适用于目标跟踪的加权增量子空间学习算法

为了克服目标物外观变化给跟踪造成的困难,提出一种基于加权增量子空间学习的目标跟踪算法.该算法构造了一个可在线更新的子空间作为视频中目标物的外观模型,根据概率转移模型预测得到一组图像样本作为目标物在当前帧中可能出现的图像区域；然后将图像样本投影到该低维子空间中估计每个图像样本为目标图像区域的似然度,以具有最高似然度的样本作为目标在当前帧中的图像区域，通过加权增量的方式调整子空间.实验结果表明：相比基于其他增量子空间学习的跟踪算法,该算法能够稳定、准确地对运动目标进行跟踪.

Weighted incremental subspace learning algorithm suitable for object tracking

An object tracking method based on weighted incremental subspace learning was proposed to overcome the difficulty in object tracking resulting from variations in appearance. The method constructs a subspace updated online to depict the appearance of the object in the video. A set of image patches are predicted based on probabilistic transformation model as candidate image regions of the object in the current frame, then these image patches are projected onto the low-dimensional subspace, and the likelihood of each image patch as the image region of the object is evaluated. The image patch with maximal likelihood is regarded as the object image region. Finally, the subspace is updated incrementally with temporal weights. Experimental results show that the method accomplishes object tracking more steadily and accurately, compared with other incremental subspace learning based object tracking methods.