基于卷积神经网络目标跟踪的AUV回收视觉导引算法

Visual Guidance Algorithm for AUV Recovery Based on CNN Object Tracking

水下自主对接回收技术是解决自主水下航行器(AUV)能源与信息传输问题,增强无人系统水下探测、隐蔽能力的主要手段。文中针对真实环境笼式水下基站回收设计水下视觉信标导引方案,提出一种改进的基于神经网络的检测-跟踪回收视觉导引算法:首先利用水下基站数据集训练卷积神经网络,进行目标检测;随后利用改进跟踪算法结合位姿空间信息实现鲁棒跟踪;最后通过改进Pn P-P3P位姿估计框架解决大偏移量下可观测信标灯数量过少的问题,有效扩展水下视觉导引作业空间,并通过作业空间仿真验证了灯阵设计与算法的有效性,提出了相关有效作业空间指标。水池光学导引实验以及在湖上真实环境下结合超短基线进行的声光联合导引实验,验证了改进检测-跟踪框架在工程上的可行性。

The development of autonomous undersea vehicle recovery technology is the main approach to solve problems pertaining to energy and information transmission and to enhance the underwater detection and concealment capabilities of unmanned systems. In this study, an underwater visual guidance scheme is designed for recovery with funnel-shaped docking stations in an actual environment. Additionally, an improved detect-by-tracking algorithm based on a convolutional neural network(CNN) is proposed. First, the CNN is trained using a docking station dataset to detect the target. Next, the improved tracking algorithm is combined with the position and attitude spatial information to achieve robust tracking. Finally, based on an improved Pn P-P3P position and attitude estimation framework, the problem of insufficient observable beacons under a large offset is solved, and the underwater visual guidance workspace is effectively expanded. The beacon array design and algorithm are validated via workspace simulation, and relevant effective workspace indexes are proposed. An optical guidance experiment is performed in a pool, and acousto–optic joint guidance is performed based on an ultrashort baseline in an actual lake test. The feasibility of the proposed framework for engineering is confirmed by the results obtained.