基于改进的MOSSE相关滤波的目标跟踪

Object Tracking Algorithm Based on Improved MOSSE Correlation Filter Algorithm

摘要:

视频目标跟踪存在如遮挡、形变、漂移等诸多挑战。虽然研究者提出了大量的算法来解决这一问题, 但大多数不具备普适性和实时性。为了实现目标有效稳定的实时跟踪, 本文在MOSSE相关滤波算法基础上提出了一种多模式的视觉目标跟踪算法, 该算法不仅具有相关算法的实时性, 还适应跟踪目标外观大幅度变化情况。同时, 为了适应跟踪过程中目标外形的复杂变化, 提出了一个控制模式更新率的算法, 利用具有多模式的跟踪算法可以同时处理极小的运动与外形突变。对基准视频数据的仿真实验结果表明, 与对应的单模型跟踪算法相比, 本文提出的算法可以明显改善跟踪精度和稳定性。

Abstract:

Object tracking algorithm has many challenges for stable and accurate performance, such as occlusion, noise, drifts and so on. Although many researches proposed many robust tracking algorithms to solve these problems, the solutions have not universality and real time. Additionally, in order to achieve robust tracking performance, a multiple model update strategy for visual tracking algorithm is proposed to adapt the target appearance changes. Moreover, for adapting complex appearance change, an update rate strategy has been proposed. The multiple models update strategy for visual tracking algorithm can process both infinitesimally small movements and abrupt changes simultaneously. Simulation experiments results have demonstrated that compared to its single model tracking algorithm, our proposed strategy can improves significantly the tracking stability and accuracy in benchmark datasets.