视觉追踪是视频监控、智能交通、无人驾驶等产业的关键技术之一，具有重要的理论和应用价值。高性能视觉追踪的核心研究内容是目标外观的鲁棒和快速建模技术，要在真实复杂环境和场景下实现鲁棒和高效的目标外观建模仍是一个未被解决的难题。本项目以解决复杂环境和场景下的目标外观的鲁棒和高效建模中存在的关键科学问题和技术难点为出发点，研究有效解决当前追踪方法对复杂目标外观的表示能力不足和计算效率低下等问题的理论模型和关键技术。针对深度特征高维性、冗余性等问题，研究相关滤波优化驱动的特征编码模型和快速学习技术；针对现有模型的表示能力不足问题，研究基于类别先验的多层次多域表示模型和并行化结构设计；针对单一传感器的成像受限问题，研究基于RGB-T多视觉融合的高性能视觉追踪技术；针对视觉歧义问题，研究基于语义-视觉推理和学习的语言引导高性能视觉追踪技术。本项目的研究将为视频分析技术的推广应用提供理论支持和技术支撑。

Visual tracking is one of the key technologies of various emerging industries such as video surveillance, intelligent transportation and self-driving system, and has important theoretical and practical value The key problem of high performance (i.e., high precision and high efficiency) visual tracking is to efficiently learn the robust models of target appearance, which is unsolved as it is difficult to describe target appearance in real complex environments and scenes. This project aims to solve the key scientific problems and technical difficulties in the robust and efficient modeling of the target appearance in complex environments and scenes. First, to reduce the dimension and rebundancy of deep features, we will study the models of correlation filter optimization driven feature coding and the fast learning algorithm. Second, aiming at improving the representation capacity of exising models, we will utilize the prior knowledge of target category to design hierarchical and parallel architectures of multi-domain model. Third, we will investigate the fusion methods of RGB and thermal modalities to overcome the imaging limitations of individual RGB sensors for high performance visual tracking. Finally, to avoid visual ambiguity in visual tracking, we will propose the algorithms of semantic-visual reasoning and learning to acheive language-guided high performance visual tracking. The research of this project will provide theoretical and technical supports for the promotion and application of video-based analysis technologies.