【作者】 王芳芳；

【导师】 杨红菊；

【作者基本信息】 山西大学 ， 计算机技术， 2018， 硕士

【摘要】 物体追踪是计算机视觉领域非常重要的研究方向,近年来备受研究者们的关注和青睐。视觉追踪是对视频第一帧图像中需要跟踪的物体进行标注,在后续的图像中标记出待追踪物体的位置和对应的边框。近年来基于卷积神经网络的视觉追踪算法取得了突破性的进展,但是这种方法仅提取图像的局部为特征,并没有考虑到不同图像中追踪物体位置之间的依赖关系。本文提出使用递归网络学习不同序列之间的依赖关系,同时使用卷积网络预测不同状态下每个动作的执行概率,使用策略梯度算法优化目标函数,提出了基于深度强化学习的视觉追踪算法。强化学习旨在构造一个马尔科夫决策过程,学习状态和动作之间的映射关系,将视觉追踪任务定义为马尔科夫决策过程,当前的物体在不同的时刻下通过一系列的动作得到待追踪物体的位置。长短记忆模型通过当前的输入和上一时刻隐含的状态输出作为网络的输入去学习不同序列的依赖关系。本文在介绍了基于深度学习的卷积神经网络和强化学习的策略梯度算法的基础上,深入分析了经典的深度学习的视觉追踪算法,并提出了基于深度强化学习的视觉追踪方法,经过实验对比分析,本文所提出的方法能够运用到实际的环境中并且提升了追踪效率。更多还原

【Abstract】 Visual tracking plays an important role in the field of computer vision and attach more attention by researchers in recent years.Visual tracking find the location of tracked object which is labeled in the first frame and annotated by human power.In recent years,the convolutional neural networks get breakthroughs in the field of visual tracking.However,which is only used to extract local image feature.The relationship between these frames in the ignored in current researches.In this thesis,we proposed a visual tracking method based on policy gradient method which utilized Convolutional Neural Networks to map the image to its corresponding actions and Long Short Time Memory to learn the relationship between history actions and current actions.Reinforcement learning is used to construct a Markov decision process to learn a map between states and actions.When introduced the reinforcement learning to visual tracking,we can apply MDP to visual tracking method: the current object is transformed by some actions to others state.The recurrent neural network utilized the current input and the previous hidden state to learn the map between history information and the current state.In this thesis we introduced the Convolutional Neural Networks based on deep learning and Policy gradient based on reinforcement learning.We also introduced visual tracking methods based on those algorithm.In this basement,we proposed a visual tracking method based on reinforcement learning and experiments demonstrate that our developed method could applied to real life and improved tracking efficiency.更多还原