Underwater moving target detection method based on echo space-time-frequency joint deep learning
Through the underwater motion target detection method based on echo space-time frequency combined with deep learning, the problems of low signal-to-noise ratio and serious false alarm in the existing technology are solved, and the high accuracy and automated detection of underwater motion targets are achieved, and the The confidence in the detection is achieved.
Patent Information
- Application Number
- CN202411491660.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-02-14
AI Technical Summary
In the detection of underwater motion targets, there are problems such as low signal-to-noise ratio, unstable Doppler frequency shift calculation and serious false alarm interference in the detection of underwater motion targets, and it is difficult to effectively utilize the consistency of motion characteristics of the moving target in the time-space-frequency joint space. .
Using the underwater motion object detection method based on echo space-time frequency combined with deep learning, the active sonar matrix data is preprocessed, and the space-time and time-frequency processing is carried out, and the deep learning detection model is constructed, combined with space-time frequency verification is achieved to achieve motion Automatic detection and false alarm suppression of targets.
It improves the accuracy and automation of underwater motion target detection, enhances the confidence in the judgment of motion targets, and effectively suppresses false alarm interference.
Smart Images
Figure CN119439169A_ABST
Abstract
Description
Technical field:
[0001] The present invention belongs to the field of sonar technology, and in particular relates to an underwater moving target detection method based on echo time-space-frequency joint deep learning. Background technology:
[0002] Active sonar detects, locates, and identifies underwater targets by emitting sound waves and then processing and analyzing the echoes. Underwater target identification has always been a difficult problem in the field of underwater acoustics. The target's movement speed can be used to judge underwater moving targets, stationary targets, or fast-moving targets and slow-moving targets. It plays an important role in judging the type of underwater targets and is also an important basis for the rapid screening of active sonar contact targets. When the underwater target moves relative to the platform, the received echo will have a Doppler shift. By calculating the Doppler shift of the echo, the radial movement speed of the target can be calculated. However, due to the influence of strong reverberation, multiple interferences, and time-space variable ocean channel propagation effects, the received echo is often unstable and the signal-to-noise ratio is low, resulting in large errors in the extraction of motion features. In addition, the current motion feature extraction does not fully utilize the consistency of the motion features of the moving target in the joint space of time, space, and frequency. Under low signal-to-noise ratio conditions, it is susceptible to false alarm interference caused by random disturbances and relies on human-computer interaction judgment, which greatly affects the autonomous detection performance of the moving target.
[0003] In recent years, artificial intelligence technologies represented by deep learning have developed rapidly, especially in the field of vision, but there is currently no record of the application of deep learning in moving target detection. Summary of the invention:
[0004] The technical problem to be solved by the present invention is to provide an underwater moving target detection method based on echo time-space-frequency joint deep learning. The method uses a deep learning method to detect moving targets, combines the consistency of the moving target's motion characteristics in time-space-frequency, and improves the confidence in the judgment of the moving target.
[0005] The technical solution of the present invention is to provide an underwater moving target detection method based on echo time-space-frequency joint deep learning, comprising the following steps:
[0006] Step 1: Select active sonar array data containing moving targets for preprocessing and beamforming to obtain beam domain data.
[0007] Step 2: Perform space-time two-dimensional processing on the full-beam data to obtain the Doppler characteristic spectrum S(l,k) in the space-time domain, l = 1:M 1 ,k=1:N 1 , where l represents the beam number, M 1 is the total number of beams, k represents the number of frequency points, N 1is the total number of frequency points, and the corresponding frequency range is [f L ,f H ] Perform moving target annotation on S(l,k) and construct a spatial-temporal Doppler signature spectrum training dataset.
[0008] Step 3: Perform two-dimensional time-frequency processing on the single-beam data to obtain the Doppler characteristic spectrum F(n,k) in the time-frequency domain, n=1:M 2 ,k=1:N 2 , where n represents the number of time sampling points, M 2 is the total number of time sampling points, k represents the number of frequency points, N 2 is the total number of frequency points, and the corresponding frequency range is [f L ,f H ] Perform moving target labeling on F(n,k) and construct a time-frequency domain Doppler characteristic spectrum training data set.
[0009] Step 4: Construct a spatial-temporal moving target detection model M based on the spatial-temporal Doppler characteristic spectrum characteristics S The model can be a commonly used deep learning detection model such as FastRCNN, FasterRCNN, YOLO series (YOLOv8, etc.). The model is trained using the spatial-temporal training dataset until convergence.
[0010] Step 5: Based on the characteristics of the Doppler characteristic spectrum in the time-frequency domain, construct a moving target detection model M in the time-frequency domain F The model can be a commonly used deep learning detection model such as FastRCNN, FasterRCNN, YOLO series (YOLOv8, etc.). The model is trained using the time-frequency domain training data set until convergence.
[0011] Step 6: For the predicted active sonar reception echo, first generate full-beam data, perform segmented two-dimensional space-time processing on the full-beam data, and obtain L space-time domain characteristic spectra S i (l,k),i=1:L, perform segmented time-frequency processing on each single beam, and generate M 1 ×K time-frequency domain feature spectra F ij (n,k),i=1:M 1 ,j=1:K, where M 1 Represents the total number of beams, and K is the number of time-frequency spectrum features generated by each single beam. S , time-frequency domain moving target detection model M F Perform moving target detection and select the detection box with a confidence threshold greater than 0.7 for output. Model M S Output a total of M detection boxes Model M F Output a total of N detection boxes
[0012] Step 7: Calculate the moving target elements of the detection frame obtained in step 6.
[0013] For the i-th model based on M S The output detection box (x i ,y i ,w i ,h i ), assuming it comes from the nth i Spatial-temporal signature spectrum generated by segment data The corresponding beam number of the moving target is bm i , distance d i and radial velocity v i The solution results are as follows:
[0014] bm i =x i +w i / 2,
[0015] Among them, N 0 is the number of sampling points for each data segment, fs is the sampling rate, c is the speed of sound, f d is the reference frequency when there is no moving target, f 0 is the transmitting signal frequency.
[0016] For the j-th model M F The output detection box (x j ,y j ,w j ,h j ), assuming that it is the time-frequency domain characteristic spectrum generated by the nj-th segment data of the mj-th single beam The corresponding beam number of the moving target is bm j , distance d j and radial velocity v j The solution results are as follows:
[0017]
[0018] Among them, M 0 The number of sampling points for each segment of data.
[0019] Step 8: Perform time-space frequency verification on the moving targets obtained in step 7. For K moving targets with the same beam number and the same distance, their radial velocity sequence is v = [v 1 ,v 2 ,L,v K ], if var(v)>δ 1, it is considered that the target has inconsistent motion characteristics in the spatiotemporal and time-frequency domains, which is caused by random disturbance or model misdetection, and is deleted; if var(v)<δ 1 And min(v)>δ 2 , they are merged into one moving target.
[0020] Compared with the prior art, the present invention has the following advantages:
[0021] (1) By utilizing the differential characteristics of moving target echoes in the time-space domain, time-frequency domain and reverberation, the moving target is automatically detected based on the deep learning detection model, which improves the accuracy and automation of moving target detection.
[0022] (2) By utilizing the space-time-frequency consistency of the moving target, the false alarm caused by random disturbances or model misdetection can be effectively suppressed through space-time-frequency verification, thereby improving the confidence in the detection of the moving target. Description of the drawings:
[0023] Figure 1 It is a work flow chart of the present invention.
[0024] Figure 2 This is a schematic diagram of underwater moving target detection based on echo time-space-frequency joint deep learning.
[0025] Figure 3 is the Doppler characteristic spectrum, where Figure 3 (a) and (b) are the spatial-temporal Doppler characteristic spectra with and without moving targets, respectively. Figure 3 (c) and (d) are the time-frequency domain Doppler characteristic spectra with and without moving targets, respectively.
[0026] Figure 4 This is an example of using time-space frequency verification to eliminate false alarms in a period of sonar reception data.
[0027] Figure 5 This is an example of a moving target detected based on deep learning for a period of sonar reception data. Specific implementation method:
[0028] The present invention will be further described below with reference to the accompanying drawings:
[0029] A method for underwater moving target detection based on echo time-space-frequency joint deep learning, such as Figure 1 The workflow diagram shown, and Figure 2 The schematic diagram of underwater moving target detection based on echo time-space-frequency joint deep learning shown in the figure specifically includes the following steps:
[0030] Step 1: Select active sonar array data containing moving targets for preprocessing and beamforming to obtain beam domain data.
[0031] Step 2: Perform space-time two-dimensional processing on the full-beam data to obtain the Doppler characteristic spectrum S(l,k) in the space-time domain, l = 1:M 1 ,k=1:N 1 , where l represents the beam number, M 1 is the total number of beams, k represents the number of frequency points, N 1 is the total number of frequency points, and the corresponding frequency range is [f L ,f H ] Perform moving target annotation on S(l,k) and construct a spatial-temporal Doppler signature spectrum training dataset.
[0032] Step 3: Perform two-dimensional time-frequency processing on the single-beam data to obtain the Doppler characteristic spectrum F(n,k) in the time-frequency domain, n=1:M 2 ,k=1:N 2 , where n represents the number of time sampling points, M 2 is the total number of time sampling points, k represents the number of frequency points, N 2 is the total number of frequency points, and the corresponding frequency range is [f L ,f H ] Perform moving target labeling on F(n,k) and construct a time-frequency domain Doppler characteristic spectrum training data set.
[0033] Figure 3 (a), (b) and (c), (d) respectively give the space-time domain and time-frequency domain Doppler characteristic spectra with and without moving targets. It can be seen that when there is no moving target, the Doppler characteristic spectrum is a straight line caused by reverberation; when there is a moving target, the moving target feature points and feature segments will appear on the Doppler characteristic spectrum at a position deviating from the reverberation band. The space-time domain Doppler characteristic spectrum reflects the difference characteristics of the target and reverberation at different directions in the same time period, and the time-frequency domain Doppler characteristic spectrum reflects the difference characteristics of the target and reverberation at different times in the same direction. The difference characteristics of the moving target and reverberation in the space-time domain and time-frequency domain Doppler characteristic spectra can be used for moving target detection.
[0034] Step 4: Construct a spatial-temporal moving target detection model M based on the spatial-temporal Doppler characteristic spectrum characteristics S The model can be a commonly used deep learning detection model such as FastRCNN, FasterRCNN, YOLO series (YOLOv8, etc.). The model is trained using the spatial-temporal training dataset until convergence.
[0035] Step 5: Based on the characteristics of the Doppler characteristic spectrum in the time-frequency domain, construct a moving target detection model M in the time-frequency domain FThe model can be a commonly used deep learning detection model such as FastRCNN, FasterRCNN, YOLO series (YOLOv8, etc.). The model is trained using the time-frequency domain training data set until convergence.
[0036] Step 6: For the predicted active sonar reception echo, first generate full-beam data, perform segmented two-dimensional space-time processing on the full-beam data, and obtain L space-time domain characteristic spectra S i (l,k),i=1:L, perform segmented time-frequency processing on each single beam, and generate M 1 ×K time-frequency domain feature spectra F ij (n,k),i=1:M 1 ,j=1:K, where M 1 Represents the total number of beams, and K is the number of time-frequency spectrum features generated by each single beam. S , time-frequency domain moving target detection model M F Perform moving target detection and select the detection box with a confidence threshold greater than 0.7 for output. Model M S Output a total of M detection boxes Model M F Output a total of N detection boxes
[0037] Step 7: Calculate the moving target elements of the detection frame obtained in step 6.
[0038] For the i-th model based on M S The output detection box (x i ,y i ,w i ,h i ), assuming it comes from the nth i Spatial-temporal signature spectrum generated by segment data The corresponding beam number of the moving target is bm i , distance d i and radial velocity v i The solution results are as follows:
[0039]
[0040] Among them, N 0 is the number of sampling points for each data segment, fs is the sampling rate, c is the speed of sound, f d is the reference frequency when there is no moving target, f 0 is the transmitting signal frequency.
[0041] For the j-th model M F The output detection box (x j ,y j ,wj ,h j ), assuming it comes from the mth j The nth single beam j Time-frequency domain feature spectrum generated by segment data The corresponding beam number of the moving target is bm j , distance d j and radial velocity v j The solution results are as follows:
[0042]
[0043] Among them, M 0 The number of sampling points for each segment of data.
[0044] Step 8: Perform time-space frequency verification on the moving targets obtained in step 7. For K moving targets from the same beam number and the same distance, their radial velocity sequence is v = [v 1 ,v 2 ,L,v K ], if var(v)>δ 1 , it is considered that the target has inconsistent motion characteristics in the spatiotemporal and time-frequency domains, which is caused by random disturbance or model misdetection, and is deleted; if var(v)<δ 1 And min(v)>δ 2 , they are merged into one moving target.
[0045] Figure 4 An example of a false alarm detection of a period of sonar receiving data is given. The radial velocity difference of the target in the time-space domain and the time-frequency domain is too large. It is considered to be caused by random noise interference or model misdetection and is deleted.
[0046] Figure 5 It is the detection result of a period of sonar receiving data. After the time-space-frequency verification, the motion characteristics of the moving target in the time-space domain and the time-frequency domain are consistent, and it is determined to be a moving target.
[0047] from Figure 4 , Figure 5 It can be seen that there is a large amount of reverberation interference in the data received by active sonar. The present invention makes full use of the differences in the moving targets in the time-space, time-frequency domain and reverberation, and performs moving target detection based on a deep learning detection model. At the same time, the consistency of the moving targets' motion characteristics in time-space-frequency is utilized, which can effectively suppress the interference caused by random disturbances, reduce the false alarms caused by model misdetection, and improve the confidence in the detection of moving targets.
[0048] It can be seen that the present invention processes the active sonar receiving data in the time-frequency domain and the time-space domain respectively, utilizes the difference characteristics of the moving target and reverberation in the time-frequency domain and the time-space domain, and performs automatic detection of the moving target on the time-frequency domain and the time-space domain Doppler characteristic spectrum based on the deep learning detection model, and then uses the consistency of the moving target's motion characteristics in the time-space-frequency joint domain to perform time-space-frequency verification, suppresses false alarms caused by random disturbances or model misdetection, and improves the accuracy, confidence and automation of moving target detection. The method is simple and flexible, and has high engineering application value.
[0049] The above is only an explanation of the preferred embodiments of the present invention, which should not be construed as a limitation on the claims. Any equivalent process changes made using the present invention specification are included in the patent protection scope of the present invention.
Claims
1. A method for underwater moving target detection based on echo time-space-frequency joint deep learning, characterized by: The following steps are included: Step 1: Select active sonar array data containing moving targets for preprocessing, and perform beam forming to obtain beam domain data; Step 2: Perform space-time two-dimensional processing on the full-beam data to obtain the Doppler characteristic spectrum S(l,k) in the space-time domain, l = 1:M1, k = 1:N1, where l represents the beam number, M1 is the total number of beams, k represents the number of frequency points, N1 is the total number of frequency points, and the corresponding frequency range is [f L ,f H ]; label the moving target of S(l,k) and construct the space-time domain Doppler characteristic spectrum training data set; Step 3: Perform two-dimensional time-frequency processing on the single-beam data to obtain the Doppler characteristic spectrum F(n,k) in the time-frequency domain, n = 1:M2, k = 1:N2, where n represents the number of time sampling points, M2 is the total number of time sampling points, k represents the number of frequency points, N2 is the total number of frequency points, and the corresponding frequency range is [f L ,f H ], mark the moving target of F(n,k) and construct the Doppler characteristic spectrum training data set in the time-frequency domain; Step 4: Construct a spatial-temporal moving target detection model M based on the spatial-temporal Doppler characteristic spectrum characteristics S , train the model using the spatial and temporal training dataset until convergence; Step 5: Based on the characteristics of the Doppler characteristic spectrum in the time-frequency domain, construct a moving target detection model M in the time-frequency domain F , train the model using the time-frequency domain training data set until convergence; Step 6: For the predicted active sonar reception echo, first generate full-beam data, perform segmented two-dimensional space-time processing on the full-beam data, and obtain L space-time domain characteristic spectra S i (l,k),i=1:L, perform segmented time-frequency processing on each single beam, and generate a total of M1×K time-frequency domain feature spectra F ij (n, k), i = 1: M1, j = 1: K, where M1 represents the total number of beams, and K is the number of time-frequency feature spectra generated by each single beam; then the spatial and temporal domain moving target detection model M S , Time-frequency domain moving target detection model M F Perform moving target detection and select the detection box with a confidence threshold greater than 0.7 for output. Model M S Output a total of M detection boxes Model M F Output a total of N detection boxes Step 7: Calculate the moving target elements of the detection frame obtained in step 6; For the i-th model based on M S The output detection box (x i ,y i ,w i ,h i ), assuming it comes from the nth i Spatial-temporal signature spectrum generated by segment data The corresponding beam number of the moving target is bm i , distance d i and radial velocity v i The solution results are as follows: bm i =x i +w i / 2, Among them, N0 is the number of sampling points in each data segment, fs is the sampling rate, c is the speed of sound, and f d is the reference frequency when there is no moving target, f0 is the transmission signal frequency; For the j-th model M F The output detection box (x j ,y j ,w j ,h j ), assuming it comes from the mth j The nth single beam j Time-frequency domain feature spectrum generated by segment data The corresponding beam number of the moving target is bm j , distance d j and radial velocity v j The solution results are as follows: bm j =m j , Among them, M0 is the number of sampling points for each segment of data; Step 8: Perform time-space frequency verification on the moving targets obtained in step 7. For K moving targets with the same beam number and the same distance, the radial velocity sequence is v = [v1, v2, L, v K ], if var(v)>δ1, it is considered that the target has inconsistent motion characteristics in the spatiotemporal domain and the time-frequency domain and is deleted; if var(v)<δ1 and min(v)>δ2, they are merged into one moving target.
2. The underwater moving target detection method based on echo time-space-frequency joint deep learning according to claim 1 is characterized by: In step 4, the model can be any deep learning detection model among FastRCNN, FasterRCNN, and YOLO.
3. The underwater moving target detection method based on echo time-space-frequency joint deep learning according to claim 1 is characterized by: In step 5, the model can be any deep learning detection model among FastRCNN, FasterRCNN, and YOLO.