传统的时空上下文跟踪算法在更新目标模型时不考虑跟踪结果的有效性, 故目标被长时间遮挡后, 目标模型容易被错误更新且难以修正。因此, 本文提出了一种基于双目标模型的改进时空上下文跟踪算法以解决错误更新问题。该算法引入一个辅助目标判别模型来评估时空上下文算法跟踪结果的有效性, 并根据评估结果对目标模型进行更新。辅助模型使用目标的局部纹理信息而不是相关性信息作为特征, 在目标被长时间遮挡后也能准确评估更新内容的有效性, 并能在遮挡结束后修正错误更新的目标模型。在多组数据集上的实验表明, 改进算法在测试数据集上的跟踪成功率为82%, 中心偏差为8 pixels; 在长时间遮挡等干扰情况下的跟踪精度比原时空上下文算法有明显提升, 实现了目标的可靠跟踪。

The original spatio-temporal context (STC) tracking algorithm does not take the tracking results into account when it updates an object model, so the object model is wrongly updated and hard to be recovered after long term occlusion. To solve this problem, an improved spatio-temporal context tracking algorithm based on a dual-object model is proposed in this paper. In this method, an auxiliary object model is introduced to evaluate the effectiveness of original STC algorithm, and to update the object model based on the tracking results accordingly. By using texture information rather than correlation information as the characteristics, the auxiliary object model evaluates exactly the updated contents and corrects wrong updates after long-term occlusion. Experimental results on several groups dada sets indicate that the success rate of the proposed algorithm is 82%, and the center location error is 8 pixels. It implements the stable tracking and the tracking precision is superior to that of the original STC algorithm in complex scenes, especially after long-term occlusions.