Due to the complex underwater environment, autonomous underwater vehicles (AUVs) may fail due to unknown obstacles in the process of performing tasks. Therefore, studying the real-time obstacle detection and avoidance methods for AUVs based on forward-looking sonar is very important. This article proposes a set of obstacle avoidance algorithms based on the images collected by the multibeam forward-looking sonar carried by AUVs. First, a new obstacle detection network based on a dual-path feature network based on a vision transformer (ViT-DPFN), which combines a convolutional neural network (CNN) and transformer, is proposed to detect obstacles in sonar images quickly and accurately. Then, combined with the current attitude data of AUVs, the obstacle position collected by the sonar image is corrected. Finally, using the fixed obstacle size and position data, the cost function is established based on the model predictive control (MPC) method, the constraint conditions are designed, and the external collision avoidance constraint is established so that AUVs can avoid obstacles in the trajectory tracking process. The simulation and field experiment results show that the proposed method improves the obstacle detection accuracy and processing speed and ensures the navigation safety of AUVs in complex obstacle environments, proving the proposed method’s advancement and effectiveness. The underwater acoustic target detection (UATD) dataset is available at https://github.com/zhenghaotianHEU/UATD-sonar-dataset .

Sonar image is the main way for underwater vehicles to obtain environmental information. The task of target detection in sonar images can distinguish multi-class targets in real time and accurately locate them, providing perception information for the decision-making system of underwater vehicles. However, there are many challenges in sonar image target detection, such as many kinds of sonar, complex and serious noise interference in images, and less datasets. This paper proposes a sonar image target detection method based on Dual Path Vision Transformer Network (DP-VIT) to accurately detect targets in forward-look sonar and side-scan sonar. DP-ViT increases receptive field by adding multi-scale to patch embedding enhances learning ability of model feature extraction by using Dual Path Transformer Block, then introduces Conv-Attention to reduce model training parameters, and finally uses Generalized Focal Loss to solve the problem of imbalance between positive and negative samples. The experimental results show that the performance of this sonar target detection method is superior to other mainstream methods on both forward-look sonar dataset and side-scan sonar dataset, and it can also maintain good performance in the case of adding noise.

In the realm of oceanic exploration, sidescan sonar's significance is indisputable. However, the inherent challenges of low resolution and robust noise interference in sidescan sonar images have presented a formidable barrier to semantic segmentation in target regions. To address this, we propose a novel CGF-Unet framework, amalgamating Unet with global features, for precise and rapid sidescan sonar image segmentation. Leveraging both Transformers and Unet, CGF-Unet strategically introduces Transformer Blocks during downsampling and upsampling, amplifying access to comprehensive global insights and synergizing Transformer's potent sequence encoding with convolutional neural network's (CNN) holistic perception and spatial invariance. The incorporation of Conv-Attention within the Transformer Block streamlines model training parameters, accelerates training pace, and bolsters learning prowess. By implementing a weighted loss function, we navigate the challenge posed by skewed positive and negative samples, thereby elevating segmentation accuracy. Demonstrating its novelty, on distinct sidescan sonar data sets, we achieve exceptional mIOU scores of 89.3% and 86.5%, surpassing existing methodologies in precision. Remarkably, even amidst noise perturbation, the method maintains robust performance.

Forward-looking sonar is a critical tool for underwater target detection, and segmentation is an essential component of forward-looking sonar image processing. Accurate segmentation of sonar images is vital, but the complexity of the underwater environment introduces challenges, such as low resolution, significant noise, and blurred target edges. These factors make real-time, precise segmentation particularly difficult. To address these challenges, we propose a novel real-time segmentation network, the multilevel mixed feature fusion network (MLMFFNet), specifically designed for forward-looking sonar images. Our approach leverages a unique MFF module and a multiscale MFF module to extract both local and contextual information using deep convolutional networks, dilated convolutions, and partial convolution combinations for effective information integration. Additionally, we incorporate a context connection module to enhance feature fusion by utilizing high-level contextual information. To further improve accuracy, we introduce three weighted loss functions designed to address imbalanced sample distributions and blurred boundaries. Experimental evaluations on two distinct forward-looking sonar data sets demonstrate that MLMFFNet significantly outperforms many state-of-the-art general and sonar-specific semantic segmentation networks, delivering superior segmentation accuracy while maintaining real-time performance.

This paper addresses the trajectory tracking control problem of underactuated autonomous underwater vehicles (AUVs) subject to model uncertainties, external disturbances, and asymmetric input saturation. A robust predefined-time controller is proposed to ensure rapid convergence of tracking errors within a predefined time, rather than the asymptotic convergence achieved by most existing methods. First, an output redefinition-based dynamic transformation (ORDT) is introduced to overcome the relative degree deficiency caused by underactuation, enabling direct control of sway and heave motions. Subsequently, a predefined-time stable integrated controller is designed, which employs a Gaussian error function to handle asymmetric input saturation and uses a neural network (NN) with minimal learning parameters for disturbance compensation. The convergence time of the proposed controller can be preset by a simple parameter, without dependence on initial conditions or complex tuning. The stability analysis based on the Lyapunov method proves that all closed-loop signals converge to a sufficiently small neighborhood of the origin within a predefined time. Finally, simulation results demonstrate the effectiveness and superiority of the proposed control scheme.