Enhancing Autonomous Recognition in Deep-Sea Environments With AI-Based Target Detection and 3D Depth Estimation

Abstract The authors aim to develop an automatic target detection system for 11,000-meter-class underwater vehicles that enables high-precision target detection and tracking in deep-sea environments. The object detection with transformers (DETR) detection model and the Byte-Track tracking algorithm have been implemented. During demonstrations in June–July 2023, real-time target detection and tracking were achieved. However, issues have become apparent, such as decreased detection accuracy when objects are partially hidden or overlapped, and differences between training data and real-world conditions. To address these issues, the authors added data of hidden targets and newly collected ocean trials to the training process, creating a new model providing superior detection accuracy. Additionally, optimizing the tracking algorithm parameters enhanced the tracking stability. Also, AI-based depth estimation (CREStereo) was introduced to improve the accuracy of 3D measurements. This method was demonstrated in October 2024. The disparity map creation method showed improvement compared to earlier approaches. These improvements increased the reliability of real-time automatic sampling tasks in deep-sea environments, marking an important step toward practical application. Nevertheless, challenges for future research remain in detecting unknown or overlapping objects. Furthermore, AI-based depth estimation requires higher processing speeds. Further optimization must be done for real-time processing.