A radar approaching extended target detection and segmentation method, device, system and medium based on multiple features
Through multi-feature detection method and maximum interchange ratio threshold calculation, combined with the time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude and modulation domain-time frequency ridge characteristics, the high-resolution radar in the sea complex is solved. The problem of low detection accuracy of adjacent extended targets under the wave background achieves higher detection accuracy and real-time performance.
Patent Information
- Application Number
- CN202311407915.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-26
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
High-resolution radars have the problem of low target segmentation accuracy in the context of sea clutter. The existing technology is difficult to effectively distinguish sea clutter from targets, especially when multiple targets are approaching, missed detection and misjudgment are prone to occur. .
Multi-feature detection method is adopted, combining four types of features: time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude and modulation domain-time frequency ridge, and is calculated through feature weighted fusion and maximum intersection ratio (IOU) Segment thresholds to achieve more accurate target detection segmentation.
It effectively improves the real-time and segmentation accuracy of radar near-expanded target detection, avoids model pre-training problems, significantly improves detection performance, and can more accurately distinguish sea clutter from targets.
Smart Images

Figure CN117351208B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of radar target detection, and in particular relates to a radar adjacent extended target detection and segmentation method based on multiple features in the field of target detection and segmentation of radar RD (Range Doppler) images. Background Art
[0002] High-resolution radar provides material for refined cognition of the marine environment, and the key to converting refined perception into detection capability is to accurately obtain complete information about the target to be detected. Several unique advantages of broadband high-resolution radar in target detection are: (1) High resolution achieves the clutter segmentation effect, and the sea clutter level is significantly reduced. For example, if the azimuth resolution remains unchanged, the range resolution is increased by 10 times, the area of the clutter scattering unit will be reduced by 1 / 10, and the clutter power will be reduced by 10dB. (2) High range resolution significantly reduces the fluctuation of the target RCS, which can overcome the target loss caused by the flickering of the target RCS in narrowband conditions. (3) High-resolution radar provides more detailed information about the observed object, which is conducive to identifying targets in complex environments.
[0003] In the high-resolution broadband radar system, the target energy is distributed in multiple adjacent range units, usually appearing as a range-extended target. Range-extended target detection is a universal problem faced by broadband high-resolution radar. The need for detection of adjacent extended targets in the background of sea clutter is particularly prominent. The main difficulty is that the range scattering function of the target is unknown and often has attitude sensitivity. In other words, a slight change in the target's attitude may cause a large change in its range scattering function, making it impossible to effectively separate adjacent targets.
[0004] The existing high-resolution target detection in the sea clutter background is mainly based on CFAR (Constant False Alarm Rate) detection based on energy features and target detection based on multiple features. CFAR detection is mainly based on the premise that the target energy is greater than the background clutter and noise energy. When the radar echo contains the target, the echo intensity is strong, and the energy amplitude information can be used to distinguish the clutter from the target to a certain extent. However, there are certain shortcomings in some aspects: (1) The algorithm assumes that the sea clutter satisfies a certain statistical model. However, it is difficult for the existing statistical model to fully describe the current sea clutter, especially in the case of high sea conditions or small and weak targets, resulting in unsatisfactory detection results; (2) When two targets are close, one target may enter the reference unit of the other target, resulting in missed detection. In particular, due to the huge difference in target energy, small ships next to large ships can easily classify the two targets as one target. Therefore, the suppression and detection effects of this method are far from meeting the actual application requirements. At the same time, the targets and clutter in different echo data have different distinguishing capabilities under different domains such as time-frequency-energy-modulation and have certain complementarity. Features that may perform well in some data may perform poorly in other data, and the features of a single transform domain are difficult to fully represent the distinguishability between sea clutter and target echoes. Therefore, it is necessary to use a multi-feature detection method, which is an effective way to improve detection performance.
[0005] Target detection based on multiple features is a new detection method that gradually replaces the traditional energy-based detection method with the development of machine learning. When the radar illuminates the sea surface, the high-resolution radar echo contains a large amount of information about the sea surface. Extracting useful information from multiple dimensions can further improve the detection performance in a single energy dimension. At present, through feature screening, we have selected four typical features from the time domain, frequency domain, energy domain and modulation domain for floating targets on the sea surface, namely time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude and modulation domain-time-frequency ridge. The above features have obvious differences in sea clutter and sea surface targets, which can effectively distinguish clutter from targets, and at the same time, the adjacent boundaries of adjacent targets can be further clarified.
[0006] Although these multi-feature based target detection algorithms have achieved excellent performance, they still have some common problems: first, multi-feature detection and segmentation rely on feature classifiers, which require highly correlated data for training. Moreover, as the feature dimension increases, the classification surface presents complex surface characteristics, which makes it difficult to achieve the function of actual training and inspection. Secondly, multi-feature detection is proposed on the basis of single feature detection, and different features are selected in turn to process the current data. If the detection results of various features are completely consistent, the basic verification test has obtained the most accurate result. However, due to the differences in the applicability and distinguishability of each feature and the proximity of multiple targets, simply selecting the intersection or union of multi-feature results will have a larger error than the actual result. Therefore, how to use multi-feature detection and segmentation algorithms in real time, quickly, accurately and effectively still has a lot of room for improvement.
[0007] In the above background, based on multi-feature target detection, this paper aims at two problems faced by the algorithm due to real-time detection and segmentation of nearby targets. The first is that the feature classifier needs to be trained offline to achieve target detection and segmentation. The second is that different features have different degrees of distinction between clutter and targets, and the randomness in the clustering process causes the detection and segmentation results to be inconsistent with the real results. A multi-feature radar nearby extended target detection and segmentation method is proposed. The obvious differences between sea clutter and targets under four types of features, time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude and modulation domain-time-frequency ridge, are used to detect and segment the target area in turn. The difference between different features is combined to achieve multi-feature weighted detection results. Finally, the maximum intersection-over-union (IOU) is introduced and the weighted segmentation threshold is approximately calculated, so as to obtain more accurate target detection and segmentation results. Experimental results show that the algorithm proposed in this paper can effectively avoid the pre-training problem of the model in the multi-feature detection process, improve the real-time performance and detection and segmentation accuracy, perform well on multiple test data, and have a significant effect in radar nearby extended target detection and segmentation. Summary of the invention
[0008] In order to solve the above technical problems, the present invention proposes a radar proximity extended target detection and segmentation method, device, system and medium based on multiple features.
[0009] The first aspect of the present invention proposes a radar proximity extended target detection and segmentation method based on multiple features, comprising the following steps:
[0010] S1. Divide the sea surface into suspicious target areas and pure sea clutter areas.
[0011] Specifically, the sparsity of sea surface targets is utilized to perform coherent accumulation of high-resolution radar echoes, reduce the detection threshold of traditional CFAR, and conduct preliminary inspection of the sea surface area to obtain suspicious target areas and pure sea clutter areas.
[0012] S2. The pulse pressure data is used to extract four features of the selected pure sea clutter area: time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude, and modulation domain-time-frequency ridge. The target-clutter division threshold of each feature is calculated respectively. The formula is as follows:
[0013]
[0014] in, is the target feature mode based on past experience under the Nth feature, is the mode of clutter features based on past experience under the Nth feature, y N is the detection and segmentation threshold of the Nth feature, X N is the vector group of the Nth feature of the selected sea clutter area, k is a floating coefficient, and the value range is 1 to 3. S3, the entire echo area except the selected sea clutter area is set as the suspicious target area, and the above four features are extracted from the range image of the suspicious target area respectively, and the eigenvalue matrix {Z N}. Use the obtained feature segmentation threshold to segment the suspicious target area
[0015] The rough division of the target distance channel is as follows:
[0016]
[0017] Among them, R N,i is the detection and segmentation result of the Nth feature in the i-th distance channel, Z N,i is the eigenvalue of the i-th distance channel in the N-th feature matrix, and the suspicious target area has a total of M channels.
[0018] S4. Since different features have different abilities to distinguish targets, it is necessary to align the clustering results of different features. N,i Clustered into {R N}, while ensuring that the number of targets detected by each feature segmentation is the same, and obtaining the position of each target under each feature.
[0019] First, the difference between the target and clutter of each feature is obtained. The formula is as follows:
[0020]
[0021] Secondly, the feature difference of each feature is calculated, and the formula is as follows:
[0022]
[0023] Finally, the multi-feature detection position channels are weighted and fused using different feature differences. The formula is as follows:
[0024]
[0025] S5. For isolated targets or targets with the strongest energy, the intersection-over-union ratio (IoU) of the cluster set is used as the basis, and the location area of the isolated target or the target with the strongest energy is obtained by using the CFAR algorithm, and the intersection-over-union ratio under different features is calculated in turn. When the comprehensive intersection-over-union ratio (IOU) is the maximum, the corresponding threshold value is the threshold value under the action of different features. When different thresholds are selected, different intersection-over-union ratios will be obtained. Combined with the feature difference, when the comprehensive intersection-over-union ratio is the maximum, the corresponding threshold value is the threshold value mentioned here. The intersection-over-union ratio is the ratio of the intersection and union of two sets, which is expressed as follows:
[0026]
[0027] in, Detect clustering results for the Nth feature of the selected target Compared with CFAR detection results R t The intersection and ratio of .
[0028] The fusion intersection ratio of the target is:
[0029] p from {min(w j ),1}, the values are incremented by 0.1 to obtain the IOU t The maximum corresponding Value, used as the segmentation threshold of the fusion result.
[0030] That is, different thresholds are selected in turn by traversal. Under the threshold conditions, the IoU of each feature of a typical target may be slightly different. The comprehensive IoU is obtained by combining the feature differences. The subsequent weighted fusion results are processed using the threshold corresponding to the maximum comprehensive IoU to obtain the precise area of each target.
[0031] To obtain the detection and segmentation results of each target, the formula is as follows:
[0032]
[0033] The second aspect of the present invention proposes a multi-feature based radar proximity extended target detection and segmentation device, including a sea surface segmentation module, which is used to perform coherent accumulation through sea surface high-resolution radar echoes, reduce the detection threshold of traditional CFAR, perform preliminary inspection on the sea surface area, and obtain suspicious target areas and pure sea clutter areas;
[0034] The feature extraction module is used to extract features of multiple dimensions from the pure sea clutter area using pulse pressure data, and calculate the target-clutter feature segmentation threshold of each feature;
[0035] A coarse segmentation module is used to extract features of multiple dimensions from the range image of the suspicious target area, form an eigenvalue matrix, and use the feature segmentation threshold to perform coarse segmentation of the target range channel; the multiple dimensions include time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude, and modulation domain-time-frequency ridge;
[0036] The weighted fusion module is used to cluster the coarse segmentation results, obtain the target and clutter difference and feature difference of each feature, and use different feature differences to weightedly fuse the multi-dimensional detection distance channels. The formula is as follows:
[0037]
[0038] Among them, R j is the detection and segmentation result of the jth feature, w j is the feature difference of the jth feature;
[0039] The detection and segmentation module is used to select the IoU of various features under different segmentation thresholds for isolated targets or targets with the strongest energy, calculate the comprehensive IoU based on the feature difference, and select the segmentation threshold corresponding to the maximum comprehensive IoU to further process the adjacent targets to obtain the optimal solution for detection and segmentation.
[0040] The detection and segmentation result formula for each target is as follows:
[0041]
[0042] in, is the final detection segmentation distance interval of the t-th target, R t is the multi-feature fusion interval of the t-th target, and p0 is the multi-feature fusion segmentation threshold.
[0043] The third aspect of the present invention proposes a system for radar proximity extended target detection and segmentation based on multiple features, including a memory and a controller, wherein the memory stores a computer-readable program, and when the computer-readable program is called by the controller, the processor can execute the steps of the radar proximity extended target detection and segmentation method based on multiple features as described in the first aspect of the present invention.
[0044] The fourth aspect of the present invention proposes a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the multi-feature-based radar proximity extended target detection and segmentation method as described in the first aspect of the present invention are implemented.
[0045] The present invention aims at the problem of low accuracy of detection and segmentation of nearby targets during high-resolution radar observation, and uses multiple features to distinguish targets from clutter, weighted feature differences and intersection-over-union ratio to obtain the optimal segmentation detection threshold. The obvious differences between sea clutter and targets under four types of features, time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude and modulation domain-time-frequency ridge, are used to detect and segment the target area in turn, and the differences of different features are combined to achieve multi-feature weighted detection results. Finally, the maximum intersection-over-union ratio (IOU) is introduced and the weighted segmentation threshold is approximately calculated, so as to obtain a more accurate target detection and segmentation result. Experimental results show that the algorithm proposed in this paper can effectively avoid the problem of pre-training of the model in the process of multi-feature detection, improve real-time performance and detection and segmentation accuracy, perform well on multiple test data, and have a more obvious effect in radar nearby extended target detection and segmentation. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a detailed schematic diagram of the radar proximity extended target detection and segmentation method based on multiple features.
[0047] Figure 2 It is a flow chart of the radar proximity extended target detection and segmentation method based on multiple features.
[0048] Figure 3 This is the target-clutter distinction effect diagram with selected features.
[0049] Figure 4 This is the implementation diagram of the energy-based (CFAR) detection algorithm.
[0050] Figure 5 This is the implementation diagram of the multi-feature detection and segmentation algorithm. DETAILED DESCRIPTION
[0051] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0052] The main idea of the target detection and segmentation method based on multiple features is: on the basis of the traditional single-feature target detection method, such as the CFAR detection method, multiple feature detection and weighting mechanisms are integrated to prevent the phenomenon of missed target detection. By determining the multi-feature segmentation threshold based on the intersection-over-union ratio of isolated or strongest signal-to-clutter targets, the effective segmentation ability of adjacent targets is further improved. The so-called proximity extension, proximity means that the two targets are close in the distance dimension, and extension means that the targets show extensibility under high-resolution observation conditions. When the two conditions exist at the same time, the two targets are difficult to distinguish in the distance dimension, and they are easily combined into the same target when clustering. Under high-resolution observation conditions, the targets show extensibility, and a single target occupies multiple distance units; when the distance between two targets is close, they are called adjacent targets. When adjacent targets are extensible, it is difficult for traditional methods to effectively distinguish them, and the two targets are often judged as one target, affecting subsequent further identification.
[0053] Example 1 proposes a multi-feature based radar proximity extended target detection and segmentation method, including the following steps, see Figure 1-Figure 2 .
[0054] S1. Divide the sea surface into suspicious target areas and pure sea clutter areas.
[0055] Specifically, the sparsity of sea surface targets is utilized to perform coherent accumulation of high-resolution radar echoes, reduce the detection threshold of traditional CFAR, and conduct preliminary inspection of the sea surface area to obtain suspicious target areas and pure sea clutter areas.
[0056] S2. The pulse pressure data is used to extract four features of the selected pure sea clutter area: time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude, and modulation domain-time-frequency ridge. The target-clutter division threshold of each feature is calculated respectively. The formula is as follows:
[0057]
[0058] in, is the target feature mode based on past experience under the Nth feature, is the mode of clutter features based on past experience under the Nth feature, y N is the detection and segmentation threshold of the Nth feature, X N is the vector group of the Nth feature of the selected sea clutter area, k is a floating coefficient, and the value range is 1 to 3. The extracted classification features include four features: time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude, and modulation domain-time-frequency ridge:
[0059] (1) Modulation domain - time-frequency ridge characteristics
[0060] Assuming that the analysis window function γ(t) is stable (pseudo-stationary) within a short time interval, the window function is moved to achieve windowing and translation near the pulse pressure signal z(t), so that z(t)γ(t) is a stable signal within different finite time widths, thereby calculating the power spectrum at different moments, and arranging the obtained series of Fourier transform results to obtain a two-dimensional representation, that is, the time-frequency distribution diagram.
[0061] The short-time Fourier transform formula of the pulse pressure signal z(t) is STFT z (t,f):
[0062]
[0063] Where z(t) is the pulse pressure signal, γ(t) is the window function, γ*(t) is the conjugate function of γ(t), t is time, and f is frequency.
[0064] Assume that the time-frequency distribution of the signal is an M×N matrix with an amplitude of f(i,j) (j∈[1,M], j∈[1,N]). The Gaussian operator is used to smooth the time-frequency image amplitude f(i,j). The Gaussian operator is G(i,j):
[0065]
[0066] Where i is the number of rows in the time-frequency distribution matrix, j is the number of columns in the time-frequency distribution matrix, and σ is the standard deviation. The smoothing process can be controlled by adjusting σ. Smoothing the time-frequency distribution graph is to convolve the original distribution graph with the Gaussian function, that is, L(i,j)=G(i,j)*f(i,j), where L(i,j) is the image after convolution.
[0067] Since the ridge features of the image are directional, in order to better extract the ridge features of the image, all operators are made non-directional and rotation-translation invariant in a certain way. The two-dimensional Laplacian operator can well meet these conditions. Define w as the gradient of the time-frequency image L(i,j), its size is the gradient modulus ||w||, and its direction is the direction with the largest gradient change, that is,
[0068]
[0069] Where L x With L y are the first-order derivatives of the image in the x and y directions, respectively.
[0070] When the point is on both sides of the image ridge, the direction of w always points to the ridge, and when the point is on the ridge, its direction is along the ridge. Define the vector ν perpendicular to the direction of w, that is
[0071]
[0072] where ||ν|| is the magnitude of the vector ν, θ ν is the direction of vector ν. For points on the ridge line, the direction of ν is perpendicular to the ridge. When ν is on the ridge, its direction changes the most. Its second-order derivative L y There will be a relatively large amplitude. Therefore, as long as the appropriate threshold is selected, the ridge features of the image can be extracted. The second-order derivative of Laplacian of ν The operation formula is:
[0073]
[0074] In the formula and are the second-order derivatives of the image in the x and y directions, L xy Calculate the partial derivatives of the image in the x and y directions respectively.
[0075] For time-frequency distribution images with obvious ridge features, after selecting appropriate standard deviations σ and When the value is , the ridge line feature extraction of the image can be performed using the above algorithm, and the ridge line of the time-frequency image can be extracted well.
[0076] (2) Time domain - fractal characteristics:
[0077] A set of acquired radar sea clutter time domain pulse pressure signals are processed and the Yule-Walker equation method is used to process the radar sea clutter time domain pulse pressure signals to obtain the sea clutter time domain AR spectrum sequence.
[0078] The structure function F(m) and scale m of the sea clutter time-domain AR spectrum sequence are calculated according to the "random walk" model, and the log-log curve is used to check whether there is a linear relationship between them. If the curve has a linear relationship, it is determined that the time-domain AR spectrum sequence of the sea clutter data has fractal characteristics; otherwise, it is determined that the time-domain AR spectrum sequence does not have fractal characteristics.
[0079] If it is determined that the time domain AR spectrum sequence of sea clutter has fractal characteristics, the Hurst index H of the AR spectrum of sea clutter can be further obtained. AR :
[0080]
[0081] Where F(m) is the structure function of the sea clutter time domain AR spectrum sequence, and m is the scale. The Hurst exponent of the AR spectrum can be used as a characteristic parameter to detect targets in the sea clutter background.
[0082] (3) Frequency domain - Doppler spectral entropy
[0083] The Doppler amplitude of the pure clutter unit is distributed in a wider bandwidth, while the Doppler amplitude of the unit where the target is located is distributed in a narrower range. Entropy can describe the degree of chaos of the system or data. Similarly, the Doppler spectrum vector entropy can be used to determine whether the target exists, which is defined as
[0084]
[0085] in The frequency domain accumulated data is obtained by coherently integrating the time domain pulse pressure data.
[0086] (4) Energy range - relative average amplitude
[0087] For a signal of length N, the average amplitude is defined as follows
[0088]
[0089] Where x represents the echo time domain pulse pressure signal. Differentiating the magnitude of echo intensity is the main basis for traditional radar to detect targets. RAA calculates the relative average amplitude by calculating the echo intensity around the unit to be estimated. and Represent the average echo of the unit to be estimated and the average echo of the surrounding reference units respectively. The relative average amplitude is calculated as follows
[0090]
[0091] For clustering algorithms, density-based clustering methods focus on determining the density core and density reach. Take any starting data point as the core point and expand the core point. According to the given radius ε, any point whose distance to the core point is less than ε is a neighboring point. If there are enough points near the core point, clustering begins, and the selected core point becomes the first point of the cluster. If there are not enough points nearby, the algorithm will mark it as clutter. In both cases, the selected point will be marked as visited. Once clustering begins, the neighboring points of the core point will be divided into the same cluster. Then these new points serve as core points, expand to the surrounding area, and continue to include qualified points in this cluster.
[0092] Compared with other clustering algorithms, density-based clustering methods have some great advantages. First, there is no need to input the number of clusters to be divided. Second, there is no bias towards the shape of the clusters. Clusters of any size and shape can be found. Even if the data distribution does not meet the convex cluster structure, they will be included in the cluster, and outliers can be identified as noise.
[0093] S3, the entire echo area except the selected sea clutter area is set as the suspicious target area, and the four features used above are extracted from the range image of the suspicious target area respectively, and the eigenvalue matrix {Z N}. The obtained feature segmentation threshold is used to roughly divide the target distance channel of the suspicious target area. The formula is as follows:
[0094]
[0095] Among them, R N,i is the detection and segmentation result of the Nth feature in the i-th distance channel, Z N,i is the eigenvalue of the i-th distance channel in the N-th feature matrix, and the suspicious target area has a total of M channels.
[0096] S4. Since different features have different abilities to distinguish targets, it is necessary to align the clustering results of different features. N,i Clustered into {R N}, while ensuring that the number of targets detected by each feature segmentation is the same, and obtaining the position of each target under each feature.
[0097] First, the difference between the target and clutter of each feature is obtained. The formula is as follows:
[0098]
[0099] Secondly, the feature difference of each feature is calculated, and the formula is as follows:
[0100]
[0101] Finally, the multi-feature detection position channels are weighted and fused using different feature differences. The formula is as follows:
[0102]
[0103] S5. For isolated targets or targets with the highest energy, the intersection-and-union ratio (IOU) of the cluster set is used as the basis, and the location area of the isolated target or the target with the highest energy is obtained by using the CFAR algorithm, and the intersection-and-union ratio under different features is calculated in turn. The threshold corresponding to the maximum comprehensive intersection-and-union ratio (IOU) is selected. When different thresholds are selected, different intersection-and-union ratios under different features are obtained. Combined with the feature difference, the threshold corresponding to the maximum comprehensive intersection-and-union ratio is the threshold mentioned here. The intersection-and-union ratio is the ratio of the intersection and union of two sets, expressed as follows,
[0104]
[0105] in, Detect clustering results for the Nth feature of the selected target Compared with CFAR detection results R t The intersection and ratio of .
[0106] The fusion intersection ratio of the target is:
[0107] p from {min(w j ),1}, the values are incremented by 0.1 to obtain the IOU t The maximum corresponding Value, used as the segmentation threshold of the fusion result.
[0108] That is, different thresholds are selected in turn by traversal. Under the threshold conditions, the IoU of each feature of a typical target may be slightly different. The comprehensive IoU is obtained by combining the feature differences. The subsequent weighted fusion results are processed using the threshold corresponding to the maximum comprehensive IoU to obtain the precise area of each target.
[0109] To obtain the detection and segmentation results of each target, the formula is as follows:
[0110]
[0111] Please see Figure 3 The invention provides a radar approaching extended target detection and segmentation method based on multiple features. For the four features of time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude and modulation domain-time-frequency ridge, sea clutter and large, medium and small ship types have obvious differences in various feature domains, but the difference degree in each feature domain is not the same. The algorithm in this paper does not need to detect and segment each distance channel. Considering that the traditional CFAR can only detect the approximate area of the target, what this paper needs to do is to accurately segment the target object. Generally speaking, after the segmentation of a single feature, only the target area is marked, but the selection method of the segmentation threshold and the difference degree of target clutter between different features are different, resulting in that the detection results of different features cannot be 100% overlapped. This precise segmentation cannot play a good effect under certain approaching target conditions. Therefore, in order to effectively use different features to achieve target detection and segmentation, this paper introduces the feature difference weighting and the typical target maximum intersection-over-union ratio distinction threshold to solve the problem of how to select a target integrity without introducing too much clutter area when the detection results of different features cannot completely overlap.
[0112] Please see Figure 4 The detection and segmentation results of the radar proximity extended target detection method based on multiple features provided by the present invention can obtain the accurate detection area of the target object after the above processing. Compared with single feature detection, the detection and segmentation accuracy can be effectively improved.
[0113] Embodiment 2 proposes a multi-feature based radar proximity extended target detection and segmentation device, including a sea surface segmentation module for dividing the sea surface into a suspicious target area and a pure sea clutter area.
[0114] The feature extraction module is used to extract features of multiple dimensions from the pure sea clutter area using pulse pressure data, and calculate the target-clutter feature segmentation threshold of each feature.
[0115] The coarse segmentation module is used to extract features of multiple dimensions from the range image of the suspicious target area, form an eigenvalue matrix, and use the feature segmentation threshold to perform coarse segmentation of the target range channel.
[0116] The weighted fusion module is used to cluster the coarse segmentation results, obtain the difference between the target and clutter of each feature, and the feature difference, and use different feature differences to weightedly fuse the multi-dimensional detection distance channels.
[0117] The detection and segmentation module is used to select the IoU of various features under different segmentation thresholds for isolated targets or targets with the strongest energy, calculate the comprehensive IoU based on the feature difference, and select the segmentation threshold corresponding to the maximum comprehensive IoU for further processing of adjacent targets to obtain the optimal solution for detection and segmentation.
[0118] Example 3 proposes a system for radar proximity extended target detection and segmentation based on multiple features, including a memory and a controller, wherein the memory stores a computer-readable program, and when the computer-readable program is called by the controller, the processor can execute the steps of the radar proximity extended target detection and segmentation method based on multiple features as described in Example 1.
[0119] Embodiment 4 proposes a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the radar proximity extended target detection and segmentation method based on multiple features as described in Embodiment 1 are implemented.
[0120] The present invention aims at the problem of inconsistent target areas obtained by clustering targets under different feature conditions, calculates the difference of each feature as a weight factor, and uses the weighted average fusion method to fuse the single feature detection segmentation results into multi-feature results. The weighted average fusion method is simple and easy to implement, has strong real-time performance, and can suppress clutter that cannot be eliminated by a single feature to a certain extent. Finally, the intersection-and-union ratio is introduced, and the segmentation threshold is selected for isolated targets or targets with the strongest energy to obtain the maximum intersection-and-union ratio result. The threshold is used to further segment adjacent targets to obtain the optimal solution for detection segmentation.
[0121] It should be understood that the above embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention without departing from the spirit and scope of the technical solution of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A multi-feature based radar approaching extended target detection and segmentation method. It is characterized in that include, S1. Coherently accumulate high-resolution radar echoes on the sea surface to reduce the detection threshold of traditional CFAR, conduct preliminary inspection of the sea surface area, and obtain suspicious target areas and pure sea clutter areas; S2. Extracting features of multiple dimensions from the pure sea clutter region using pulse pressure data, and calculating the target-clutter feature segmentation threshold of each feature, wherein the multiple dimensions include time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude, and modulation domain-time-frequency ridge; S3, extracting features of multiple dimensions from the range image of the suspicious target area to form an eigenvalue matrix {Z N }, use the feature segmentation threshold to roughly divide the target distance channel; S4. Cluster the rough division results, calculate the target and clutter difference and feature difference of each feature, and use different feature differences to weightedly fuse the multi-dimensional detection distance channels. The formula is as follows: Among them, R j is the detection and segmentation result of the jth feature, w j is the feature difference of the jth feature; S5. For isolated targets or targets with the strongest energy, select the IoU ratio of various features under different segmentation thresholds, calculate the comprehensive IoU ratio based on the feature difference, select the segmentation threshold corresponding to the maximum comprehensive IoU ratio to further process the adjacent targets, and obtain the optimal solution for detection and segmentation. The detection and segmentation result formula for each target is as follows: in, is the final detection segmentation distance interval of the t-th target, R t is the multi-feature fusion interval of the t-th target, p 0 is the multi-feature fusion segmentation threshold.
2. The radar proximity extended target detection and segmentation method as claimed in claim 1, It is characterized in that The calculation formula of the target-clutter division threshold in S2 is: in, is the target feature mode based on past experience under the Nth feature, is the mode of clutter features based on past experience under the Nth feature, y N is the detection and segmentation threshold of the Nth feature, X N is the vector group of the Nth feature of the selected sea clutter area, k is a floating coefficient ranging from 1 to 3.
3. The radar proximity extended target detection and segmentation method as claimed in claim 2, It is characterized in that The rough division in S3 is as follows: Among them, R N,i is the detection and segmentation result of the Nth feature in the i-th distance channel, Z N,i is the eigenvalue of the i-th distance channel in the N-th feature matrix, and the suspicious target area has a total of M channels.
4. The radar approaching extended target detection and segmentation method as claimed in claim 3, It is characterized in that In S4, the target and clutter difference formula of each feature is as follows: Among them, Z N,i is the eigenvalue of the i-th distance channel in the N-th feature matrix, and the area to be detected has a total of M channels. is the mean of the vector group of the Nth feature of the selected sea clutter area; The feature difference formula for each feature is as follows: Among them, σ j is the difference between the target and clutter of the jth feature.
5. The radar proximity extended target detection and segmentation method as claimed in claim 4, It is characterized in that The S5 further includes: The calculation formula for the intersection-over-union ratio of different feature detection results is as follows: The weighted fusion formula of the intersection-and-union ratio of comprehensive multiple features is as follows: Among them, w j is the feature difference of the jth feature, is the intersection-over-union ratio of the jth feature of the tth target; The corresponding segmentation threshold under the comprehensive maximum intersection-over-union ratio condition is as follows:
6. A radar proximity extended target detection and segmentation device based on multiple features, It is characterized in that It includes a sea surface segmentation module, which is used to perform coherent accumulation through high-resolution radar echoes on the sea surface, reduce the detection threshold of traditional CFAR, conduct preliminary inspection of the sea surface area, and obtain suspicious target areas and pure sea clutter areas; A feature extraction module is used to extract features of multiple dimensions from the pure sea clutter area using pulse pressure data, and calculate the target-clutter feature segmentation threshold of each feature; the multiple dimensions include time domain-fractal, frequency domain-Doppler entropy, energy domain-relative average amplitude and modulation domain-time-frequency ridge; A coarse segmentation module is used to extract features of multiple dimensions from the range image of the suspicious target area, form an eigenvalue matrix, and use the feature segmentation threshold to perform coarse segmentation of the target range channel; The weighted fusion module is used to cluster the coarse segmentation results, obtain the target and clutter difference and feature difference of each feature, and use different feature differences to weightedly fuse the multi-dimensional detection distance channels. The formula is as follows: Among them, R j is the detection and segmentation result of the jth feature, w j is the feature difference of the jth feature; The detection and segmentation module is used to select the IoU of various features under different segmentation thresholds for isolated targets or targets with the strongest energy, calculate the comprehensive IoU based on the feature difference, and select the segmentation threshold corresponding to the maximum comprehensive IoU to further process the adjacent targets to obtain the optimal solution for detection and segmentation. The detection and segmentation result formula for each target is as follows: in, is the final detection segmentation distance interval of the t-th target, R t is the multi-feature fusion interval of the t-th target, p 0 is the multi-feature fusion segmentation threshold.
7. A system for radar proximity extended target detection and segmentation based on multiple features, comprising a memory and a controller, wherein the memory stores a computer-readable program, and when the computer-readable program is called by the controller, the processor can execute the steps of the radar proximity extended target detection and segmentation method based on multiple features as described in any one of claims 1 to 5.
8. A computer-readable storage medium having a computer program stored thereon, Features: When the computer program is executed by a processor, the steps of the radar proximity extended target detection and segmentation method based on multiple features as described in any one of claims 1 to 5 are implemented.
Citation Information
Patent Citations
SRDF (Space-Range-Doppler Focus) moving object accumulation detection method for frequency diverse array radar
CN108693509A
Multi-Bernoulli tracking-before-detection method for multi-distance extended target
CN111580084A
Time-frequency constant false alarm high-frequency radar target detection method and system
CN113093119A
Radar target detection method and system based on sea surface scene perception, and storage medium
CN116482635A
Method of target feature extraction based on millimeter-wave radar echo
US20220155432A1
