在线多示例学习目标跟踪方法研究

Research on object tracking method based on online multiple instance learning

多示例学习是不同于传统机器学习的一种新的学习模式,近年来被应用于图像检索、文本分类等领域。提出一种基于在线学习的多示例学习算法,将其应用于目标跟踪。该算法通过构造一个在线学习的多示例分类器作为检测器,无需制作大量的样本进行离线的训练,只需在第一帧手动选中目标,便可以自动生成正样本和负样本,并在随后的帧序列中,根据跟踪到的目标自动更新分类器,在跟踪器丢失目标或者目标从场景中消失后,它能够重新检测到目标并更新跟踪器,从而有效地支持了跟踪器跟踪目标。实验证明该方法在背景复杂,光线变化,摄像机抖动等复杂条件下,可以很好地跟踪到目标,且对遮挡具有较好的鲁棒性。

Multiple instance learning is a new mode of learning, which is different from the traditional machine learning. It has been used in the field of image retrieval, text classification in recent years. This paper proposes an online multi-instance learning algorithm, and applies it to the object tracking. The algorithm by constructing an online learning multi-instance classifier as the detector, without making a large number of samples for off-line training, and only needs to manually select the object in the first frame. It can automatically generate the positive samples and negative samples, and in the subse-quent sequence of frames, automatically update the classifier based on tracking target. If the tracker misses the target or the target disappears from the scene, it can re-detect target and update the tracker, so it can effectively support the tracker to track the object. Experiments show that the proposed approach can track the target well in the complex background, the light change, camera jitter and some other complex conditions, and it has a better robustness for the occlusion.