avoided, and performance matching the time and frequency domain features can be obtained; a combined convex hull algorithm is proposed to realize a 3D non-convex decision area. Compared with the original convex hull decision area, the volume of the non-convex decision area is smaller, so higher detection performance can be obtained; this invention improves the detection probability of small targets on the sea surface under low signal-to-clutter ratio conditions, and provides a new idea for the design of non-convex decision detectors. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 is a flow chart of the method of the present invention;
[0059] Figure 2 It is the histogram of the three features in simulation 2;
[0060] Figure 3 It is a comparison diagram of two decision regions in the 3D feature space in simulation 3;
[0061] Figure 4 The detection results of the measured IPIX data based on the Hurst exponent detector in simulation 4, the original three-feature based detector and the detector proposed by the basic method of the present invention. DETAILED DESCRIPTION
[0062] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0063] like Figure 1 The method for detecting small sea surface targets based on combined convex hulls includes the following steps:
[0064] Step 1: Get the observation vector
[0065] Assume that the radar receives N consecutive pulses in a range unit. These N pulses constitute an observation vector, denoted as z = [z(1), z(2), ..., z(N)] T This distance cell is called the cell under test (CUT). There are P reference cells around the CUT, and their observation vectors are denoted by z p ,p=1,2,...,P. Then, the target detection problem can be classified as the following binary hypothesis testing problem:
[0066]
[0067] Where c represents the sea clutter vector, s represents the target vector, and c p Indicates that the pth unit is a sea clutter vector. The H0 hypothesis indicates that the observation vector contains only sea clutter; the H1 hypothesis indicates that the observation vector contains the target. The essence of detection is to determine which type of hypothesis the CUT observation vector belongs to.
[0068] Step 2: Frequency domain preprocessing
[0069] 2.1) For the CUT observation vector z in step (1), calculate its Doppler amplitude spectrum as:
[0070]
[0071] Where f represents the frequency variable, f r is the pulse repetition frequency, and exp(.) represents the exponential function.
[0072] 2.2) Perform whitening preprocessing on the spectrum of CUT to suppress sea clutter. The whitened spectrum (WS) is calculated according to the following formula:
[0073]
[0074] Among them, Z p (f) represents the Doppler amplitude frequency of the pth reference unit.
[0075] Step 3: Extract three features
[0076] 3.1) Time Domain Characteristics
[0077] Based on the assumption that sea clutter and target short time series remain unchanged, the observation vector in step 1) is first divided into Q segments
[0078]
[0079] Where L represents the number of pulses per segment, satisfying L×Q=N. Then, the weighted average amplitude (WAA) is defined as:
[0080]
[0081] Among them, ω q represents the weighting factor of the qth segment. Finally, the relative weighted average amplitude (RWAA) is defined as:
[0082]
[0083] 3.2) Frequency domain characteristics
[0084] 3.2a) In order to measure the geometric differences between the sea clutter and the target echo spectrum, the relative coefficient variance (RCV) is defined as:
[0085]
[0086] Among them, CV is the coefficient of variation (CV), which is used to describe the difference in spectrum fluctuations and is calculated according to the following formula:
[0087]
[0088] 3.2b) For the whitened spectrum in step (2.2), define the whitened peak height ratio (WPHR) as:
[0089]
[0090] where #Θ represents the operator for the number of elements in a set Θ, and the set Θ is given by:
[0091] Θ=[-f r / 2,f max -Δ]∪[f max +Δ,f r / 2]
[0092]
[0093] Among them, f max is the frequency value with the maximum whitening spectrum, Δ is f max Surrounding protected area.
[0094] Step 4: Construct a three-dimensional test statistic
[0095] The three features in step (3), RWAA, RCV, WPHR, are recorded as ξ1, ξ2, ξ3 respectively. Each feature is a dimension to construct a three-dimensional feature space. Then, the CUT observation vector in step 1) is converted into a 3D feature vector:
[0096]
[0097] This characteristic quantity is used as a test statistic for discrimination.
[0098] Step 5: Design non-convex decision area
[0099] In order to further minimize the convex decision region, the combined convex hull algorithm is implemented through the K-means clustering algorithm to construct the non-convex decision region.
[0100] 5.1) After the radar is turned on, a large number of sea clutter samples are obtained, and a 3D training sample set is constructed according to steps 2 to 4. Assume that according to the original convex hull algorithm, a convex decision area under which false alarms can be controlled has been obtained, and the area contains A samples.
[0101] 5.2) In the sample set A, randomly select K samples as cluster centroids, denoted as The remaining samples are divided into K categories according to the criterion of the closest Euclidean distance to the cluster centroid:
[0102]
[0103] 5.3) In each cluster, calculate the new cluster centroid:
[0104]
[0105] Among them, N k is the number of samples in the kth cluster. Then, all samples are redistributed by the above formula.
[0106] 5.4) Repeat step (5.3) until all samples in A have no cluster transformation. Finally, the set A is completely divided into K clusters, that is, A = A1∪A2∪...∪A K .
[0107] 5.5) In the kth category, the convex hull is obtained by G k The polyhedron composed of triangular faces encloses N k The K clusters are wrapped by the convex hull algorithm, so that the K convex hulls obtain the combined convex hull, which is the final non-convex decision area Ω
[0108] Ω={CH(A1),CH(A2),...,CH(A K )}
[0109] Among them, CH(A) represents the convex hull formed for the sample set of A.
[0110] Step 6: Determine the test result
[0111] Calculate the test statistic ξ and the position of the decision area Ω to determine whether there is a target in the CUT.
[0112] If ξ∈Ω, it means there is no target echo in the CUT.
[0113] if This indicates that the CUT contains target echo.
[0114] The effects of the present invention are further described below in conjunction with the experimental results of measured data.
[0115] This example uses a database collected by the IPIX radar in 1993, which is publicly available on the Internet. This database is currently internationally recognized as small target measured data. The data name used in the experiment is 19931118_023604_stareC0000280.mat. The pulse repetition frequency of the X-band radar is 1000Hz, the range resolution is 30m, the radar works in the dwell mode, and the HH polarization. The test target is a steel ball with a diameter of 1m, and the signal-to-noise ratio is 4dB, which is a low signal-to-noise ratio.
[0116] Simulation 1, such as Figure 1 As shown, a flowchart of a three-feature detector based on a combined convex hull is given according to the present method.
[0117] Simulation 2 analyzes the ability of the three new features to separate sea clutter and target echoes. The results are as follows: Figure 2 As shown, where:
[0118] Figure 2 (a) is a histogram of the RWAA feature. RWAA measures the power difference between sea clutter and target echoes. As the signal-to-clutter ratio increases, this difference becomes more obvious.
[0119] Figure 2 (b) is the histogram of the RCV feature, which measures the difference in the fluctuation of the two echo spectra. Since the target echo spectrum contains sea clutter, the spectrum difference between the two cases is small under low signal-to-clutter ratio conditions.
[0120] Figure 2 (c) is the histogram of WPHR features. It can be seen that the two hypotheses under WPHR have the farthest center distance, so it has the best feature separation ability.
[0121] The combined use of three effective features can achieve robust detection performance under different signal-to-noise ratio conditions.
[0122] Simulation 3 demonstrates the acquisition of the decision area where false alarms can be controlled. The results are as follows: Figure 3 As shown in the figure. The decision area is constructed using 10210 sea clutter samples, of which 10 samples are false alarm points. First, the sample set is divided into 3 clusters using the K-means clustering algorithm, and then the clusters are transformed into combined convex hulls using the convex hull algorithm. As can be seen from the figure, the combined convex hull algorithm can further reduce the area of the convex hull and achieve a non-convex decision area for accurately controlling false alarm points.
[0123] Simulation 4 verifies the detection algorithm proposed in this invention. The cumulative number of pulses is 256 and the false alarm rate is set to 10. -3 , the results are as follows Figure 4 shown. Figure 4 The detection results based on the Hurst exponent detector, the original three-feature based detector and the proposed detector are included.
[0124] exist Figure 4 In (a), the performance of the detection based on the Hurst exponent is the worst, with a detection probability of 0.131, and the detector cannot work. The main reason for the poor performance loss of this detector is that the detection ability of a single feature is poor, and the performance of this detector is seriously lost below the second level.
[0125] exist Figure 4 In (b), the original three-feature based detector performs suboptimally, with a detection probability of 0.469. Compared with the Hurst exponent based detector, the key to the performance improvement lies in the joint use of the three features, which exploits the characteristics of the time domain and the frequency domain. However, the performance of this detector is highly dependent on the signal-to-clutter ratio, and the performance loss is large under low signal-to-clutter ratio conditions.
[0126] exist Figure 4 In (c), the proposed three-feature detector based on combined convex hull has the best performance with a detection probability of 0.941. Compared with the Hurst exponent-based detector and the original three-feature-based detector, the proposed detector has significantly improved performance under low signal-to-clutter ratio conditions. In fact, this advantage mainly comes from the carefully designed three features and non-convex decision regions. Therefore, considering the actual detection environment, the proposed detector can be used to replace the original three-feature-based detector.
Claims
1. A method for detecting small targets on the sea surface based on combined convex hull. It is characterized in that The following steps are involved: S1. Assume that the radar receives N consecutive pulses in a range unit. These N pulses constitute an observation vector z = [z(1), z(2), ..., z(N)] T , called the unit to be detected CUT; at the same time, the observation vector z of P reference units around CUT is obtained p ,p=1,2,...,P, construct the target detection problem; S2, the CUT observation vector z in S1 and the reference unit observation vector z p Convert to the frequency domain and calculate the Doppler amplitude spectrum; use the spectrum vector of the reference unit to whiten the spectrum vector of the CUT to obtain the whitened spectrum; S3, in the time domain, the observation vector in S1 is processed in segments, the average amplitude value of each pulse echo is set as the weighting factor, and the relative weighted average amplitude is calculated; In the frequency domain, the relative coefficient of variation and the whitening peak height ratio are extracted for the spectrum in S2 and the whitened spectrum respectively to measure the geometric difference of the spectrum; S4, construct a three-dimensional feature space using the three features extracted in S3 to form a feature vector ξ as a test statistic; S5. Under the condition of a given false alarm rate, the sea clutter training samples are first divided into K clusters using a clustering algorithm, and then the convex hull algorithm is performed on each cluster. Finally, the K convex hulls are combined to form the final non-convex decision region Ω. S6. Calculate the test statistic ξ and the position of the decision area Ω to determine whether there is a target in the observation vector z: If ξ∈Ω, it means that the observation vector z contains the target echo; if This indicates that the observation vector z does not contain the target echo; This includes extracting time domain features: Based on the assumption that sea clutter and target short time series remain unchanged, the observation vector in S1 is first divided into Q segments Where L represents the number of pulses per segment, satisfying L×Q=N, and the weighted average amplitude WAA is defined as: Among them, ω q represents the weighting factor of the qth segment, and the relative weighted average amplitude RWAA is defined as: Extract frequency domain features: In order to measure the geometric differences between the sea clutter and the target echo spectrum, the relative coefficient of variation RCV is defined as: Among them, CV is the coefficient of variation, which is used to describe the difference in spectrum fluctuations and is calculated according to the following formula: For the whitened spectrum in S2, the whitened peak height ratio WPHR is defined as: where #Θ represents the operator for the number of elements in a set Θ, and the set Θ is given by: Θ=[-f r / 2,f max -Δ]∪[f max +Δ,f r / 2] Among them, f max is the frequency value with the maximum whitening spectrum, Δ is f max surrounding protected areas; The three features extracted from S3, RWAA, RCV, and WPHR, are denoted as ξ 1 ,ξ 2 ,ξ 3 ; Take each feature as a dimension to construct a three-dimensional feature space, then the CUT observation vector in S1 is converted into a 3D feature vector: This eigenvector is used as a test statistic for discrimination; S5 specifically includes the following steps: S5.1 radar obtains a large number of sea clutter samples. According to S2 to S4, a 3D training sample set is constructed. It is assumed that according to the original convex hull algorithm, a convex judgment area under which false alarms can be controlled has been obtained, and the area contains A samples; S5.2 In sample set A, randomly select K samples as cluster centroids, denoted as The remaining samples are divided into K categories according to the criterion of the closest Euclidean distance to the cluster centroid. S5.3 In each cluster, calculate the new cluster centroid Among them, N k is the number of samples in the kth cluster, and all samples are redistributed by the above formula; S5.4 Repeat step S5.3 until all samples in A have no cluster transformation and the set A is completely divided into K clusters, that is, A = A 1 ∪A 2 ∪...∪A K ; S5.5 In the kth category, the convex hull is obtained by G k The polyhedron composed of triangular faces encloses N k samples, and wrap the K clusters with the convex hull algorithm, so that the K convex hulls obtain the combined convex hull, which is the final non-convex decision area Ω Ω={CH(A 1 ),CH(A 2 ),...,CH(A K )} Among them, CH(A) represents the convex hull formed for the sample set of A.
2. According to claim 1, the method for detecting small sea surface targets based on combined convex hull, Features: In S1, the target detection problem is classified into the following binary hypothesis testing problem: Where c represents the sea clutter vector, s represents the target vector, and c p Indicates that the pth unit is the sea clutter vector, H 0 Assume that the observation vector contains only sea clutter, H 1 Assume that the observation vector contains the target; The essence of detection is to determine which category the CUT observation vector belongs to. The detection problem is regarded as a two-classification problem. 0 Assuming the first type, H 1 Assume the second category.
3. According to claim 2, the method for detecting small sea surface targets based on combined convex hull, Features: For the CUT observation vector z in S1, the calculated Doppler amplitude spectrum is: Where f represents the frequency variable, f r is the pulse repetition frequency, exp(.) represents the exponential function; The spectrum of CUT is whitened, and the whitened spectrum WS is calculated according to the following formula: Among them, Z p (f) represents the Doppler amplitude spectrum of the pth reference unit.
