In this paper, we propose a LiDAR-Inertial-Visual multi-sensor fusion SLAM system for dynamic environments. While many state of the art multi-sensor fusion SLAM systems can achieve excellent performance in static environment, achieving high accuracy and robustness in dynamic environment remains a significant challenge. In order to address the dynamic object elimination issue during mapping, a dynamic voxel judgment method based on octree map is proposed. Spe-cifically, we perform template matching between the current frame and submap to calculate the dynamic voxel occupancy rate of point clouds in Box, so as to effectively detect dynamic objects in the environment. At the same time, we continue to track dynamic objects and eliminate them in the map. Exten-sive experiments have been conducted on the KITTI dataset and our own dataset. Results shows the proposed method can effectively eliminate dynamic objects and improve the SLAM accuracy in dynamic environments.

Deep neural networks (DNNs) have been widely utilized in synthetic aperture radar (SAR) for automatic target recognition (ATR), demonstrating remarkable performance. Nevertheless, the vulnerability of DNNs to adversarial examples, particularly in SAR ATR tasks with high safety requirements, necessitates a critical examination. Existing adversarial attacks primarily concentrate on scenarios where classifier inputs consist solely of intensity information, leaving a significant research gap for attacks on classifiers that utilize polarimetric SAR (PolSAR) data. To address this gap, we propose a novel attack method aimed at PolSAR classifiers, where we manipulate the scattering matrix rather than its transformed real value images. Moreover, we incorporate a polarimetric feature constraint to enhance the stealth of the adversarial perturbations. This technique enables the generation of subtle yet effective perturbations concentrated on SAR object regions. Our experiments demonstrate high success rates in cheating state-of-the-art PolSAR classifiers and effectively evading advanced adversarial example detection methods.

This paper introduces a novel plug-and-play Multi-scale Feature Spatial Attention Fusion (MFSAF) module, aiming at enhancing the capabilities of convolutional neural networks (CNNs) in Synthetic Aperture Radar (SAR) ship classification tasks. The MFSAF module integrates spatial attention mechanisms and feature alignment strategies, providing a seamless integration into general CNNs to better capture ship features of different scales. The experimental results on the OpenSARShip2.0 and FUSARShip datasets demonstrate a significant improvement of the "baseline+MFSAF" model compared to baseline model, highlighting the effectiveness of the MFSAF module in capturing SAR ship features and its adaptability across different networks.

To address the problem of imbalanced datasets in SAR ship classification, this paper presents a novel cosine affinity (CA) loss that enhances the Gaussian affinity (GA) loss. The CA loss focuses on the angular relationship between feature vectors, prioritizing their direction over their magnitude, which is advantageous for high-dimensional space analysis. In addition, class weights are incorporated to compute weighted distances. Importantly, the proposed CA loss does not increase the computational complexity of algorithm, nor does it lead to overfitting problems associated with data-level techniques. Through various experiments, its effectiveness has been demonstrated by achieving the highest F1 score and recall compared to other existing loss functions, highlighting its superior ability to classify minority classes in FUSARShip.

Flood caused by torrential rain is one of the most influential meteorological disasters in the world. Currently, most of the flood change detection networks are based on homogeneous images. However, due to the influence of bad weather and satellite revisit cycle, the acquisition of homogeneous images is greatly limited. In this paper, a novel heterogeneous image change detection network based on prior feature estimation is proposed. In order to better guide the network to find the solution space related to change, we propose feature enhancement module to strengthen the water body features and introduce auxiliary information. At the same time, the fusion module is designed to solve the problem that the feature space of heterogeneous images is difficult to align while mapping water features into changing features. Experimental results on the CAU-Flood dataset demonstrate the effectiveness of our network.

Segmentation of Synthetic Aperture Radar (SAR) imagery constitutes the cornerstone of SAR image analysis, with sea-land segmentation in SAR images playing a crucial role in determining the precision of subsequent sea surface target detection. This study introduces an integrated approach for sea-land segmentation and coastline detection in SAR imagery, aiming to overcome the limitations posed by the traditional separation of these two tasks. In essence, the proposed method merges a segmentation module and an edge detection module, employing a hollow convolution and a global context mechanism. Additionally, the approach utilizes a cross-entropy loss function incorporating multiple losses with adaptive weighting, thereby enhancing the richness of the extracted feature information. To validate the algorithm’s efficacy, a specialized dataset for sea-land segmentation and coastline detection is constructed, utilizing the GRD data format from the Sentinel-1 satellite. Experimental outcomes show that the presented algorithm achieves scores of 0.988 and 0.981 on the Intersection over Union (IOU) metrics for sea-land segmentation, and 0.569 and 0.401 on the Optimal Dataset Scale (ODS) F1 and ODS IOU metrics for coastline detection.

Abstract Self‐supervised learning‐based depth completion is a cost‐effective way for 3D environment perception. However, it is also a challenging task because sparse depth may deactivate neural networks. In this paper, a novel Sparse‐Dense Depth Consistency Loss (SDDCL) is proposed to penalize not only the estimated depth map with sparse input points but also consecutive completed dense depth maps. Combined with the pose consistency loss, a new self‐supervised learning scheme is developed, using multi‐view geometric constraints, to achieve more accurate depth completion results. Moreover, to tackle the sparsity issue of input depth, a Quasi Dense Representations (QDR) module with triplet branches for spatial pyramid pooling is proposed to produce more dense feature maps. Extensive experimental results on VOID, NYUv2, and KITTI datasets show that the method outperforms state‐of‐the‐art self‐supervised depth completion methods.

Self-supervised monocular depth estimation from video sequences is promising for 3D environments perception. However, most existing methods use complicated depth networks to realize monocular depth estimation, which are often difficultly applied to resource-constrained devices. To solve this problem, in this letter, we propose a novel encoder-decoder-based lightweight depth network (LD-Net). Briefly speaking, the encoder is composed of six efficient downsampling units and the Atrous Spatial Pyramid Pooling (ASPP) module. The decoder consists of some novel upsampling units that adopt the sub-pixel convolutional layer (SP). Experiments tested on the KITTI dataset show that the proposed LD-Net can reach nearly 150 frames per second (FPS) on GPU, and remarkably decreases the model parameters while maintaining competitive accuracy compared with other state-of-the-art self-supervised monocular depth estimation methods.

The self-supervised learning-based depth and visual odometry (VO) estimators trained on monocular videos without ground truth have drawn significant attention recently. Prior works use photometric consistency as supervision, which is fragile under complex realistic environments due to illumination variations. More importantly, it suffers from scale inconsistency in the depth and pose estimation results. In this paper, robust geometric losses are proposed to deal with this problem. Specifically, we first align the scales of two reconstructed depth maps estimated from the adjacent image frames, and then enforce forward-backward relative pose consistency to formulate scale-consistent geometric constraints. Finally, a novel training framework is constructed to implement the proposed losses. Extensive evaluations on KITTI and Make3D datasets demonstrate that, i) by incorporating the proposed constraints as supervision, the depth estimation model can achieve state-of-the-art (SOTA) performance among the self-supervised methods, and ii) it is effective to use the proposed training framework to obtain a uniform global scale VO model.