多模型融合验证的视觉跟踪算法

Multi-model Integrated and Verified for Visual Tracking

为提升孪生网络和学习判别模型跟踪算法的精度和鲁棒性,提出了一种多模型融合验证的视觉跟踪方法。首先用置信反馈调节评估目标搜索区域特征与分类器之间的相似性响应图,反映目标定位的可靠度,然后结合孪生网络生成一系列候选对象并估计其相似性得分,最后验证网络评估候选框,并输出最佳候选框。在VOT2018、VOT2019、OTB100三种数据集上进行测试,所提算法在快速运动、严重遮挡等干扰下仍能较好地跟踪目标。

In order to improve both the accuracy and robustness of the siamese network and the learning discriminative model tracking algorithm,a multi-model fusion verification visual tracking method is proposed.Firstly,the average peak-to-correlation energy and peak-to-sidelobe ratio are used to determine the similarity response map between the target search area feature and the classifier,which reflects the reliability of target position,and then combines the siamese network to generate a series of candidates and estimate their similarity scores effectively.Finally,the verification network evaluates these candidates and outputs the optimal one as the tracked object.The proposed algorithm is evaluated on the standard datasets VOT2018,VOT2019,and OTB,which can track the object stably under the interference of fast motion and severe occlusion and so on.