基于可变形卷积的孪生网络目标跟踪算法

为解决多数基于孪生网络的跟踪算法存在骨干网络特征提取能力弱、模板不适应目标变化等问题,在SiamFC的基础上提出基于可变形卷积的孪生网络算法(DCSiam).首先,采用可变形卷积模块在不同方向上学习多层特征数据的自适应偏移量,增大卷积过程中的有效感受野;然后,通过多层可变形互相关融合得到最终响应图,以增强骨干网络的深层语义特征提取能力;最后,采用一种高置信度的模板在线更新策略,每隔固定帧计算响应图的峰值旁瓣比与最大值作为更新依据,使用加权的方式融合特征以更新模板.使用OTB2013、 OTB2015、VOT2016和VOT2017四个公共基准数据集对所提出算法进行跟踪性能评估,实验结果表明,在OTB2015数据集上, DCSiam算法整体精确率、成功率较基线分别提高9.5%和7.5%,很好地实现了复杂情况下的目标跟踪,验证了所提出算法的有效性.

Target tracking based on deformable convolution siamese network

In order to solve the problem that most tracking algorithms based on siamese networks have weak feature extraction ability of backbone networks and template is not suitable for target transformation, this paper proposes the siamese network algorithm based on deformable convolution(DCSiam) on the basis of SiamFC. Firstly, the deformable convolution module is used to learn the adaptive offset of multi-layer feature data in different directions, and the effective receptive field in the convolution process is increased. The final response map is obtained by multi-layer deformable cross-correlation fusion to enhance the capability of deep semantic feature extraction of the backbone network. Finally, an online template updating strategy with high confidence is adopted, the peak sidelobe ratio and the maximum value of the response graph are calculated in every fixed frame as the basis for updating, and the features are fused in a weighted way to update the template. The tracking performance of the proposed algorithm is evaluated using four common benchmark datasets: OTB2013, OTB2015, VOT2016 and VOT2017. The experimental results show that the overall accuracy and success rate of the DCSiam algorithm are increased by 9.5% and 7.5% respectively compared with the baseline on the OTB2015 dataset,which well realizes the target tracking in complex situations and verifies the effectiveness of the proposed algorithm.