Modern millimeter-wave (mmWave) radar heat map-based object detection techniques have shown significant potential. However, radar heat maps still pose challenges for fine-grained object differentiation, and fusing camera maps with radar heat maps remains challenging. This paper proposes a multi-sensor radar object segmentation (ROS) method that fuses radar range-angle (RA) heat maps with camera RGB maps, leveraging the advantages of camera sensors for extracting semantic information and mmWave radar for extracting object position information. A method for mapping RGB to the RA coordinate system (RMTR) is proposed, eliminating the need for additional training branches and sensor calibration. A Transformer is utilized to predict the range and angle matrices of each object separately from a global perspective. These matrices then undergo an element-wise multiplication (Hadamard product) operation to generate the final pseudo-RA map. For the fusion method, this paper introduces the cross-attention fusion module (CAF), which uses the cross-attention mechanism to achieve more efficient fusion by treating the pseudo-RA feature map as the Query and the RA feature map as the Key and Value. Extensive experiments show that the proposed method achieves state-of-the-art (SOTA) performance on the CRUW and CARRADA datasets, with fewer parameters and reduced computational complexity. Our code for RC-ROSNet is available at https://github.com/Zhuanglong2/RCROSNet.

When registering synthetic aperture radar (SAR) images with inferior quality, general methods based on a single feature usually suffer from feature scarcity. To address the problem, a multi-feature registration algorithm is proposed in this paper. This algorithm utilizes scale-invariant feature transform (SIFT) feature points and obvious edge points as the keypoints of SAR images. Meanwhile, the ratio of exponentially weighted average (ROEWA) is adopted to calculate image gradient and extract edge features. Then, it employs gradient location-orientation histograms (GLOH) to construct feature vectors and matches keypoints based on Euclidean distance. Experiments on SAR image registration demonstrate that the proposed registration algorithm can provide more matching feature points and improve the registration accuracy compared with SIFT, SAR-SIFT and SIFT-Harris algorithms.

Wide angle synthetic aperture radar (WASAR) receives the echo data from a large azimuth angle, which causes aspect dependent scattering. Azimuth-range decouple sparse WASAR imaging based on group sparsity can accommodate the aspect dependent scattering. However, in the iterations of this method, the azimuth-range decouple operations are implemented, which causes huge computational costs. To solve this problem, this paper presents a novel complex image-based WASAR sparse imaging method. The proposed method can achieve identical images to those obtained via the azimuth-range decouple algorithms with fully sampled raw data. The computational complexity is also decreased dramatically. The image enhancement performances are improved compared with the sparse SAR imaging method based on complex images without group sparsity. Experimental results on real data validate the proposed method.

Semi-supervised domain adaptation methods can decrease the cost in producing training samples for scene classification. But they may deliver unsatisfactory performances because of the feature distribution bias between the previously labeled data and new aerial images caused by different imaging conditions. In order to address this problem, an original semi-supervised, rotation-robust domain adaptation (SRDA) framework is proposed to decrease the effect of feature distribution bias from the aspect of reducing the effect of rotation variance on representing scenes in aerial images and decreasing the influence of intra-class diversity on land-cover classification. The SRDA framework and some state-of-the-art domain adaptation methods are experimented on the UC Merced dataset to prove its superiority. The experimental results show the superiority of the proposed SRDA method over most of the previous domain adaptation approaches by at least 2% in overall accuracy.

Traditional polar format algorithm (PFA) is widely used in high-resolution and highly-squinted (HRHS) synthetic aperture radar (SAR) systems, because of its unique decoupling ability for any squint angle. However, the image size limit in traditional PFA, affected by the wavefront curvature, is too small to meet the modern wide-scene requirements. To break the limit, a new two-step PFA is proposed in this letter. It uses PFA twice. The first PFA is to obtain a coarse focused big image, from which a number of subimages can be filtered out. Each subimage could be recovered to the original phase history domain and reprocessed by another modified PFA, i.e., second PFA, to get a precisely focused subimage. The final full-image can be obtained easily by mosaicking subimages together. This two-step approach is derived in detail and could obviously enlarge the valid PFA scene. It is validated by simulated and real data.

Convolutional neural network (CNN)-based methods have become the mainstream in radar ship recognition. However, these methods suffer from two common problems. First, the training samples consist largely of common ship types, giving them an overwhelming numerical advantage over rare ship types. As a result, CNN-based recognition algorithms fail to classify rare ship types correctly. Second, huge high-resolution slices result in heavy computational burdens. To solve the first problem, namely, the class imbalance problem, this letter proposes a CNN training method that combines deep metric learning (DML) with gradually balanced sampling. DML obtains the center of each class in the feature space and performs clustering equally. Gradually balanced sampling adopts a smooth transition from instance-aware resampling to class-aware resampling to improve the recognition rate drop caused by traditional resampling methods. As for the second problem, to reduce the computational complexity of high-resolution synthetic aperture radar (SAR) images, a lightweight CNN is also proposed.

Correlation Filters (CFs) are widely applied to visual tracking because of their effectiveness and efficiency. However, the online learning of CFs is problematic because of the increasing number of training samples as the tracking process goes on. To solve this issue, most CFs-based methods decouple the learning phase and updating phase of the correlation filters and then use a simple linear interpolation to fuse the newly learned correlation filters with the old ones for fast computing. Nevertheless, the linear interpolation may be an unwise way to update the model. In this letter, we propose a smooth incremental learning framework of CFs. In our method, the increments of the correlation filters are smoothly learned by minimizing an optimization problem in each frame, thus avoiding the ad hoc linear interpolation. The optimization problem can be efficiently solved via the Alternating Direction Method of Multipliers (ADMM). Furthermore, a rotation estimation strategy is introduced to enable the correlation filters to accurately estimate the rotation angle of the target. Experiments show the superiority of the incremental learning framework and the rotation estimation method.

The approximation of planar wavefront in traditional Polar Format Algorithm(PFA) limits the effective scene size of PFA for high-resolution squinted Synthetic Aperture Radar (SAR). In this paper, a new wide-field PFA based on quadtree sub-image segmentation is proposed. The original phase history of the full-beam is filtered recursively to generate multiple sub-beams by sub-image recursive segmentation. As long as each sub-beam is filtered narrow enough, standard PFA could be implemented to produce a number of fully focused sub-image. Finally, all fully focused sub-images are mosaicked to get a big image perfectly focused. This divide-and-conquer approach breaks the image size limit in traditional PFA, extensively enlarges the effective focused scene. The processing flows are derived in detail and the algorithm is validated by measured data.