基于元注意力机制的水下机器人多任务跟踪决策研究

The underwater robot plays a rather important role in many areas such as oceanographic examination, intelligence detection and so on. Its working scene is shaking and low visibility because of motion blur, marine snow, inhomogeneous illumination and so on. These factors result in the unstable underwater tracking and drift. Another reason is that the current traking methods only consider the local and short-term tracking, lack the global and long term tracking prediction, and interaction with the tracking scene. So, in order to overcome these problems, this application proposes deep multi-task underwater tracking decision methods based on meta attentional mechanism, from the view of meta deep multi-task reinforcement learning. The research of this application focuses on the decision problems of tracking feature and tracking action, attentional mechanism and tracking drift correction. Firstly, the application designs a deep multi-task tracking feature decision network, which adaptively selects the combination of different kinds of features and the depth of the network, based on the current tracking environment. Secondly, the application develops meta deep multi-task tracking action decision network and its optimizing method, in order to control the sequentially iterative movement of the tracking rectangle to locate the target region accurately. In the following, the application develops deep multi-task Actor-Critic learning network for target tracking, which critics the tracking policy according to multiple views such as tracking and verify, meta valued network, Actor-Critic GAN, in order to active correct the tracking drift, and improve the tracking robustness. Next, the project designs a deep multi-task correlation filter tracking decision network based on attentional mechanism, which adaptively chooses the subset of correlation filters set, focuses on and updates the combination of best multi-task correlation filter modals. Finally, lightweight network design methods are developed to accelerate the above proposed meta deep multi-task tracking decision network, and improve its efficiency. The visual target tracking experiment of underwater robot is carried out. The above research will give theoretical and technological support to the navigation of underwater robot.

水下机器人在海洋科考、情报侦测等领域具有重要的应用价值，其低可见度的晃动工作环境如模糊、海雪、光照不均导致水下跟踪不稳定、易漂移。原因之一是当前水下跟踪器仅考虑局部帧、短时跟踪，缺乏全局长时跟踪预测，与场景交互不足。为此，本申请从元多任务强化学习的视角，提出基于元注意力机制的深度多任务水下跟踪决策方法，研究特征和跟踪动作决策、漂移纠正及注意力机制。首先，设计多任务跟踪特征决策网络，自适应选择跟踪网络深度。其次，提出元跟踪动作决策网络及优化方法，使跟踪框顺序迭代定位目标。再次，发展多任务Actor-Critic跟踪网络，通过跟踪验证、元指导网络、Actor-Critic生成对抗网络等多视角评判跟踪策略，主动纠正漂移。然后，设计时空、通道注意力机制的多任务相关滤波决策网络，聚焦最佳跟踪模块。最后，设计轻量化决策网络，开展水下机器人跟踪实验验证。该研究有望为水下机器人导航提供新的理论和技术支持。

本项目从元深度多任务强化学习的视角研究水下目标跟踪，首先构建了水下弱小目标跟踪与检测真实图像数据集，之后分别提出了多任务目标跟踪特征决策网络，孪生元多任务跟踪决策策略，多任务目标跟踪模板更新时机和策略机制选择，基于空间注意力机制相关滤波跟踪决策网络、可变形注意力机制的Transformer决策网络等跟踪决策模型，通过水下目标跟踪器与跟踪场景的交互，提高了跟踪器的鲁棒性、定位精度与成功率。最后对发展的跟踪模型进行轻量化，并利用水下机器人测试所发展元多任务跟踪决策模型的有效性。项目组在机器人领域及图像视频领域顶级期刊及会议如IEEE TIP、TMM、TCSVT、RAL、PR、ECCV、ICRA等发表高水平学术论文10余篇，申请授权国家发明专利8项，培养研究生7名。