为了解决背景嘈杂、遮挡、形变和尺度变化情况下目标跟踪问题, 提出利用视觉显著性和扰动模型的上下文感知跟踪。本文以相关滤波算法为基础, 将目标周围的上下文信息引入到分类器学习过程中, 构造了上下文感知相关跟踪, 提高了算法鲁棒性; 同时引入直方图扰动模型, 利用加权融合的方法获得目标响应图, 以此估计目标位置变化; 最后利用视觉显著性构建目标稀疏显著性图, 解决严重遮挡情况下的目标重定位问题, 并利用尺度估计策略解决目标尺度变化问题。利用公开数据集测试算法性能, 并与8种流行跟踪算法进行比较。实验结果表明, 本文算法的跟踪精确度得分和成功率得分分别为0.695和0.708, 均优于其它算法。与传统的相关滤波算法相比, 所提算法能很好地解决背景嘈杂、遮挡、形变和尺度变化等复杂下的目标跟踪问题, 具有一定理论研究价值和工程实用价值。

To solve the problem of target tracking in the presence of background noise, occlusion, deformation and scale variation, a context-aware tracking algorithm based on a visual saliency and perturbation model was proposed. First, the proposed algorithm was based on the correlation filtering algorithm. The contextual information of the target was introduced into the classifier learning process. The context-aware correlation filter was then constructed, which improves the robustness of the algorithm. Meanwhile, the histogram perturbation model was introduced. The target response map was calculated using the weighted fusion method to estimate the target position change. Finally, the target saliency map was constructed using visual saliency to solve the target relocation problem under occlusion problem. The scale estimation strategy was used to solve the problem of target scale variation. The algorithm performance was tested using open-source datasets and was compared with eight popular tracking algorithms. The experimental results demonstrate that the accuracy and success rate of the algorithm are 0.695 and 0.708, respectively, which are better than other algorithms. Compared with the traditional correlation filtering algorithm, the proposed algorithm can solve the target tracking problem with complex background noise, occlusion, deformation and scale changes. It has a certain theoretical research value and practical value of engineering.