模拟人脑多重储存记忆模型的在线视觉跟踪研究

The last three years have witnessed an expeditious development in online visual tracking with several benchmarks proposed. The soundness and fairness of these evaluation systems attract increasing attention from researchers in the tracking field. This also gives rise to many excellent tracking methods. Especially introducing multi-expert strategy, correlation filters and deep learning into the tracking framework proves to be superior to the traditional generative or discriminative appearance model based methods. This project stands on the aforementioned academic frontier, and focuses on proposing a new human perception and cognition inspired online visual tracking framework. This will be mainly based on the studies of Atkinson-Shiffrin human memory model and human semi-supervised learning mechanism in the cognitive science, and also fit the characteristics of online visual tracking. The main research contents are tree-fold: (1) designing the whole tracking framework, mainly focusing on long-term and short-term memories with sample information stored and the control processes in these two stores; (2) constructing the control processes in the short-term store based on Gaussian processes regression for making the initial tracking decision and providing feedback for the long-term store; (3) constructing the coding processes in the long-term store based on the theories of sparse coding or correlation filters, for the correction of the initial tracking decision. This project has both the theoretical and application significances, aims to improve the tracking performance significantly, and provides innovative theories and key techniques for some related research areas.

近三年，在线视觉单目标跟踪领域提出了多个评测库，其系统、公平的评测体系吸引了更多研究者参与，也涌现出了众多优秀跟踪算法。特别是多决策融合策略、相关滤波以及深度学习的引入，对比传统产生式或判别式方法优势明显。本项目立足于该领域前沿技术理论，遵循该领域应用特点，基于认知学理论中有关人脑的多重储存记忆模型和半监督学习机制的研究，拟提出一种模拟人脑感知的在线视觉单目标跟踪框架。主要研究内容包括：（1）构建整体跟踪框架，包括长短时样本信息的储存结构以及在跟踪过程中依赖于此结构的控制过程；（2）模拟人脑的半监督学习机制，依赖于短时储存结构拟构建基于高斯过程回归的控制过程，用于初始的跟踪决策以及对长时储存结构的信息反馈；（3）依赖于长时储存结构基于稀疏表示或相关滤波进行信息再加工编码，用于初始跟踪决策的校正。本项目兼顾理论创新和实际应用，期望显著提升跟踪算法性能，并为相关研究提供创新性理论和关键技术。

近些年，在线视觉单目标跟踪领域提出了多个评测库，其系统、公平的评测体系吸引了更多研究者参与，也涌现出了众多优秀跟踪算法。特别是多决策融合策略、相关滤波以及深度学习的引入，对比传统产生式或判别式方法优势明显。本项目立足于该领域前沿技术理论，遵循该领域应用特点，基于认知学理论中有关人脑的多重储存记忆模型和半监督学习机制的研究，提出了一种模拟人脑感知实现多决策融合的在线视觉单目标跟踪框架。主要研究内容包括：（1）构建整体跟踪框架，包括长短时样本信息的储存结构以及在跟踪过程中依赖于此结构的控制过程；（2）模拟人脑的半监督学习机制，依赖于短时储存结构拟构建基于高斯过程回归的控制过程，用于初始的跟踪决策以及对长时储存结构的信息反馈；（3）依赖于长时储存结构基于相关滤波或孪生全卷积神经网络互相关操作进行信息再加工编码，用于初始跟踪决策的校正。本项目兼顾理论创新和实际应用，在显著提升跟踪算法性能的同时，为其它相关研究提供了创新性理论和关键技术支撑，相关理论和算法发表在多个国际重要学术会议和学术期刊上。比如，发表顶级国际期刊IEEE Trans. Pattern Anal. Mach. Intell.（IEEE TPAMI；1/132；IF：17.730）论文1篇，顶级国际会议Proc. IEEE Conf. Comput. Vis. Pattern Recognit.（CVPR） 2018、Proc. Eur. Conf. Comput. Vis.（ECCV） 2018、Proc. Int. Joint Conf. Artif. Intell.（IJCAI）2018论文各1篇，重要国际会议Proc. IEEE Int. Conf. Pattern Recognit.（ICPR）2018论文2篇。