Submit Manuscript
Peer Review
Editor Work2023 Vol. 12, No. 5
Previous Issue
2023, 12(5): 923-970.
Abstract
10003KB
As one of the core components of Advanced Driver Assistance Systems (ADAS), automotive millimeter-wave radar has become the focus of scholars and manufacturers at home and abroad because it has the advantages of all-day and all-weather operation, miniaturization, high integration, and key sensing capabilities. The core performance indicators of the automotive millimeter-wave radar are distance, speed, angular resolution, and field of view. Accuracy, cost, real-time and detection performance, and volume are the key issues to be considered. The increasing performance requirements pose several challenges for the signal processing of millimeter-wave radar systems. Radar signal processing technology is crucial for improving radar performance to meet more stringent requirements. Obtaining dense radar point clouds, generating accurate radar imaging results, and mitigating mutual interference among multiple radar systems are the key points and the foundation for subsequent tracking, recognition, and other applications. Therefore, this paper discusses the practical application of the automotive millimeter-wave radar system based on the key technologies of signal processing, summarizes relevant research results, and mainly discusses the topics of point cloud imaging processing, synthetic aperture radar imaging processing, and interference suppression. Finally, herein, we summarize the research status at home and abroad. Moreover, future development trends for automotive millimeter-wave radar systems are forecast with the hope of enlightening readers in related fields.
As one of the core components of Advanced Driver Assistance Systems (ADAS), automotive millimeter-wave radar has become the focus of scholars and manufacturers at home and abroad because it has the advantages of all-day and all-weather operation, miniaturization, high integration, and key sensing capabilities. The core performance indicators of the automotive millimeter-wave radar are distance, speed, angular resolution, and field of view. Accuracy, cost, real-time and detection performance, and volume are the key issues to be considered. The increasing performance requirements pose several challenges for the signal processing of millimeter-wave radar systems. Radar signal processing technology is crucial for improving radar performance to meet more stringent requirements. Obtaining dense radar point clouds, generating accurate radar imaging results, and mitigating mutual interference among multiple radar systems are the key points and the foundation for subsequent tracking, recognition, and other applications. Therefore, this paper discusses the practical application of the automotive millimeter-wave radar system based on the key technologies of signal processing, summarizes relevant research results, and mainly discusses the topics of point cloud imaging processing, synthetic aperture radar imaging processing, and interference suppression. Finally, herein, we summarize the research status at home and abroad. Moreover, future development trends for automotive millimeter-wave radar systems are forecast with the hope of enlightening readers in related fields.
2023, 12(5): 971-985.
Single snapshot forward-looking imaging technology with high performance and resolution is crucial for enabling the development of automotive radars. However, range migration issues can limit the implementation of coherent integration methods, and improving system resolution is generally difficult due to hardware parameter limitations. Based on the Time-Division Multiplexing Multiple-Input-Multiple-Output (TDM-MIMO) forward-looking imaging systems of automotive millimeter wave radar, this paper proposes Doppler domain compensation and point-to-point echo correction measures for achieving multidomain signal decoupling. However, the accuracy of traditional single-dimension range and angle imaging is limited by the number of finite array elements and significant noise interference. Therefore, this paper proposes a multidomain joint estimation algorithm based on the Improved Bayesian Matching Pursuit (IBMP) method. The Bayesian method is based on the Bernoulli-Gaussian (BG) model, and the estimated parameters and support domain are iteratively updated in this method while adhering to the Maximum a Posteriori (MAP) criterion constraint to achieve the high-precision reconstruction of multidimensional joint signals. The final set of simulation and actual measurement results demonstrate that the proposed method can effectively solve the problem of range migration and improve the angle resolution of radar forward-looking imaging while exhibiting excellent noise robustness.
Single snapshot forward-looking imaging technology with high performance and resolution is crucial for enabling the development of automotive radars. However, range migration issues can limit the implementation of coherent integration methods, and improving system resolution is generally difficult due to hardware parameter limitations. Based on the Time-Division Multiplexing Multiple-Input-Multiple-Output (TDM-MIMO) forward-looking imaging systems of automotive millimeter wave radar, this paper proposes Doppler domain compensation and point-to-point echo correction measures for achieving multidomain signal decoupling. However, the accuracy of traditional single-dimension range and angle imaging is limited by the number of finite array elements and significant noise interference. Therefore, this paper proposes a multidomain joint estimation algorithm based on the Improved Bayesian Matching Pursuit (IBMP) method. The Bayesian method is based on the Bernoulli-Gaussian (BG) model, and the estimated parameters and support domain are iteratively updated in this method while adhering to the Maximum a Posteriori (MAP) criterion constraint to achieve the high-precision reconstruction of multidimensional joint signals. The final set of simulation and actual measurement results demonstrate that the proposed method can effectively solve the problem of range migration and improve the angle resolution of radar forward-looking imaging while exhibiting excellent noise robustness.
2023, 12(5): 986-999.
Super-resolution Direction of Arrival (DOA) estimation is a critical problem related to vehicle-borne Millimeter-wave radars that needs to be solved to realize accurate target positioning and tracking. Based on the common conditions of limited array aperture, low snapshot, low signal-to-noise ratio, and coherent sources with respect to vehicle-borne scenarios, a super-resolution DOA estimation method for a moving target with an MMW radar based on Range-Doppler Atom Norm Minimize (RD-ANM) is proposed herein. First, an array for receiving signals in the range-Doppler domain is constructed based on the radar echo of the moving target. Then, the compensation vector for the Doppler coupling phase of the moving target is designed to reduce the influence of target motion on DOA estimation. Finally, a multitarget super-resolution DOA estimation method based on the atomic norm framework is proposed herein. Compared to the existing DOA estimation algorithm, the proposed algorithm can achieve higher angular resolution and estimation accuracy owing to low signal-to-noise ratio and single snapshot processing conditions, as well as robust performance in processing coherent sources without sacrificing array aperture. The effectiveness of the proposed algorithm is proven via theoretical analyses, numerical simulations, and experiments.
Super-resolution Direction of Arrival (DOA) estimation is a critical problem related to vehicle-borne Millimeter-wave radars that needs to be solved to realize accurate target positioning and tracking. Based on the common conditions of limited array aperture, low snapshot, low signal-to-noise ratio, and coherent sources with respect to vehicle-borne scenarios, a super-resolution DOA estimation method for a moving target with an MMW radar based on Range-Doppler Atom Norm Minimize (RD-ANM) is proposed herein. First, an array for receiving signals in the range-Doppler domain is constructed based on the radar echo of the moving target. Then, the compensation vector for the Doppler coupling phase of the moving target is designed to reduce the influence of target motion on DOA estimation. Finally, a multitarget super-resolution DOA estimation method based on the atomic norm framework is proposed herein. Compared to the existing DOA estimation algorithm, the proposed algorithm can achieve higher angular resolution and estimation accuracy owing to low signal-to-noise ratio and single snapshot processing conditions, as well as robust performance in processing coherent sources without sacrificing array aperture. The effectiveness of the proposed algorithm is proven via theoretical analyses, numerical simulations, and experiments.
2023, 12(5): 1000-1013.
In recent years, millimeter-wave radar has been widely used in safety detection, nondestructive detection of parts, and medical diagnosis because of its strong penetration ability, small size, and high detection accuracy. However, due to the limitation of hardware transmission bandwidth, achieving ultra-high two-dimensional resolution using millimeter-wave radar is challenging. Two-dimensional high-resolution imaging of altitude and azimuth can be realized using radar platform scanning to form a two-dimensional aperture. However, during the scanning process, the millimeter-wave radar produces sparse tracks in the height dimension, resulting in a sparse sampling of the altitude echo, thus reducing the imaging quality. In this paper, a high-resolution three-dimensional imaging algorithm for millimeter-wave radar based on Hankel transformation matrix filling is proposed to solve this problem. The matrix filling algorithm restores the sparse sampling echo, which guarantees the imaging accuracy of the millimeter-wave radar in the scanning plane. First, the low-rank prior characteristics of the millimeter-wave radar's elevation-range section were analyzed. To solve the problem of missing whole rows and columns of data during sparse trajectory sampling, the echo data matrix was reconstructed using the Hankel transform, and the sparse low-rank prior characteristics of the constructed matrix were analyzed. Furthermore, a matrix filling algorithm based on truncated Schatten-p norm combining low-rank and sparse priors was proposed to fill and reconstruct the echoes to ensure the three-dimensional resolution of the sparse trajectory millimeter-wave radar. Finally, using simulation and several sets of measured data, the proposed method was proved to achieve high-resolution three-dimensional imaging even when only 20%–30% of the height echo was used.
In recent years, millimeter-wave radar has been widely used in safety detection, nondestructive detection of parts, and medical diagnosis because of its strong penetration ability, small size, and high detection accuracy. However, due to the limitation of hardware transmission bandwidth, achieving ultra-high two-dimensional resolution using millimeter-wave radar is challenging. Two-dimensional high-resolution imaging of altitude and azimuth can be realized using radar platform scanning to form a two-dimensional aperture. However, during the scanning process, the millimeter-wave radar produces sparse tracks in the height dimension, resulting in a sparse sampling of the altitude echo, thus reducing the imaging quality. In this paper, a high-resolution three-dimensional imaging algorithm for millimeter-wave radar based on Hankel transformation matrix filling is proposed to solve this problem. The matrix filling algorithm restores the sparse sampling echo, which guarantees the imaging accuracy of the millimeter-wave radar in the scanning plane. First, the low-rank prior characteristics of the millimeter-wave radar's elevation-range section were analyzed. To solve the problem of missing whole rows and columns of data during sparse trajectory sampling, the echo data matrix was reconstructed using the Hankel transform, and the sparse low-rank prior characteristics of the constructed matrix were analyzed. Furthermore, a matrix filling algorithm based on truncated Schatten-p norm combining low-rank and sparse priors was proposed to fill and reconstruct the echoes to ensure the three-dimensional resolution of the sparse trajectory millimeter-wave radar. Finally, using simulation and several sets of measured data, the proposed method was proved to achieve high-resolution three-dimensional imaging even when only 20%–30% of the height echo was used.
2023, 12(5): 1014-1030.
In radar-based road target recognition, the increase in target feature dimension is a common technique to improve recognition performance when targets become diverse, but their characteristics are similar. However, the increase in feature dimension leads to feature redundancy and dimension disasters. Therefore, it is necessary to optimize the extracted high-dimensional feature set. The Adaptive Genetic Algorithm (AGA) based on random search is an effective feature optimization method. To improve the efficiency and accuracy of the AGA, the existing improved AGA methods generally utilize the prior correlation between features and targets for pre-dimensionality reduction of high-dimensional feature sets. However, such algorithms only consider the correlation between a single feature and a target, neglecting the correlation between feature combinations and targets. The selected feature set may not be the best recognition combination for the target. Thus, to address this issue, this study proposes an improved AGA via pre-dimensionality reduction based on Histogram Analysis (HA) of the correlation between different feature combinations and targets. The proposed method can simultaneously improve the efficiency and accuracy of feature selection and target recognition performance. Comparative experiments based on a real dataset of the millimeter-wave radar showed that the average accuracy of target recognition of the proposed HA-AGA method could reach 95.7%, which is 1.9%, 2.4%, and 10.1% higher than that of IG-GA, ReliefF-IAGA, and improved RetinaNet methods, respectively. Comparative experiments based on the CARRADA dataset showed that the average accuracy of target recognition of the proposed HA-AGA method could reach 93.0%, which is 1.2% and 1.5% higher than that of IG-GA and ReliefF-IAGA methods, respectively. These results verify the effectiveness and superiority of the proposed method compared with existing methods. In addition, the performance of different feature optimization methods coupled with the integrated bagging tree, fine tree, and K-Nearest Neighbor (KNN) classifier was compared. The experimental results showed that the proposed method exhibits evident advantages when coupled with different classifiers and has broad applicability.
In radar-based road target recognition, the increase in target feature dimension is a common technique to improve recognition performance when targets become diverse, but their characteristics are similar. However, the increase in feature dimension leads to feature redundancy and dimension disasters. Therefore, it is necessary to optimize the extracted high-dimensional feature set. The Adaptive Genetic Algorithm (AGA) based on random search is an effective feature optimization method. To improve the efficiency and accuracy of the AGA, the existing improved AGA methods generally utilize the prior correlation between features and targets for pre-dimensionality reduction of high-dimensional feature sets. However, such algorithms only consider the correlation between a single feature and a target, neglecting the correlation between feature combinations and targets. The selected feature set may not be the best recognition combination for the target. Thus, to address this issue, this study proposes an improved AGA via pre-dimensionality reduction based on Histogram Analysis (HA) of the correlation between different feature combinations and targets. The proposed method can simultaneously improve the efficiency and accuracy of feature selection and target recognition performance. Comparative experiments based on a real dataset of the millimeter-wave radar showed that the average accuracy of target recognition of the proposed HA-AGA method could reach 95.7%, which is 1.9%, 2.4%, and 10.1% higher than that of IG-GA, ReliefF-IAGA, and improved RetinaNet methods, respectively. Comparative experiments based on the CARRADA dataset showed that the average accuracy of target recognition of the proposed HA-AGA method could reach 93.0%, which is 1.2% and 1.5% higher than that of IG-GA and ReliefF-IAGA methods, respectively. These results verify the effectiveness and superiority of the proposed method compared with existing methods. In addition, the performance of different feature optimization methods coupled with the integrated bagging tree, fine tree, and K-Nearest Neighbor (KNN) classifier was compared. The experimental results showed that the proposed method exhibits evident advantages when coupled with different classifiers and has broad applicability.
2023, 12(5): 1031-1043.
With the substantial improvement of Synthetic Aperture Radar (SAR) regarding swath width and spatial and temporal resolutions, a time series obtained by registering SAR images acquired at different times can provide more accurate information on the dynamic changes in the observed areas. However, inherent speckle noise and outliers along the temporal dimension in the time series pose serious challenges for subsequent interpretation tasks. Although existing state-of-the-art methods can effectively suppress the speckle noise in a SAR time series, outliers along the temporal dimension will interfere with the denoising results. To better solve this problem, this paper proposes an additive signal decomposition method in the logarithm domain that can suppress the speckle noise and separate stable data and outliers along the temporal dimension in a time series, thus eliminating the impact of outliers on the denoising results. When the simulated data are disturbed by outliers, the proposed method can achieve an approximately 3 dB improvement in the Peak Signal-to-Noise Ratio (PSNR) compared to the other state-of-the-art methods. On Sentinel-1 data, the proposed method robustly suppresses the speckle noise in a time series, and the obtained outliers along the temporal dimension provide reference data for subsequent interpretation tasks.
With the substantial improvement of Synthetic Aperture Radar (SAR) regarding swath width and spatial and temporal resolutions, a time series obtained by registering SAR images acquired at different times can provide more accurate information on the dynamic changes in the observed areas. However, inherent speckle noise and outliers along the temporal dimension in the time series pose serious challenges for subsequent interpretation tasks. Although existing state-of-the-art methods can effectively suppress the speckle noise in a SAR time series, outliers along the temporal dimension will interfere with the denoising results. To better solve this problem, this paper proposes an additive signal decomposition method in the logarithm domain that can suppress the speckle noise and separate stable data and outliers along the temporal dimension in a time series, thus eliminating the impact of outliers on the denoising results. When the simulated data are disturbed by outliers, the proposed method can achieve an approximately 3 dB improvement in the Peak Signal-to-Noise Ratio (PSNR) compared to the other state-of-the-art methods. On Sentinel-1 data, the proposed method robustly suppresses the speckle noise in a time series, and the obtained outliers along the temporal dimension provide reference data for subsequent interpretation tasks.
2023, 12(5): 1044-1055.
An improved Synthetic Aperture Radar (SAR) imaging algorithm is proposed to address the issues of low azimuth resolution and noise interference in the sparse sampling condition. Based on the existing L1/2 regularization theory and iterative threshold algorithm, the gradient operator is modified, which can improve the solution accuracy of the reconstructed image and reduce the load of calculation. Then, under full sampling and under-sampling conditions, the original and improved L1/2 iterative threshold algorithm are combined with the approximate observation model to image SAR echo signals and compare their imaging performance. The experimental findings demonstrate that the improved algorithm improves the azimuth resolution of SAR images and has higher convergence performance.
An improved Synthetic Aperture Radar (SAR) imaging algorithm is proposed to address the issues of low azimuth resolution and noise interference in the sparse sampling condition. Based on the existing L1/2 regularization theory and iterative threshold algorithm, the gradient operator is modified, which can improve the solution accuracy of the reconstructed image and reduce the load of calculation. Then, under full sampling and under-sampling conditions, the original and improved L1/2 iterative threshold algorithm are combined with the approximate observation model to image SAR echo signals and compare their imaging performance. The experimental findings demonstrate that the improved algorithm improves the azimuth resolution of SAR images and has higher convergence performance.
2023, 12(5): 1056-1068.
Synthetic Aperture Radar Tomography (TomoSAR) has emerged as a hot research topic in the field of SAR imaging, particularly for three-dimensional (3D) urban imaging in recent years. However, in TomoSAR 3D reconstruction, due to the phase unwrapping difficulty, periodic spectral peaks appear in the reconstruction results of the reflectivity profile along the elevation. This results in errors in estimating the elevation locations of the scatterers and causing layering effects in 3D imaging results, which is the elevation ambiguity. In light of this phenomenon observed in TomoSAR, a method for the adaptive adjustment of the elevation search range is proposed to improve the accuracy of the elevation estimation and reduce elevation ambiguity. In this method, the height of the scene is first estimated, the linear function of the elevation sampling center is subsequently constructed based on the height pre-estimations, and the search radius is finally calculated. Thereafter, the elevation search range of each pixel in the SAR image is determined and updated, preserving the true spectral peaks while isolating the ambiguity peaks. The experimental results for airborne and spaceborne measured data demonstrate that the proposed method significantly improves elevation ambiguity and artifacts-related issues while also improving the spatial concentration and continuity of 3D point clouds.
Synthetic Aperture Radar Tomography (TomoSAR) has emerged as a hot research topic in the field of SAR imaging, particularly for three-dimensional (3D) urban imaging in recent years. However, in TomoSAR 3D reconstruction, due to the phase unwrapping difficulty, periodic spectral peaks appear in the reconstruction results of the reflectivity profile along the elevation. This results in errors in estimating the elevation locations of the scatterers and causing layering effects in 3D imaging results, which is the elevation ambiguity. In light of this phenomenon observed in TomoSAR, a method for the adaptive adjustment of the elevation search range is proposed to improve the accuracy of the elevation estimation and reduce elevation ambiguity. In this method, the height of the scene is first estimated, the linear function of the elevation sampling center is subsequently constructed based on the height pre-estimations, and the search radius is finally calculated. Thereafter, the elevation search range of each pixel in the SAR image is determined and updated, preserving the true spectral peaks while isolating the ambiguity peaks. The experimental results for airborne and spaceborne measured data demonstrate that the proposed method significantly improves elevation ambiguity and artifacts-related issues while also improving the spatial concentration and continuity of 3D point clouds.
2023, 12(5): 1069-1080.
Most high-resolution Synthetic Aperture Radar (SAR) images of real-life scenes are complex due to clutter, such as grass, trees, roads, and buildings, in the background. Traditional target detection algorithms for SAR images contain numerous false and missed alarms due to such clutter, adversely affecting the performance of SAR images target detection. Herein we propose a feature decomposition-based Convolutional Neural Network (CNN) for target detection in SAR images. The feature extraction module first extracts features from the input images, and these features are then decomposed into discriminative and interfering features using the feature decomposition module. Furthermore, only the discriminative features are input into the multiscale detection module for target detection. The interfering features that are removed after feature decomposition are the parts that are unfavorable to target detection, such as complex background clutter, whereas the discriminative features that are retained are the parts that are favorable to target detection, such as the targets of interest. Hence, an effective reduction in the number of false and missed alarms, as well as an improvement in the performance of SAR target detection, is achieved. The F1-score values of the proposed method are 0.9357 and 0.9211 for the MiniSAR dataset and SAR Aircraft Detection Dataset (SADD), respectively. Compared to the single shot multibox detector without the feature extraction module, the F1-score values of the proposed method for the MiniSAR and SADD datasets show an improvement of 0.0613 and 0.0639, respectively. Therefore, the effectiveness of the proposed method for target detection in SAR images of complex scenes was demonstrated through experimental results based on the measured datasets.
Most high-resolution Synthetic Aperture Radar (SAR) images of real-life scenes are complex due to clutter, such as grass, trees, roads, and buildings, in the background. Traditional target detection algorithms for SAR images contain numerous false and missed alarms due to such clutter, adversely affecting the performance of SAR images target detection. Herein we propose a feature decomposition-based Convolutional Neural Network (CNN) for target detection in SAR images. The feature extraction module first extracts features from the input images, and these features are then decomposed into discriminative and interfering features using the feature decomposition module. Furthermore, only the discriminative features are input into the multiscale detection module for target detection. The interfering features that are removed after feature decomposition are the parts that are unfavorable to target detection, such as complex background clutter, whereas the discriminative features that are retained are the parts that are favorable to target detection, such as the targets of interest. Hence, an effective reduction in the number of false and missed alarms, as well as an improvement in the performance of SAR target detection, is achieved. The F1-score values of the proposed method are 0.9357 and 0.9211 for the MiniSAR dataset and SAR Aircraft Detection Dataset (SADD), respectively. Compared to the single shot multibox detector without the feature extraction module, the F1-score values of the proposed method for the MiniSAR and SADD datasets show an improvement of 0.0613 and 0.0639, respectively. Therefore, the effectiveness of the proposed method for target detection in SAR images of complex scenes was demonstrated through experimental results based on the measured datasets.
2023, 12(5): 1081-1096.
Vehicle targets in urban scenes have the characteristics of random distribution and can be easily disturbed by environmental factors during the detection process. Given the above issues, this paper proposes a detection method that utilizes multi-aspect Synthetic Aperture Radar (SAR) images for stationary vehicle target extraction. In the feature extraction stage, a novel feature extraction method called Multiscale Rotational Gabor Odd Filter-based Ratio Operator (MR-GOFRO) is designed for vehicle targets in multi-aspect SAR images, where the original GOFRO features are improved from four aspects—filter form, feature scale, feature direction and feature level. The improvement allows MR-GOFRO to adapt to possible variations in the target direction, scale, morphology, etc. In the image fusion stage, a Weighted-Non-negative Matrix Factorization (W-NMF) method is developed to adjust the feature weights from various images according to the feature quality. This method can reduce the quality degradation of the fusion features due to mutual interference between different aspects. The proposed method is verified on various airborne multi-aspect image datasets. The experimental results revealed that the feature extraction and feature fusion methods proposed in this paper enhance the detection accuracy by an average of 3.69% and 4.67%, respectively, compared with similar methods.
Vehicle targets in urban scenes have the characteristics of random distribution and can be easily disturbed by environmental factors during the detection process. Given the above issues, this paper proposes a detection method that utilizes multi-aspect Synthetic Aperture Radar (SAR) images for stationary vehicle target extraction. In the feature extraction stage, a novel feature extraction method called Multiscale Rotational Gabor Odd Filter-based Ratio Operator (MR-GOFRO) is designed for vehicle targets in multi-aspect SAR images, where the original GOFRO features are improved from four aspects—filter form, feature scale, feature direction and feature level. The improvement allows MR-GOFRO to adapt to possible variations in the target direction, scale, morphology, etc. In the image fusion stage, a Weighted-Non-negative Matrix Factorization (W-NMF) method is developed to adjust the feature weights from various images according to the feature quality. This method can reduce the quality degradation of the fusion features due to mutual interference between different aspects. The proposed method is verified on various airborne multi-aspect image datasets. The experimental results revealed that the feature extraction and feature fusion methods proposed in this paper enhance the detection accuracy by an average of 3.69% and 4.67%, respectively, compared with similar methods.
2023, 12(5): 1097-1111.
This study aims to address the unreasonable assignment of positive and negative samples and poor localization quality in ship detection in complex scenes. Therefore, in this study, a Synthetic Aperture Radar (SAR) ship detection network (A3-IOUS-Net) based on adaptive anchor assignment and Intersection over Union (IOU) supervise in complex scenes is proposed. First, an adaptive anchor assignment mechanism is proposed, where a probability distribution model is established to adaptively assign anchors as positive and negative samples to enhance the ship samples’ learning ability in complex scenes. Then, an IOU supervise mechanism is proposed, which adds an IOU prediction branch in the prediction head to supervise the localization quality of detection boxes, allowing the network to accurately locate the SAR ship targets in complex scenes. Furthermore, a coordinate attention module is introduced into the prediction branch to suppress the background clutter interference and improve the SAR ship detection accuracy. The experimental results on the open SAR Ship Detection Dataset (SSDD) show that the Average Precision (AP) of A3-IOUS-Net in complex scenes is 82.04%, superior to the other 15 comparison models.
This study aims to address the unreasonable assignment of positive and negative samples and poor localization quality in ship detection in complex scenes. Therefore, in this study, a Synthetic Aperture Radar (SAR) ship detection network (A3-IOUS-Net) based on adaptive anchor assignment and Intersection over Union (IOU) supervise in complex scenes is proposed. First, an adaptive anchor assignment mechanism is proposed, where a probability distribution model is established to adaptively assign anchors as positive and negative samples to enhance the ship samples’ learning ability in complex scenes. Then, an IOU supervise mechanism is proposed, which adds an IOU prediction branch in the prediction head to supervise the localization quality of detection boxes, allowing the network to accurately locate the SAR ship targets in complex scenes. Furthermore, a coordinate attention module is introduced into the prediction branch to suppress the background clutter interference and improve the SAR ship detection accuracy. The experimental results on the open SAR Ship Detection Dataset (SSDD) show that the Average Precision (AP) of A3-IOUS-Net in complex scenes is 82.04%, superior to the other 15 comparison models.
2023, 12(5): 1112-1124.
The realization of anti-jamming technologies via beamforming for applications in Frequency-Diverse Arrays and Multiple-Input and Multiple-Output (FDA-MIMO) radar is a field that is undergoing intensive research. However, because of limitations in hardware systems, such as component aging and storage device capacity, the signal covariance matrix data computed by the receiver system may be missing. To mitigate the impact of the missing covariance matrix data on the performance of the beamforming algorithm, we have proposed a covariance matrix data recovery method for FDA-MIMO radar based on deep learning and constructed a two-stage framework based on missing covariance matrix recovery-adaptive beamforming. Furthermore, a learning framework based on this two-stage framework and the Generative Adversarial Network (GAN) is constructed, which is mainly composed of a discriminator (D) and a generator (G). G is primarily used to output complete matrix data, while D is used to judge whether this data is real or filled. The entire network closes the gap between the samples generated by G and the distribution of the real data via a confrontation between D and G, consequently leading to the missing data of the covariance matrix being recovered. In addition, considering that the covariance matrix data is complex, two independent networks are constructed to train the real and imaginary parts of the matrix data. Finally, the numerical experiment results reveal that the difference in the root square mean error levels between the real and recovery data is 0.01 in magnitude.
The realization of anti-jamming technologies via beamforming for applications in Frequency-Diverse Arrays and Multiple-Input and Multiple-Output (FDA-MIMO) radar is a field that is undergoing intensive research. However, because of limitations in hardware systems, such as component aging and storage device capacity, the signal covariance matrix data computed by the receiver system may be missing. To mitigate the impact of the missing covariance matrix data on the performance of the beamforming algorithm, we have proposed a covariance matrix data recovery method for FDA-MIMO radar based on deep learning and constructed a two-stage framework based on missing covariance matrix recovery-adaptive beamforming. Furthermore, a learning framework based on this two-stage framework and the Generative Adversarial Network (GAN) is constructed, which is mainly composed of a discriminator (D) and a generator (G). G is primarily used to output complete matrix data, while D is used to judge whether this data is real or filled. The entire network closes the gap between the samples generated by G and the distribution of the real data via a confrontation between D and G, consequently leading to the missing data of the covariance matrix being recovered. In addition, considering that the covariance matrix data is complex, two independent networks are constructed to train the real and imaginary parts of the matrix data. Finally, the numerical experiment results reveal that the difference in the root square mean error levels between the real and recovery data is 0.01 in magnitude.
微信 | 公众平台 
