改进特征和时间融合框架的区分式跟踪算法

Discriminative Tracker with Advanced Features and Temporal Fusion Framework

文章基于多示例学习的跟踪框架,引入改进型的分布场特征并融合目标时间维度信息,提出了一种新的跟踪算法。新的特征能够更为有效地描述目标的空间结构信息,对于目标模糊、局部遮挡以及细微形变有良好的鲁棒性。加入的目标时间维度信息融合方法,包含了目标的历史信息,同时也能响应目标的外观变化,提高了跟踪器从跟踪异常中恢复的能力。通过对比新算法与其他先进算法在多组测试视频上的跟踪结果,可以发现本文提出的算法具有更为优异的性能,能够在各种复杂情况下对目标进行稳定的跟踪。

A new tracker based on the multiple instance learning framework is proposed in this work, which introduces modified distribution fields features and a temporal fusion framework. The new distribution field features can describe the spatial information of the object more efficiently, and gain robustness to motion blur, minor occlusion and deformation of the object. Our temporal information fusion framework can contain the previous information of the object, and respond to the appearance variation of the object simultaneously, which improves the tracker＇s ability of recovering from tracking outliers. This new algorithm obtains a better performance on several test sequences compared with other state-of-the-art methods, and can track the object stably under various complicated situations.