基于模板更新的深层孪生网络跟踪算法

针对跟踪过程中因尺度变化、遮挡及运动模糊等造成的目标定位不准确问题,在SiamFC(fully-convolutional siamese network)的跟踪框架基础上提出了一种具有高置信度模板更新机制的深层孪生网络目标跟踪算法.首先,主干网络采用ResNet-50残差网络进行特征提取,并融合多层特征图进行目标预测;其次,为避免模板频繁更新带来的模板漂移问题,构建了高置信度的模板更新模块.在OTB100数据集上的实验结果表明,相比基准算法,文中算法的跟踪成功率和精确度分别提高了3.4%和2.6%;在多种挑战因素下的对比实验表明,文中算法可以较好地抵抗目标遮挡、尺度变化、运动模糊等多种复杂因素带来的影响,有很好的鲁棒性.

A deep siamese network algorithm with template updating for object tracking

To solve the problem of inaccurate target location caused by scale variation, occlusion and motion blur during the tracking process, a deep Siamese network target tracking algorithm with high confidence template updating mechanism is proposed based on the SiamFC(fully convolutional siamese network). First, the main network uses ResNet-50 residual network for feature extraction and multi-layer feature maps for target prediction; second, a high confidence template updateing module is constructed to avoid the template drift caused by frequent updating. Experimental results indicate that the success rate and tracking accuracy of the proposed algorithm are increased by about 3.4% and 2.6% respectively compared to the benchmark algorithm when running on the dataset of OTB100. The experiments under various challenging factors show that the proposed algorithm has good robustness, which can resist the effects of various complex factors effectively such as target occlusion, scale variation and motion blur.