基于高效多示例学习的目标跟踪

Object tracking with efficient multiple instance learning

基于多示例学习(MIL)的跟踪算法能在很大程度上缓解漂移问题。然而,该算法的运行效率相对较低,精度也有待提高,这是由于MIL算法采用的强分类器更新策略效率不高,以及分类器更新速度与目标外观变化速度不一致引起的。为此提出一种新的强分类器更新策略,以大幅提升MIL算法的运行效率;同时提出一种动态更新分类器学习率的机制,使更新后的分类器更符合目标的外观,提高跟踪算法的精度。通过实验将该算法和MIL算法以及基于加权多示例学习的跟踪算法(WMIL)进行对比,实验结果表明,所提出算法的运行效率和跟踪精度都是三者中最好的,在背景中没有与被跟踪目标外观相似的干扰物体存在时有较好的跟踪优势。

The method based on Multiple Instance Learning （MIL） can alleviate the drift problem to a certain extend. However, MIL method has relatively poor performance in running efficiency and accuracy, because the update strategy efficiency of the strong classifiers is low, and the update speed of the classifiers is not same with the appearance change speed of the targets. To solve this problem, a new update strategy for strong classifier was proposed to improve the running efficiency of MIL method. In addition, to improve the tracking accuracy of the MIL method, a new dynamic mechanisim for learning rate renewal of the classifier was given to make the updated classifier would more conform to the appearance of the target. The experimental results on comparison with MIL method and the Weighted Multiple Instance Learning （WMIL） method show that, the proposed method has the best performance in running efficiency and accuracy among the three methods, and has an advantage over tracking when there is no similar interference objects to target objects in background.