【摘要】 目标跟踪是计算机视觉领域中具有挑战性的问题.提出了一种基于稀疏表示的判别式目标跟踪算法,用于在复杂场景中对运动目标进行鲁棒跟踪.该算法首先对目标进行滑动窗口稠密采样,构建目标的稀疏表示字典,然后将目标表示为该字典的稀疏编码,从而构造具有判别力的目标特征表示.在跟踪过程中,将目标跟踪问题看作是背景与目标的判别性问题,使用目标和背景的特征表示在线训练朴素贝叶斯分类器,根据分类结果得到目标的跟踪结果.为了适应场景及目标外观变化,设计动态更新机制对字典与分类器进行在线更新.和传统基于稀疏表示的跟踪方法相比,该算法将稀疏表示与判别式分类器结合,利用稀疏表示获得具有判别力的目标特征表示,而在线的朴素贝叶斯分类器则确保了目标跟踪的快速有效.与流行的多种跟踪算法比较结果表明,本算法能够在复杂条件下实现目标的鲁棒跟踪.更多还原

【Abstract】 Visual tracking in complex backgrounds is a challenging problem in computer vision.In this paper,we propose a novel tracking method based on discriminative classifier with sparse representation.First,local image patches within an object region are sampled by using sliding windows and spare representation dictionary of object is constructed.Then the tracking object is represented by the sparse coding of this dictionary,and forms a feature vector.Therefore,the tracking problem is formulated as a binary classification problem to separate the target object from the background.Features of object and background are used to train an online nave Bayesian classifier.The tracking result is formed based on the classifier output.Sparse representation and discriminative classifier are combined in this algorithm.The discriminative feature representation is obtained by the sparse representation.On the other hand,the speed and effectiveness of object tracking are guaranteed by using online nave Bayesian classifier.Experiments have shown that our algorithm can track objects in complex background and outperform the state-of-the-art algorithms in terms of robustness and effectiveness.更多还原