结合PLS表示与随机梯度的目标优化跟踪

针对实际视觉跟踪中目标表观与前背景的非线性变化，论文提出一种基于偏最小二乘分析(PLS)表示与随机梯度的目标优化跟踪方法。该方法将目标跟踪转化为表示误差与分类损失的联合优化问题。首先，为了提高算法对前背景表观变化的稳定性，利用PLS理论的非线性对目标区域的前背景信息进行表达，并通过空间聚类构造多个线性外观模型来描述目标区域的动态变化，建立带约束条件的表观特征库；然后，提出一种确定性搜索机制，构造联合优化目标函数，使表示误差与分类损失最小化；结合表观建模特点，构建随机梯度分类器，对模型进行增量特征更新，最终实现对目标的稳定准确跟踪。经多场景对比实验验证，该算法能有效应对目标前背景的多种复杂变化。

Object Optimization Tracking via PLS Representation and Stochastic Gradient

In order to improve the stability and accuracy of the object tracking under nonlinear conditions, an object tracking algorithm based on Partial Least Squares (PLS) representation and stochastic gradient object optimization tracking is proposed. In this method, object tracking is defined as an optimization task that minimizes the representation error and classification loss. Firstly, it expresses object appearance and background information by PLS theory, learns multiple low dimensional and discriminative subspaces to describe the nonlinear appearance changes of the object. Then, a joint optimization objective function based on deterministic search mechanism is proposed. Furthermore, an stochastic gradient classifier based on incremental features updating is proposed, and make sure that it can achieve a stable tracking. Experiments show favorable performance of the proposed algorithm on sequences where the targets undergo a variety complex changes on foreground and background.