基于注意力与特征融合的双分支跟踪算法

Dual-branch tracking algorithm based on attention and feature fusion

针对全卷积孪生网络(SiamFC)特征表达不足导致在目标形变、复杂背景干扰条件下跟踪效果不佳等问题,提出了一种基于注意力与特征融合的双分支跟踪算法。通过引入通道注意力机制和小卷积核的思想,在增加网络深度的同时,动态调节模型权重,提高了网络的特征提取能力和辨别能力。在原有孪生网络的基础上,提出采用多层特征融合策略构建新的跟踪分支并用于辅助决策。通过不同视觉层级间目标特征的融合,进一步改善网络的精确定位与适应能力。在目标跟踪基准库OTB50、OTB100、VOT2017进行试验,试验结果表明:所提算法在保持实时性的情况下,性能指标优于多个基准算法,验证所提算法的有效性。

Aiming at the problem of insufficient feature expression of the fully convolutional siamese network(SiamFC),which leads to poor tracking performance in the conditions of target deformation and complex background interference,a dual-branch tracking algorithm based on attention and feature fusion is proposed.By introducing the idea of a channel attention mechanism and a small convolution kernel,while increasing the depth of the network,the weight of the model is dynamically adjusted to improve the feature extraction ability and discrimination ability of the network.On the basis of the original siamese network,a multi-layer feature fusion strategy is proposed to construct a new tracking branch and use to assist decision-making.Through the fusion of target features between different visual levels,the precise positioning and adaptability of the network are further improved.Tests are conducted on the target tracking benchmark libraries OTB50,OTB100,and VOT2017.The test results show that the proposed algorithm has better performance indicators than multiple benchmark algorithms while maintaining real-time performance,verifying the effectiveness of the proposed algorithm.