基于随机投影和稀疏表示的跟踪算法

Visual Tracking Based on Random Projection and Sparse Representation

针对目标跟踪过程中存在的诸多技术问题,该文提出一种鲁棒的目标跟踪方法。首先,该文采用基于稀疏表示的全局模板描述目标的表观状态,通过构造正负模板以区分目标和背景;然后采用随机投影法对表示模板和候选目标进行降维,以降低算法的时间复杂度;采用粒子滤波法作为目标的运动模型,通过多项式重采样方法进行粒子重采样,以保持粒子的多样性;设计了正负模板更新策略,将正模板分为固定集和更新集,对这两部分在相似度计算和正模板更新时采取不同的处理方法,并且在其中加入目标遮挡的判决机制,从而可以有效避免遮挡的影响;实验结果表明,该算法能够准确跟踪受遮挡、运动模糊等多种复杂场景的目标,与现有跟踪方法相比,所提算法具有更好的准确性和稳定性。

A robust object tracking method is proposed to deal with technical issues during tracking. Firstly, the global template based on sparse representation is used to describe object appearance, while positive and negative modules are built to separate the object from the background. Then, Random Projection（RP） is used to reduce the dimension of modules and candidate objects, which could release the calculation burden. Furthermore, the Particle Filter（PF） is used as the object motion model, and the multi-normal resample method is used to maintain the diversity of particles. To alleviate module drift problem, the positive module is divided into static module and changeable module, while different modules are dealt with different ways, and sparse reconstruction error is used to determine whether the object is occluded. Experiment results on numerous challenging videos show that the proposed method has better performance in accuracy and stability in comparison with state-of-the-art tracking methods.