基于稀疏编码网络结构的目标迁移跟踪算法研究

The implementation of object tracking task using deep learning model needs to deal with the challenges of limited samples, long training time and real-time tracking processing requirement, a sparse coding network structure based object transfer tracking algorithm research will be carried out in this project. First of all, the structural feature of the object in video frames is extracted, which is used to model the structured sparse representation via structured regularization term constraints. Then we introduce problem formulation of sparse representation hidden in deep learning model. We design the deep learning model of sparse dictionary pair and discuss its convergence. Secondly, to reduce the risk of model training overfitting, the project will further studies semi-supervised learning method of sparse dictionary pair deep learning model under the limited samples circumstance. We will define a deep feature object transfer similarity measure. The labels of unknown samples are achieved via transfer learning and similar samples are found to train sparse dictionary pair deep model. Based on those, we will further propose deep learning model based domain adaptation object transfer tracking algorithms used to track the object steadily by sharing information in different scenarios and overcome difficulties in complicated scene. Finally, a complex and diverse video tracking dataset is established. A simulation verification platform for deep learning based video tracking algorithms are constructed, which is used to verify the robustness of related key algorithms.

针对深度学习模型执行目标跟踪任务需要应对有限样本、模型训练时间长以及跟踪处理需要实时性等挑战，本项目拟开展基于稀疏编码网络结构的目标迁移跟踪算法研究。首先提取视频帧中目标的结构性信息，利用结构化的正则项约束表征目标的结构化稀疏表示。然后引出深度学习中潜藏的稀疏表达的问题阐述，设计稀疏字典对的深度学习模型，并对模型的收敛性进行讨论。其次，为了降低模型发生过拟合的风险，本项目将需要进一步研究有限样本下稀疏字典对深度模型的半监督学习方法，拟定义一种深度特征的目标迁移相似度量，通过迁移学习实现对未知样本的标注，寻找相似样本完成稀疏字典对深度模型的训练。在此基础上，进一步将提出基于深度学习模型下的域自适应目标迁移跟踪算法，使其在不同的场景下通过共享信息来稳定地跟踪目标，克服复杂场景下跟踪的困难。最后，建立复杂多样的视频跟踪数据库，构建基于深度学习的视频跟踪算法仿真验证平台，验证相关关键算法的鲁棒性。

智能视频监控在提高公共安全保障能力，推动智能防控构建，维护国家安全和社会安定方面发挥着日益重要的作用。视频目标跟踪是智能视频监控的关键技术之一。本项目深入研究了基于稀疏编码网络结构的目标迁移跟踪算法，不仅可以从理论角度提高了已有目标跟踪算法的性能，也从另一方面拓展了视觉目标跟踪算法的应用场景，有着明确的理论研究价值和工程应用需求。主要开展如下研究：(1）提出了基于自编码器对的相关滤波目标跟踪方法，采用密集循环采样策略增加训练样本数目，防止深度网络模型过拟合。(2）提出了基于多任务学习的目标跟踪方法，以克服单一任务对目标表征的局限性；(3）提出多任务多源融合的遥感数据处理方法，利用不同源数据和任务之间的相互依赖性来联合学习深度网络模型参数和滤波器参数，可用于目标跟踪和遥感数据标记。(4）提出了基于空间感知的多级损失目标跟踪对抗攻击方法，从高层类别和底层特征考虑设计了欺骗损失、漂移损失和基于注意力机制的特征损失来联合训练生成器，生成对抗样本，用于干扰目标跟踪器，导致目标运动轨迹发生偏移，实现了复杂场景下对目标快速有效的攻击。本项目的研究成果为当前目标跟踪方法关键问题的解决提供新思路与新方法，支撑智能视频监控系统建设，推进国家与社会安全保障的高效化。