基于视觉字典的在线多示例目标跟踪

Visual dictionary and online multi-instance learning based object tracking

在线多示例目标跟踪算法无法判别目标丢失以及无法适应目标尺度的变化。提出了一种基于视觉字典的在线多示例目标跟踪算法。算法将视觉字典和多示例跟踪分别作为检测器和跟踪器,利用互反馈技术提高跟踪性能。跟踪器完成目标的跟踪并为视觉字典的构建和更新提供训练样本;检测器则对跟踪器的结果(候选样本)进行判定,目标丢失时,暂停跟踪并重新检测目标,目标未丢失时,利用Ransac算法获得目标的尺度变换系数并在新尺度下更新跟踪器。为了提高目标丢失判别的准确性,提出了一种局部随机抽样的直方图相似性度量技术,采用局部划分思想和Noisy-NR模型计算候选样本与训练样本特征直方图的相似性,减少了传统直方图匹配由于受目标局部遮挡影响造成的误判。实验结果表明,该算法能够适应目标的尺度变化,检测目标的丢失,提高了跟踪稳定性。

A novel object tracking algorithm fused with the visual dictionary and online multiple instance learning tracking （MILTrack） is proposed to solve the problem of tracking failure detection and scale changes in MILTraek algorithm. It regards the visual dictionary and MILTrack as detector and tracker respectively. Mutu- al feedback technology is employed for improving the tracking performance. The dictionary is constructed and updated by the training sample obtained from the tracker, while the detector make decision whether the object is lost or tracked. If we lost the object, a detection is implemented in a larger area. Otherwise, Ransac algorithm is utilized to obtain the scaling factors of the target, under which the tracker is updated. In order to improve the accuracy of the loss decision of the target, we propose a local random sampling of histogram similarity measure technique. The idea of local division and Noisy-NR model is employed for the measurement of similarity between the histograms of candidate sample and training target samples. The results shows that our algorithm makes the MILTrack algorithm adaptively adjust the scale of the target, and the detection of tracking failure is possible. The stability of tracking is improved.