Sea surface moving target detection method applying prior information

Abstract

The invention relates to a sea surface moving target detection method applying prior information, and belongs to the technical field of radar signal processing and target detection. The method comprises the following steps: 1) extracting and preprocessing test statistics; (2) a judgment threshold is solved through priori distribution of clutter test statistics; the method aims at solving the problem that a traditional single-frame target detection method is insufficient in sea surface small target and moving target detection performance under the high sea condition, and the sea surface target is detected through prior information provided by historical frames.

Description

Technical Field

[0001] The invention relates to a method for detecting sea surface moving targets using prior information, and belongs to the technical field of radar signal processing and target detection. Background Art

[0002] In the field of radar detection, the detection of small floating targets has always been a key and difficult point. The first difficulty is that the small radar scattering cross-section of small targets generally leads to weak radar echo energy. The second difficulty is that under low ground-grazing angle observation conditions, the non-uniform and non-stationary characteristics of sea clutter make it difficult to recognize the characteristics of sea clutter, and the strong echo generated by sea peaks will also confuse the resolution boundary between clutter and small targets.

[0003] In existing detection algorithms, features often require a long accumulation time to have considerable separability. However, in actual radar systems, short accumulation pulses are the mainstream pulse processing method. In order to solve the problem that the features and test statistics of short accumulation pulses are insufficient to distinguish between clutter and targets, existing methods often use multi-frame averaging or multi-frame forgetting factors to smooth or optimize the current frame test statistics using the information of historical frames. Although the existing multi-frame processing methods have achieved certain improvements, the limitation of the existing methods is that they do not quantify and distinguish the importance of the information contained in the current frame and the historical frame, which leads to the problem of insufficient use of historical frame information.

[0004] As an existing parameter estimation method, Bayesian estimation can integrate historical frame information and current frame information, make more reasonable estimates of distribution parameters, and has certain anti-interference capabilities. The Bayesian estimation method uses the discrete degree of historical frame information and current frame information to define the importance of current frame and historical frame information. The smaller the variance, the more important the information. The technology proposed in the present invention uses Bayesian estimation as a theoretical support. In the context of feature detection, it uses feature iteration to obtain the prior distribution of the features of the unit to be detected, and uses the clutter feature level estimated by the reference unit to distinguish the prior distribution of the target and clutter, and completes the detection work based on the distinction results.

[0005] In view of this feature, taking the SO-CFAR detection statistic as an example, a target detection method with stronger separability and sea spike suppression ability can be obtained, which improves the detection performance for small targets. However, in actual detection situations, the background of the unit to be detected is often disturbed, such as a moving ship entering or a stationary ship leaving. The iteration of the prior distribution often uses the historical frame information of the current unit. In order to ensure that the prior distribution can reflect the difference between the target and the clutter, the historical frame information should only contain clutter prior information or target prior information. When abnormal environmental data participates in the iteration of the prior distribution, the prior distribution will contain both target feature prior information and clutter feature prior information, which will make the prior distribution lose the ability to distinguish between targets and clutter. Therefore, the singularity of the prior distribution must be guaranteed so that the distribution can support the detection.

[0006] In view of the above phenomenon, the iterative strategy of the prior distribution must be adjusted to prevent sudden environmental mutation factors from affecting the iterative perception of the prior distribution, so as to adapt to the needs of actual target detection scenarios. Summary of the invention

[0007] The purpose of the present invention is to solve the shortcomings of the above-mentioned prior art and propose a method for detecting sea surface moving targets using prior information, aiming to improve the iterative method of prior information in the original Bayesian estimation method so that the test statistic after Bayesian estimation has better ability to distinguish clutter and targets.

[0008] The method for detecting sea surface moving targets using prior information of the present invention is special in that it comprises the following steps: Step 1: Extraction and preprocessing of test statistics The clutter echo time series of multiple clutter units within a period of time are collected, the mean-class CFAR test statistic series is extracted, and the Gaussian preprocessing is performed using the Box-Cox transformation as training information.

[0009] Step 2: Use the prior distribution of the clutter test statistic to obtain the decision threshold The initial parameters of the prior distribution are set using the Gaussianized test statistics of multiple clutter units, and the clutter test statistic sequences of multiple clutter units are iteratively estimated using the prior distribution iterative perception algorithm to obtain the Bayesian estimates of the prior distribution and test statistics that can adapt to environmental changes in real time. The Monte Carlo method and false alarm rate are used to obtain the appropriate prior distribution quantile as the decision threshold.

[0010] Step 3: Selective iteration of prior distribution during detection During the detection phase, the test statistics and Box-cox transformation are extracted from the echo time series of the unit to be detected and the reference unit. The prior distribution of the unit to be detected is obtained based on the input of the test statistic. The distribution position of the unbiased estimate of the test statistic of the unit to be detected in the prior distribution of the unit to be detected is observed. The number of sample points that deviate from the prior distribution of the unit to be detected within a period of history is counted. If the number of estimated value points that deviate from the prior distribution is too large, it is regarded as a sudden change in the environment of the unit to be detected at that time point. If no sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is calculated and compared with the threshold to obtain the detection result. If a sudden change in the environment occurs, the prior distribution parameters are fixed to prevent abnormal data from affecting the iteration of the prior distribution. During this time period, the probability density of the unbiased estimate of the unit to be detected in the prior distribution of the unit to be detected is used instead of the probability density of the Bayesian estimated mean of the reference unit, which is compared with the threshold to obtain the detection result.

[0011] Step 4: Use measured data to verify the proposed method, taking channel buoys and moving ships as the targets to be detected to verify the performance of the proposed method.

[0012] The proposed method has three advantages. First, it makes full use of the historical frame test statistics, which not only completes the correction of the prior distribution, but also completes the fusion of the historical frame test statistics and the current frame test statistics; second, based on the prior distribution containing historical frame information, a more discriminatory test statistic is extracted, thereby improving the performance of the detector; third, in view of the algorithm defects of using historical frames to calculate the prior distribution, the method proposed in this patent not only has excellent slow-moving small target detection capabilities, but also has the ability to detect moving targets that suddenly appear in the scene, which is more in line with the needs of actual detection scenarios.

[0013] Preferably, the specific steps of step 1 are: The radar collects sea clutter echo data of multiple range units in a certain direction in scanning mode, which is shown in the following form: (1); in, represents the complex echo sequence of the i-th clutter range unit, Represents the first frame echo sequence of the i-th clutter range unit.

[0014] Classic mean-type CFAR methods include but are not limited to CA-CFAR, GO-CFAR, and SO-CFAR. Since the method proposed in the present invention is sensitive to the same distribution of data, the reference unit needs to select the clutter unit as much as possible, so the test statistic selection strategy of the SO-CFAR method is selected.

[0015] First, the echo signal Perform square-law detection.

[0016] (2); In the formula, refer to The real part of the echo signal, refer to The imaginary part of the echo signal; Secondly, traverse in the fast time dimension, select the unit to be detected and the corresponding reference unit, and calculate the test statistic of the unit to be detected according to the following formula.

[0017] (3); In the formula represents the test statistic time series of the reference unit, It represents the echo time series of the unit to be detected after square law detection. , Represents the echo time series of the reference units on both sides of the unit to be detected after square-law detection.

[0018] Preferably, the specific steps of step 2 are: The application background of the method proposed in the present invention is a Gaussian background, and the data needs to be Gaussianized first to ensure that subsequent iterations of the prior distribution can be carried out stably.

[0019] The Box-cox transformation can be performed according to the following formula.

[0020] (4); In the formula, is the power transformation parameter, is the test statistic after power transformation.

[0021] This step can make the data easier to be accepted by the iterator without changing the size relationship of the test statistic, reduce the impact of extreme data on the iterator, and allow the iteration to proceed more smoothly.

[0022] The time series of the test statistic of the unit to be tested after Box-Cox transformation is recorded as , the time series of the test statistic of the reference unit is denoted by .

[0023] Step a: Take the time series of some clutter units , the total number of sequences is K, the sequence is divided into K / N sequence segments of length N, the mean of the sequence segments is counted, and K / N mean sequences are obtained. The mean and variance of these K / N mean sequences are estimated by maximum likelihood as the mean of the prior distribution and variance , complete the setting of prior information.

[0024] Step b: Delimit the unit to be detected and the reference unit based on the time series of another part of the clutter unit. and the unit to be tested , iterate using the iterative perceptual algorithm of the Bayesian estimator and obtain the Bayesian estimate of the reference unit test statistic , the specific steps are as follows Ⅰ~Ⅴ.

[0025] Ⅰ: Slide the time series of the unit, the sliding window step is step=1, the time series length of each unit is M, and the mean of the test statistic in the current window is and variance Make an unbiased estimate.

[0026] (6); Where N refers to the number of test statistic samples in the window.

[0027] Ⅱ: Secondly, using the mean and variance obtained by unbiased estimation, combined with the prior distribution, we can get the Bayesian estimate of the mean of the test statistic in the current window: , which is the mean of the posterior probability density function. Similarly, the variance of the posterior probability density function can also be calculated .

[0028] (7); Where N refers to the number of test statistic samples in the window, and are the mean and variance of the prior distribution, and are the mean and variance of the test statistic, respectively.

[0029] III: According to the Bayesian estimate obtained by the current sliding window, the mean and variance of the Bayesian estimate are sequentially estimated, and the mean and variance of the Bayesian estimate are updated.

[0030] (8); In the formula, k refers to the number of iterations, Refers to the Bayesian estimate obtained based on the sample at the current moment.

[0031] IV: According to the Bayesian estimate of the current sliding window, the mean and variance of the Bayesian estimate obtained by sequential estimation, and the variance of the current frame, the mean and variance of the prior distribution are corrected. The correction method is as follows.

[0032] (9); Where N refers to the number of test statistic samples in the window, and are the mean and variance of the Bayesian estimate, and are the mean and variance of the test statistic, respectively.

[0033] V: In the prior distribution iteration process, the Bayesian estimated mean sequence of the reference unit test statistic is obtained .

[0034] Step c: Compute the mean of the Bayesian estimates for the reference unit The probability density under the prior distribution density function of the current unit .

[0035] (5); Will The quantile threshold under the current unit prior distribution The corresponding probability density is compared, if it is greater than the quantile threshold The corresponding probability density indicates that the current unit is a clutter unit, otherwise it is a target unit. Suppose that we get The judgment result is obtained by continuously adjusting the quantile threshold. , let the number of samples judged as targets in the training set be , and output the threshold as the decision threshold for the detection step.

[0036] Preferably, the specific steps of step 3) are: In the detection phase, the echo time series of the unit to be detected and the reference unit are used to extract the test statistic and Box-Cox transformation. Based on the input of the test statistic, the prior distribution of the unit to be detected is obtained. If the detection background of the moving target is not considered, the Bayesian estimated mean of the reference unit can be used. and the prior distribution quantile threshold of the unit to be detected For comparison, if If the value is within the quantile threshold, it means that the current unit is a clutter unit, otherwise it is a target unit.

[0037] However, if the unit to be detected is a clutter unit and a ship passes by in the middle, the change in the test statistic caused by the ship will also participate in the iteration process of the prior distribution. The participation of abnormal samples in the iteration will inevitably change the prior distribution of the test statistic of the unit to be detected under the original single background, causing it to deviate from the prior distribution of the original background. And because the ship has a certain speed, it is often greater than the iteration speed of the prior distribution. When the prior distribution iterates to the prior distribution that conforms to the characteristics of the ship's test statistic, the ship may have left the unit to be detected. At this time, the prior distribution needs to spend some time to restore the prior distribution of the clutter background. In the process of this distribution iteration, a large number of missed detections and false alarms will occur. Therefore, it is necessary to avoid the test statistic that appears for a short time and does not match the original background from participating in the iteration of the prior distribution, so as to ensure that the detection method using the prior distribution can correctly reflect the background information of the current unit.

[0038] Therefore, some discriminant methods must be used to distinguish the test statistics that appear briefly and do not match the background, so as to avoid using this part of the data to iterate the prior distribution. Observe the distribution position of the unbiased estimate of the test statistic of the unit to be tested in the prior distribution of the unit to be tested. Set the length to The observation sliding window contains the total time from the current time to a certain period of time. The Bayesian estimate of the test statistic for the units to be tested. Statistical observation sliding window intrinsic prior distribution quantile threshold The number of Bayesian estimates other than . Assume a positive proportionality coefficient ,in .

[0039] If the number of estimated value points outside the threshold satisfies , it is considered that the unit to be detected has a sudden change in the environment at that time point. If no sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is calculated and compared with the threshold to obtain the detection result. If a sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is replaced by the probability density of the prior distribution of the unbiased estimate of the unit to be detected during this time period, and the detection result is obtained by comparing it with the threshold. During the period of sudden change in the environment, the iteration of the prior distribution needs to be stopped to prevent the contamination of the prior distribution by samples that do not match the environment during this period. For the specific processing flow, please refer to the attached Figure 2 .

[0040] The two parameters mentioned above use a length of The observation window and scale factor The target entering the unit to be detected and the target leaving the unit to be detected are identified and distinguished from the sea peak. Compared with the sea peak, the target echo sequence amplitude is more stable and the amplitude value is larger. Therefore, a length of The observation window and scale factor This difference is described and identified, thus distinguishing it from sea pinnacles. The longer it is, the higher the detection delay. When the time it takes for a moving ship to pass through the unit to be detected is less than the delay time Sometimes there will be missed detection, which will also make Sample points that do not match the environment participate in the iteration of the prior distribution. The longer the delay, the smaller the delay. However, if the delay is shorter than the duration of the sea peak, the sea peak will be regarded as an abnormal environmental change. It is not suitable in an environment with high sea conditions and frequent sea peaks. It is a measure of the stability of the features within the observation window. When it is 1, it means that all unbiased estimates of the test statistics in the window are shifting the prior distribution of the current unit. The larger it is, the more stringent the requirements on the stability of the features within the window.

[0041] Preferably, the specific steps of step 4) are: The measured data used include two nearly stationary channel buoys and a ship that entered the observation direction. In the intercepted data, the ship entered the 33rd distance unit at 16.5s after the start of the observation, and moved away along the observation direction, reaching the 100th distance unit at 65.5s, with a speed of about 8.3m / s along the observation direction. In addition, the remaining two floats were located at the 17th and 182nd distance units, respectively, and their positions did not change during the observation time.

[0042] In the detection phase, the unit to be detected is used and reference unit , The time echo sequence is extracted according to formula (2) to obtain the test statistic of the unit to be tested. and the test statistic for the reference unit , perform power transformation according to formula (3) to obtain the test statistic after power transformation of the unit to be tested and the test statistic after power transformation of the unit to be tested After step c, the probability density function of the unit to be detected is obtained , see formula (8) and the mean of the Bayesian estimates of the reference unit , and calculate As a new test statistic output, by With threshold For comparison, if This means that the unit to be detected is the target unit, otherwise it is clutter.

[0043] As can be seen from Appendix 1, the target detection method that only uses the Bayesian prior distribution iterative perception has a greatly improved detection probability for stationary floats compared to the original test statistic, but has no detection capability for moving targets. This is because when the test statistic with a higher amplitude generated by the moving target participates in the iteration of the prior distribution, the prior distribution of the current unit is between the target and the clutter, and the ability to use the prior distribution to distinguish between the target and the clutter is lost. The original test statistic has a responsiveness to moving targets with strong echoes due to the lack of any prior information support, but cannot achieve good detection results for stationary targets with weak echoes. This patent uses historical frame information to support the current frame information, so that slow weak echo small targets can also achieve excellent detection results in a low signal-to-clutter ratio environment. By using the perception of abnormal environmental changes, abnormal data is avoided from participating in the iteration of the prior distribution. On the basis of the target detection method of the Bayesian prior distribution iterative perception, the stability of the prior distribution is improved. It not only has the ability to detect moving targets, but also has excellent detection effects for slow small targets.

[0044] Compared with the prior art, the present invention provides a method for detecting sea surface moving targets using prior information, which has the following beneficial effects: (1) By making full use of the historical frame test statistics, not only the correction of the prior distribution is completed, but also the fusion of the historical frame test statistics and the current frame test statistics is completed.

[0045] (2) The algorithm defect of selectively using historical frames to calculate the prior distribution not only enables the method proposed in this patent to have excellent slow-moving small target detection capabilities, but also has the ability to detect moving targets that suddenly appear in the scene, which is more in line with the needs of actual detection scenarios.

[0046] (3) Under the existing data conditions, the average accuracy of the method proposed in this invention for detecting channel buoys is improved by more than 60% compared with the original method at an observation time of 0.064s, and the average detection accuracy of moving ships is improved by more than 30%. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The figure shows the change of the test statistic distribution histogram before and after the Box-cox transformation. In the figure, (a) is the test statistic distribution histogram before the Box-cox transformation; (b) is the test statistic distribution histogram after the Box-cox transformation. Figure 2 It is a detection flow chart of the method proposed by the present invention under the background of moving targets; Figure 3 The figure is a comparison diagram of the detection results of the method of the present invention and the existing method on a multi-target data set. In the figure, (a) is the traditional SO-CFAR target detection method; (b) is the detection method of the present invention. DETAILED DESCRIPTION

[0048] For better understanding and implementation, a specific embodiment is given below to describe in detail a method for detecting sea surface moving targets using prior information of the present invention, including the following steps: 1) Extraction and preprocessing of test statistics The radar collects sea clutter echo data of multiple range units in a certain direction in scanning mode, which can be expressed as follows: (1); in, represents the complex echo sequence of the i-th clutter range unit, Represents the first frame echo sequence of the i-th clutter range unit.

[0049] Classic mean-type CFAR methods include but are not limited to CA-CFAR, GO-CFAR, and SO-CFAR. Since the method proposed in the present invention is sensitive to the same distribution of data, the reference unit needs to select the clutter unit as much as possible, so the test statistic selection strategy of the SO-CFAR method is selected.

[0050] First, the echo signal Perform square-law detection.

[0051] (2); In the formula, refer to The real part of the echo signal, refer to Imaginary part of the echo signal.

[0052] Secondly, traverse in the fast time dimension, select the unit to be detected and the corresponding reference unit, and calculate the test statistic of the unit to be detected according to the following formula.

[0053] (3); In the formula represents the test statistic time series of the reference unit, It represents the echo time series of the unit to be detected after square law detection. , Represents the echo time series of the reference units on both sides of the unit to be detected after square-law detection.

[0054] 2) Using the prior distribution of the clutter test statistic to obtain the decision threshold The application background of the method proposed in this patent is Gaussian background, and the data needs to be Gaussianized first to ensure that the subsequent iteration of the prior distribution can be carried out stably.

[0055] The Box-cox transformation can be performed according to the following formula.

[0056] (4); In the formula, is the power transformation parameter, is the test statistic after power transformation.

[0057] This step can make the data easier to be accepted by the iterator without changing the size relationship of the test statistic, reduce the impact of extreme data on the iterator, and allow the iteration to proceed more smoothly. The results before and after the transformation are shown in the attached figure. Figure 1 As shown, Figure 1 It describes the distribution of the test statistic before and after the Box-cox transformation. After the Box-cox transformation, the distribution of the test statistic is more concentrated.

[0058] The time series of the test statistic of the unit to be tested after Box-Cox transformation is recorded as , the time series of the test statistic of the reference unit is denoted by .

[0059] Step a: Take the time series of some clutter units , the total number of sequences is K, the sequence is divided into K / N sequence segments of length N, the mean of the sequence segments is counted, and K / N mean sequences are obtained. The mean and variance of these K / N mean sequences are estimated by maximum likelihood as the mean of the prior distribution and variance , complete the setting of prior information.

[0060] Step b: Delimit the unit to be detected and the reference unit based on the time series of another part of the clutter unit. and the unit to be tested , iterate using the iterative perceptual algorithm of the Bayesian estimator and obtain the Bayesian estimate of the reference unit test statistic , the specific steps are as follows Ⅰ~Ⅴ.

[0061] Ⅰ: Slide the time series of the unit, the sliding window step is step=1, the time series length of each unit is M, and the mean of the test statistic in the current window is and variance Make an unbiased estimate.

[0062] (6); In the formula, and are the mean and variance of the prior distribution.

[0063] Ⅱ: Secondly, using the mean and variance obtained by unbiased estimation, combined with the prior distribution, we can get the Bayesian estimate of the mean of the test statistic in the current window: , which is the mean of the posterior probability density function. Similarly, the variance of the posterior probability density function can also be calculated .

[0064] (7); Where N refers to the number of test statistic samples in the window, and are the mean and variance of the prior distribution, and are the mean and variance of the test statistic, respectively.

[0065] III: According to the Bayesian estimate obtained by the current sliding window, the mean and variance of the Bayesian estimate are sequentially estimated, and the mean and variance of the Bayesian estimate are updated.

[0066] (8); In the formula, k refers to the number of iterations, Refers to the Bayesian estimate obtained based on the sample at the current moment.

[0067] IV: According to the Bayesian estimate of the current sliding window, the mean and variance of the Bayesian estimate obtained by sequential estimation, and the variance of the current frame, the mean and variance of the prior distribution are corrected. The correction method is as follows.

[0068] (9); Where N refers to the number of test statistic samples in the window, and are the mean and variance of the Bayesian estimate, and are the mean and variance of the test statistic, respectively.

[0069] V: In the prior distribution iteration process, the Bayesian estimated mean sequence of the reference unit test statistic is obtained .

[0070] Step c: Compute the mean of the Bayesian estimates for the reference unit The probability density under the prior distribution density function of the current unit .

[0071] (5); Will The quantile threshold under the current unit prior distribution The corresponding probability density is compared, if it is greater than the quantile threshold The corresponding probability density indicates that the current unit is a clutter unit, otherwise it is a target unit. Suppose that we get The judgment result is obtained by continuously adjusting the quantile threshold. , let the number of samples judged as targets in the training set be , and output the threshold as the decision threshold for the detection step.

[0072] 3) Selective iteration of prior distribution during detection In the detection phase, the echo time series of the unit to be detected and the reference unit are used to extract the test statistic and Box-Cox transformation. Based on the input of the test statistic, the prior distribution of the unit to be detected is obtained. If the detection background of the moving target is not considered, the Bayesian estimated mean of the reference unit can be used. and the prior distribution quantile threshold of the unit to be detected For comparison, if If the value is within the quantile threshold, it means that the current unit is a clutter unit, otherwise it is a target unit.

[0073] However, if the unit to be detected is a clutter unit and a ship passes by in the middle, the change in the test statistic caused by the ship will also participate in the iteration process of the prior distribution. The participation of abnormal samples in the iteration will inevitably change the prior distribution of the test statistic of the unit to be detected under the original single background, causing it to deviate from the prior distribution of the original background. And because the ship has a certain speed, it is often greater than the iteration speed of the prior distribution. When the prior distribution iterates to the prior distribution that conforms to the characteristics of the ship's test statistic, the ship may have left the unit to be detected. At this time, the prior distribution needs to spend some time to restore the prior distribution of the clutter background. In the process of this distribution iteration, a large number of missed detections and false alarms will occur. Therefore, it is necessary to avoid the test statistic that appears for a short time and does not match the original background from participating in the iteration of the prior distribution, so as to ensure that the detection method using the prior distribution can correctly reflect the background information of the current unit.

[0074] Therefore, some discriminant methods must be used to distinguish the test statistics that appear briefly and do not match the background, so as to avoid using this part of the data to iterate the prior distribution. Observe the distribution position of the unbiased estimate of the test statistic of the unit to be tested in the prior distribution of the unit to be tested. Set the length to The observation sliding window contains the total time from the current time to a certain period of time. The Bayesian estimate of the test statistic for the units to be tested. Statistical observation sliding window intrinsic prior distribution quantile threshold The number of Bayesian estimates other than . Assume a positive proportionality coefficient ,in .

[0075] If the number of estimated value points outside the threshold satisfies , it is considered that the unit to be detected has a sudden change in the environment at that time point. If no sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is calculated and compared with the threshold to obtain the detection result. If a sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is replaced by the probability density of the prior distribution of the unbiased estimate of the unit to be detected during this time period, and the detection result is obtained by comparing it with the threshold. During the period of sudden change in the environment, the iteration of the prior distribution needs to be stopped to prevent the contamination of the prior distribution by samples that do not match the environment during this period. For the specific processing flow, please refer to the attached Figure 2 .

[0076] The two parameters mentioned above use a length of The observation window and scale factor The target entering the unit to be detected and the target leaving the unit to be detected are identified and distinguished from the sea peak. Compared with the sea peak, the target echo sequence amplitude is more stable and the amplitude value is larger. Therefore, a length of The observation window and scale factor This difference is described and identified, thus distinguishing it from sea pinnacles. The longer it is, the higher the detection delay. When the time it takes for a moving ship to pass through the unit to be detected is less than the delay time Sometimes there will be missed detection, which will also make Sample points that do not match the environment participate in the iteration of the prior distribution. The longer the delay, the smaller the delay. However, if the delay is shorter than the duration of the sea peak, the sea peak will be regarded as an abnormal environmental change. It is not suitable in an environment with high sea conditions and frequent sea peaks. It is a measure of the stability of the features within the observation window. When it is 1, it means that all unbiased estimates of the test statistics in the window are shifting the prior distribution of the current unit. The larger it is, the more stringent the requirements on the stability of the features within the window.

[0077] 4) The proposed method is verified using measured data.

[0078] The measured data used include two nearly stationary channel buoys and a ship that entered the observation direction. In the intercepted data, the ship entered the 33rd distance unit at 16.5s after the start of the observation, and moved away along the observation direction, reaching the 100th distance unit at 65.5s, with a speed of about 8.3m / s along the observation direction. In addition, the remaining two floats were located at the 17th and 182nd distance units, respectively, and their positions did not change during the observation time.

[0079] In the detection phase, the unit to be detected is used and reference unit , The time echo sequence is extracted according to formula (2) to obtain the test statistic of the unit to be tested. and the test statistic for the reference unit , perform power transformation according to formula (3) to obtain the test statistic after power transformation of the unit to be tested and the test statistic after power transformation of the unit to be tested After step c, the probability density function of the unit to be detected is obtained , see formula (8) and the mean of the Bayesian estimates of the reference unit , and calculate As a new test statistic output, by With threshold For comparison, if This means that the unit to be detected is the target unit, otherwise it is clutter.

[0080] As can be seen from Appendix 1, the target detection method that only uses the Bayesian prior distribution iterative perception has a very large improvement in the detection probability of stationary floats compared to the original test statistic, but has no detection ability for moving targets. This is because when the test statistic with a higher amplitude generated by the moving target participates in the iteration of the prior distribution, the prior distribution of the current unit is between the target and the clutter, and the ability to use the prior distribution to distinguish between the target and the clutter is lost. The original test statistic has a responsiveness to the moving target with strong echo due to the lack of any prior information support, but it cannot achieve a good detection effect for the stationary target with weak echo. The present invention uses historical frame information to support the current frame information, so that the slow weak echo small target can also achieve excellent detection results in a low signal-to-clutter ratio environment. By using the perception of abnormal environmental changes, the abnormal data is avoided from participating in the iteration of the prior distribution. On the basis of the target detection method of the Bayesian prior distribution iterative perception, the stability of the prior distribution is improved, and it not only has the ability to detect moving targets, but also has excellent detection effects for slow small targets. The actual test results of the original method and the method proposed by the present invention are shown in the attached Figure 3 and as shown in Table 1. Figure 3 It can be clearly observed that the detection results of both ships and floats are more obvious with the proposed method than with the original test statistic detection method.

[0081] Table 1 The original test statistics and the detection probability of the patented method on the measured data set ; The detection probability of the present invention for channel buoys in the measured data set is above 60%, which is significantly improved compared with the original method. The detection probability of ships is increased from 68.14% of the original method to 99.17%.

Claims

1. A method for detecting sea surface moving targets using prior information, characterized in that The following steps are involved: Step 1: Extraction and preprocessing of test statistics Collect clutter echo time series of multiple clutter units within a period of time, extract the mean-class CFAR test statistic series, and use Box-Cox transformation to perform Gaussian preprocessing as training information; Step 2: Use the prior distribution of the clutter test statistic to obtain the decision threshold The initial parameters of the prior distribution are set by using the Gaussian test statistics of multiple clutter units, and the clutter test statistic sequences of multiple clutter units are iteratively estimated using the prior distribution iterative perception algorithm to obtain the Bayesian estimation of the prior distribution and test statistics that can adapt to environmental changes in real time. The Monte Carlo method and false alarm rate are used to obtain the appropriate prior distribution quantile as the decision threshold. Step 3: Selective iteration of prior distribution during detection In the detection stage, the test statistics and Box-cox transformation are extracted from the echo time series of the unit to be detected and the reference unit. According to the input of the test statistics, the prior distribution of the unit to be detected is obtained, and the distribution position of the unbiased estimate of the test statistic of the unit to be detected in the prior distribution of the unit to be detected is observed. The number of sample points that deviate from the prior distribution of the unit to be detected within a period of history is counted. If the number of estimated value points that deviate from the prior distribution is too large, it is regarded as a sudden change in the environment of the unit to be detected at this time point. If no sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is calculated and compared with the threshold to obtain the detection result. If a sudden change in the environment occurs, the prior distribution parameters are fixed to prevent abnormal data from affecting the iteration of the prior distribution. In this time period, the probability density of the unbiased estimate of the unit to be detected in the prior distribution of the unit to be detected is used instead of the probability density of the Bayesian estimated mean of the reference unit, and compared with the threshold to obtain the detection result. Step 4: Use measured data to verify the proposed method.

2. A method for detecting sea surface moving targets using prior information according to claim 1, characterized in that The specific steps of step 1 are: the radar collects sea clutter echo data of multiple range units in a certain direction in scanning mode, which is expressed as follows: （1）； in, represents the complex echo sequence of the i-th clutter range unit, Represents the first frame echo sequence of the i-th clutter range unit; The strategy for selecting the test statistic for the SO-CFAR method is: First, the echo signal Perform square-law detection: （2）； In the formula, refer to The real part of the echo signal, refer to The imaginary part of the echo signal; Secondly, traverse in the fast time dimension, select the unit to be detected and the corresponding reference unit, and calculate the test statistic of the unit to be detected according to the following formula: （3）； In the formula represents the test statistic time series of the reference unit, It represents the echo time series of the unit to be detected after square law detection. , Represents the echo time series of the reference units on both sides of the unit to be detected after square-law detection.

3. A method for detecting sea surface moving targets using prior information according to claim 1, characterized in that The specific steps of step 2 are: Prioritize Gaussianization of the data to ensure that subsequent iterations of the prior distribution can proceed stably; The Box-cox transformation is performed as follows: （4）； In the formula, is the power transformation parameter, is the test statistic after power transformation; The time series of the test statistic of the unit to be tested after Box-Cox transformation is recorded as , the time series of the test statistic of the reference unit is denoted by ; Step a: Take the time series of some clutter units , the total number of sequences is K, the sequence is divided into K / N sequence segments of length N, the mean of the sequence segments is counted, and K / N mean sequences are obtained. The mean and variance of these K / N mean sequences are estimated by maximum likelihood as the mean of the prior distribution and variance , complete the setting of prior information; Step b: Delimit the detection unit and the reference unit based on the time series of another part of the clutter unit. and the unit to be tested , iterate using the iterative perceptual algorithm of the Bayesian estimator and obtain the Bayesian estimate of the reference unit test statistic , Step c: Compute the mean of the Bayesian estimates for the reference unit The probability density under the prior distribution density function of the current unit : （5）； In the formula, and are the mean and variance of the prior distribution; Will The quantile threshold under the current unit prior distribution The corresponding probability density is compared, if it is greater than the quantile threshold The corresponding probability density indicates that the current unit is a clutter unit, otherwise it is a target unit. Suppose that we get The judgment result is obtained by continuously adjusting the quantile threshold. , let the number of samples judged as targets in the training set be , and output the threshold as the decision threshold for the detection step.

4. A method for detecting sea surface moving targets using prior information according to claim 1, characterized in that The specific steps of step 3 are: In the detection phase, the echo time series of the unit to be detected and the reference unit are used to extract the test statistic and Box-Cox transformation. Based on the input of the test statistic, the prior distribution of the unit to be detected is obtained. If the detection background of the moving target is not considered, the Bayesian estimated mean of the reference unit is used. and the prior distribution quantile threshold of the unit to be detected For comparison, if If the current unit is within the quantile threshold, it means that the current unit is a clutter unit, otherwise it is a target unit; Distinguish the test statistics that appear for a short time and do not match the background, avoid using this part of the data to iterate the prior distribution, observe the distribution position of the unbiased estimate of the test statistic of the unit to be tested in the prior distribution of the unit to be tested, and set the length to The observation sliding window contains the total time from the current time to a certain period of time. The Bayesian estimate of the test statistic for the units to be tested, the quantile threshold of the inherent prior distribution of the statistical observation sliding window The number of Bayesian estimates other than , assume a positive proportionality coefficient ,in ; If the number of estimated value points outside the threshold satisfies , it is considered that the unit to be detected has a sudden change in the environment at that time point; if no sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is calculated and compared with the threshold to obtain the detection result; if a sudden change in the environment occurs, the probability density of the Bayesian estimated mean of the reference unit is replaced by the probability density of the prior distribution of the unbiased estimate of the unit to be detected during this time period, and compared with the threshold to obtain the detection result; during the period of sudden change in the environment, the iteration of the prior distribution needs to be stopped to prevent the contamination of the prior distribution by samples that do not match the environment during this period.

5. A method for detecting sea surface moving targets using prior information according to claim 3, characterized in that The specific steps of step b are as follows: Ⅰ: Slide the time series of the unit, the sliding window step is step=1, the time series length of each unit is M, and the mean of the test statistic in the current window is and variance Make an unbiased estimate: （6）； Where N refers to the number of test statistic samples in the window; Ⅱ: Using the mean and variance obtained by unbiased estimation, combined with the prior distribution, we can get the Bayesian estimate of the mean of the test statistic in the current window. , which is the mean of the posterior probability density function. Similarly, the variance of the posterior probability density function is It can also be calculated: （7）； Where N refers to the number of test statistic samples in the window, and are the mean and variance of the prior distribution, and are the mean and variance of the test statistic, respectively; III: Based on the Bayesian estimate obtained by the current sliding window, the mean and variance of the Bayesian estimate are sequentially estimated to update the Bayesian estimate The mean and variance of : （8）； In the formula, k refers to the number of iterations, Refers to the Bayesian estimate obtained based on the sample at the current moment; IV: According to the Bayesian estimate of the current sliding window, the mean and variance of the Bayesian estimate obtained by sequential estimation, and the variance of the current frame, the mean and variance of the prior distribution are corrected. The correction method is as follows: （9）； Where N refers to the number of test statistic samples in the window, and are the mean and variance of the Bayesian estimate, and are the mean and variance of the test statistic, respectively; V: In the prior distribution iteration process, the Bayesian estimated mean sequence of the reference unit test statistic is obtained .

Images