再生核Hilbert空间中两阶段稀疏表示目标跟踪算法

Two-stage sparse representation objective tracking algorithm in reproducing kernel Hilbert space

中文关键词  目标跟踪  再生核Hilbert空间  核方法  稀疏表示  两阶段框架  加速近端梯度算法

英文关键词  objective tracking  reproducing kernel Hilbert space  kernel method  sparse representation  two-stage framework  accelerating proximate gradient algorithm

中文摘要

      在强干扰复杂环境下, 有效的特征选择对于目标跟踪模型的可解释性至关重要. 针对这一问题, 本文基于 再生核Hilbert空间(RKHS)理论, 对特征空间构建生成式的两阶段稀疏表示(TSSR)模型, 从而描述图像样本与字典 之间的非线性关系, 避免了在字典中引入大量的琐碎模板. 在第1阶段, 首先建立图像样本与字典在原始低维空间中 的关系, 然后利用批处理最小二乘算法求得稀疏表示系数的初值, 根据观测模型确定初始跟踪位置的分布; 在第2阶 段, 首先利用核方法将原始低维空间映射到高维特征空间, 然后提出一种基于核的加速近端梯度算法(KAPG), 从而 求得字典元素系数的核稀疏表示, 最终确定跟踪目标. 最后实验结果证明了本文所提出的TSSR方法在面对视角变 化和部分遮挡时的有效性.

英文摘要

      In the complex environment with strong interference, the effective feature selection is crucial for the objective tracking model interpretability. To tackle this issue, based on the reproducing kernel Hilbert space (RKHS) theory, this paper constructs a generative two-stage sparse representation (TSSR) model in the feature space to describe the nonlinear relationship between the image sample and the dictionary, while avoiding the introduction of a large size of trivial templates. In the first stage, we establish the relationship between the image sample and the dictionary in the original low-dimensional space, then the batch least square algorithm is used to obtain the initial value of the sparse representation coefficient. The distribution of the initial tracking position is determined according to the observation model. In the second stage, we use the kernel method to map the low-dimensional original space to the high-dimensional feature space, and then propose a kernel-based accelerated proximal gradient algorithm (KAPG) to obtain the sparse representation of the dictionary element coefficients. On this basis, the tracking target can be determined. The experimental results show the effectiveness of the proposed TSSR method in the face of viewing angle changes and partial occlusion.