毕业生知识库

仿生机器鱼作为一种新型的水下机器人，具有良好的机动性和游动效率，应用前景广阔。然而在仿生机器鱼上应用视觉系统却面临着诸多的困难，基于视觉传感的仿生机器鱼目标跟随控制更是面临着重重挑战。但视觉传感对于实现仿生机器鱼的智能化，增强仿生机器鱼的自主作业能力，有着非常重要的意义，更是仿生机器鱼工程化应用的基础。

 

本文围绕着仿生机器鱼的目标跟随控制任务，开发了一条具有三维运动能力的小型仿生机器鲨鱼，基于该仿生机器鲨鱼，设计了适用于仿生机器鱼的视觉稳定系统和主动视觉跟踪系统，并最终利用强化学习算法实现了仿生机器鱼的目标跟随控制。本文的主要研究成果如下：

 

第一，开发了具有视觉稳定功能的小型仿生机器鲨鱼平台。具体来说，设计了一个三关节驱动的尾部以及一对单自由度的胸鳍，用以实现仿生机器鱼的基本运动功能；设计了一个可沿偏航方向旋转的图像增稳云台，用以实现视觉稳定功能；设计了仿生机器鱼的电控系统，用以实现传感器信号采集、控制信号输出和通信等功能；设计了仿生机器鱼的软件系统，作为实现仿生机器鱼各项功能的基本架构。基于此平台，进一步设计了基于中枢模式发生器（Central Pattern Generator，CPG）模型的仿生机器鱼三维运动控制方法，并通过仿生机器鱼运动控制实验验证了所提控制方法的有效性，分析了特征参数对仿生机器鱼运动性能的影响。

 

第二，设计了仿生机器鱼的视觉稳定系统。仿生机器鱼的波动推进方式带来的头部晃动为仿生机器鱼的视觉信号的采集带来困难。为了解决这一问题，本文基于所设计的图像增稳云台，以摄像头上的惯性测量单元（Inertial Measurement Unit，IMU）作为反馈信号，以机器鱼身体上的IMU作为前馈信号，构建了一个前馈反馈控制器用以保持摄像头姿态稳定，从而实现视觉稳定功能。通过实验，验证了所设计控制器的效果，并通过对比实验，验证了前馈反馈控制结构的优越性。

 

第三，设计了仿生机器鱼的主动视觉跟踪系统。视觉稳定系统使得仿生机器鱼的摄像头能够在一定范围内自由旋转，以此为基础，利用核化相关滤波（Kernelized Correlation Filter，KCF）算法对机器鱼所采集图像中的目标物体进行视觉跟踪，同时根据图像目标位置推断目标物体相对摄像头的方位，并以此方位控制主动视觉跟踪系统令摄像头趋向于目标物体旋转，从而实现对目标物体的主动视觉跟踪。通过实验，验证了主动视觉跟踪系统的有效性。

主动视觉跟踪系统大幅提高了对目标物体进行视觉跟踪的可靠性，从而为目标跟随控制奠定了基础。

 

第四，基于强化学习算法，实现了仿生机器鱼的目标跟随控制。以主动视觉跟踪系统得到的目标物体相对于仿生机器鱼鱼体的方位为基础，设计了基于深度确定策略梯度（Deep Deterministic Policy Gradient，DDPG）强化学习算法的目标跟随控制器。该控制器以目标物体的方位为输入，以仿生机器鱼运动参数为输出，在连续的状态空间和动作空间的环境中实现了对目标物体的跟随控制。通过仿真实验，探讨了该控制器的鲁棒性、在系统发生改变时的自适应能力以及系统中的滞后环节对控制器的影响。在物理实验中，利用该强化学习算法，成功地实现了对目标物体的跟随控制。

 

最后，对本文的工作和研究成果进行了总结，并指出了进一步工作开展的方向。

As a new kind of underwater robot, robotic fish is characterized by high maneuverability and high propulsive efficiency, which has broad application prospects. However, applying vision systems to robotic fish could be really challenging, letting along controlling the robotic fish to track a target object based on embedded vision. Despite the challenge, vision sensing is the key to the intelligence of robotic fish. It can make the robotic fish able to accomplish a complicated task autonomously, which is the base of applying the robotic fish to engineering applications.

 

This thesis mainly focuses on the target tracking task for a robotic fish with a vision system.

A small-sized robotic shark with 3-D maneuverability is developed. A camera stabilizing system and an active tracking system are designed for the robotic shark. A tracking controller based on reinforcement learning is proposed to guide the robotic fish to track the target object. The main research achievements of this thesis are listed as follows.

 

Firstly, a small-sized robotic shark with a camera stabilizing system is developed. Specifically, the mechanism of a robotic shark with a three-joints tail and a pair of single-joint pectoral fins is designed to realize basic 3-D maneuverability. A pan-tilt inside the robotic fish, which has one degree of freedom along the yaw axis, is designed to stabilize the camera. The electronic system of the robotic shark is designed to sample the signal of sensors, to output the control signals and to communicate with upper-computer. The software structure is designed to make the robotic shark intelligent. Based on the small-sized robotic shark, a 3-D motion control method based on the Central Pattern Generator (CPG) model is proposed. A series of experiments are carried out to verify the effectiveness of the proposed motion control method and to analyze the influence of the intrinsic parameters of CPG model on the swimming performance.

 

Secondly, the camera stabilizing system is proposed. The swimming motion of the robotic fish causes its head to shake inevitability, resulting in serious degradation of the image captured by the camera in the head of robotic fish. To solve the problem, a feedback-feedforward controller is proposed for the pan-tilt inside the robotic fish. The controller uses the signal of the Inertial Measurement Unit (IMU) on the camera as the feedback signal and the signal of the IMU on the body of robotic shark as the feedforward signal to maintain the attitude angle of the camera stable and to maintain the image captured by the camera stable. A series of experiments are carried out to verify the designed controller and its advantages over the traditional controller.

 

Thirdly, the active tracking system is proposed. With the camera stabilizing system, the camera is able to point at any direction over a specific range. Based on that, an active tracking controller is proposed. This controller uses the position of the target object, which is detected by the Kernelized Correlation Filter (KCF) using the image captured by the onboard camera, to control the camera stabilizing system to point at the target object. Experiments are carried out to verify the effectiveness of the active tracking system. The active tracking system greatly reduces the risk of losing the target object while the robotic fish is swimming, which is the base of the target tracking control system.

 

Fourthly, a target tracking control system based on reinforcement learning is proposed. Using the position of the target object provided by the active tracking system as input and the parameters of the CPG model as output, a target tracking controller based on Deep Deterministic Policy Gradient (DDPG) is proposed. This reinforcement learning controller works in environments where both state space and action space are continuous. Through simulation experiments, the robustness of the controller, the adaptivity of the controller and the influence of the delay on the system are discussed. Finally, experiments on the actual robotic shark are carried out, in which the robotic shark successfully tracks the target object.

 

In the end of this thesis, conclusions are given, and further research works are listed.