Accurate detection of sea-surface objects is vital for the safe navigation of autonomous ships. With the continuous development of artificial intelligence, electro-optical (EO) sensors such as video cameras are used to supplement marine radar to improve the detection of objects that produce weak radar signals and small sizes. In this study, we propose an enhanced convolutional neural network (CNN) named VarifocalNet * that improves object detection in harsh maritime environments. Specifically, the feature representation and learning ability of the VarifocalNet model are improved by using a deformable convolution module, redesigning the loss function, introducing a soft non-maximum suppression algorithm, and incorporating multi-scale prediction methods. These strategies improve the accuracy and reliability of our CNN-based detection results under complex sea conditions, such as in turbulent waves, sea fog, and water reflection. Experimental results under different maritime conditions show that our method significantly outperforms similar methods (such as SSD, YOLOv3, RetinaNet, Faster R-CNN, Cascade R-CNN) in terms of the detection accuracy and robustness for small objects. The maritime obstacle detection results were obtained under harsh imaging conditions to demonstrate the performance of our network model.

Object detection by shipborne unmanned aerial vehicles (UAVs) equipped with electro-optical (EO) sensors plays an important role in maritime rescue and ocean monitoring. However, high-precision and low-latency maritime environment small-object-detection algorithms remain a major challenge. To address this problem, this paper proposes the YOLO-BEV (“you only look once”–“bird’s-eye view”) model. First, we constructed a bidirectional feature fusion module—that is, PAN+ (Path Aggregation Network+)—adding an extremely-small-object-prediction head to deal with the large-scale variance of targets at different heights. Second, we propose a C2fSESA (Squeeze-and-Excitation Spatial Attention Based on C2f) module based on the attention mechanism to obtain richer feature information by aggregating features of different depth layers. Finally, we describe a lightweight spatial pyramid pooling structure called RGSPP (Random and Group Convolution Spatial Pyramid Pooling), which uses group convolution and random channel rearrangement to reduce the model’s computational overhead and improve its generalization ability. The article compares the YOLO-BEV model with other object-detection algorithms on the publicly available MOBDrone dataset. The research results show that the mAP_0.5 value of YOLO-BEV reached 97.1%, which is 4.3% higher than that of YOLOv5, and the average precision for small objects increased by 22.2%. Additionally, the YOLO-BEV model maintained a detection speed of 48 frames per second (FPS). Consequently, the proposed method effectively balances the accuracy and efficiency of object-detection in shipborne UAV scenarios, outperforming other related techniques in shipboard UAV maritime object detection.

Accurate positioning and state estimation of surface vessels are prerequisites to achieving autonomous navigation. Recently, the rapid development of 3D LiDARs has promoted the autonomy of both land and aerial vehicles, which has aroused the interest of researchers in the maritime community accordingly. In this paper, the state estimation schemes based on 3D LiDAR scan matching are explored in depth. Firstly, the iterative closest point (ICP) and normal distribution transformation (NDT) algorithms and their variants are introduced in detail. Besides, ten representative registration algorithms are selected from the variants for comparative analysis. Two types of experiments are designed by utilizing the field test data of an ASV equipped with a 3D LiDAR. Both the accuracy and real-time performance of the selected algorithms are systemically analyzed based on the experimental results. It follows that ICP and Levenberg–Marquardt iterative closest point (LMICP) methods perform well on single-frame experiments, while the voxelized generalized iterative closest point (FastVGICP) and multi-threaded optimization generalized iterative closest point (FastGICP) methods have the best performance on continuous-frame experiments. However, all methods have lower accuracy during fast turning. Consequently, the limitations of current methods are discussed in detail, which provides insights for future exploration of accurate state estimation based on 3D LiDAR for ASVs.

Image and video stitching techniques represent a significant research direction in the field of computer vision. They aim to generate images or videos with a wide field of view from multiple overlapping images or videos. These techniques find widespread application in virtual reality, autonomous driving, surveillance systems, digital cultural heritage preservation, and other fields. However, image and video stitching still encounter challenges, including large parallax, occlusion, weak texture, shape distortion, video jitter, and low efficiency. In recent years, these challenges have spurred the development of various new image and video stitching techniques. This work offers a comprehensive review of the latest advancements in image and video stitching techniques. First, we analyze traditional stitching methods and deep learning-based methods, discussing key techniques such as shape correction in image stitching and video stabilization. Second, we review datasets and evaluation metrics for image and video stitching tasks, and analyze the performance of mainstream algorithms on representative datasets. Finally, we discuss typical application scenarios and future research directions for image and video stitching. This work aims to provide researchers with a detailed overview of advancements in image and video stitching techniques, and to inspire improved stitching methods, thereby advancing the development of image and video stitching.