An Underwater Object Detection and Grasping System Based on an Improved YOLOv8 Model

At present, the Underwater Vehicle-Manipulator System(UVMS) is a effective tool for underwater resource exploration in the marine field. Accurate underwater target detection is a prerequisite for autonomous operations of UVMS and holds significant research value. However, challenges such as high noise levels in underwater environments, small target scales, and low background distinction increase the difficulty of detection. Therefore, the YOLO-MCA target algorithm is proposed, building upon improvements to YOLOv8. A Multi-Convolutional Context Attention (MCA) mechanism is designed that uses convolutional kernels of different sizes for contextual feature extraction and learning, thereby enhancing the ability to learn small targets. Additionally, a novel loss function named DA-CIOU is designed to dynamically evaluate the loss between the predicted box and the ground truth box, improving the learning capability for small target objects. Experiments demonstrate that YOLO-MCA performs well on the public UROD dataset and a self-constructed dataset, effectively detecting blurry, dense, and small-sized targets. Furthermore, we conducted autonomous grasping experiments in the Gazebo environment based on YOLO-MCA to validate the autonomous grasping performance of the UVMS.