基于自适应特征学习和表观建模的目标跟踪算法研究

本课题将结合信号处理、统计分析、机器学习和计算机视觉等交叉领域在图像全局和局部特征表示、非线性特征变换、信号压缩、分类器构造、信息融合和算法复杂度分析方面最新进展和热点研究，将自适应非线性特征学习、基于哈希（hashing）的二值特征分析、混合产生式和判别式统计模型设计、特征表示和表观建模联合学习、多任务判别学习以及算法运行复杂度分析等难点问题作为突破点进行研究，建立基于自适应特征学习和鲁棒表观建模的目标跟踪框架。通过有机集成相关研究成果与技术，面向公共开放视频集和互联网视频，研发能在复杂条件下进行鲁棒目标跟踪的原型系统，对相关算法进行对比、验证和完善。

This project is motivated by the recent development trends and hot research topics in several interdisciplinary fields (e.g., signal processing, statistical analysis, machine learning, computer vision, etc.), which typically concentrate on the following aspects: local or global feature representation, nonlinear feature transform, signal compression, classifier construction,information fusion,algorithmic complexity analysis, etc. Furthermore, this project aims to establish an effective object tracking framework based on adaptive feature learning and robust appearance modeling, which pay much attention to the studies of several challenging problems in the aspects of adaptive nonlinear feature learning, hashing-based binary feature analysis, hybrid generative and discriminative statistical model design, joint learning of feature representation and appearance modeling, multi-task discriminative learning, and algorithm-running complexity analysis. Using publicly available video datasets and Internet video data, this project generates the output pf a robust object tracking system by effectively integrating a number of relevant research outcomes and techniques, and further refines the system through empirical comparisons and validations with several related algorithms.