基于多注意力融合的抗遮挡目标跟踪

Anti-occlusion Target Tracking Based on Multi-attention Fusion

针对基于孪生网络的目标跟踪算法在相似目标干扰和发生遮挡时容易丢失目标的问题,提出一种基于多注意力融合的抗遮挡目标跟踪算法(anti-occlusion target tracking based on multi-attention fusion, AOTMAF)。为更好地模拟遮挡图片,引入渐进式随机遮挡模块,由易到难地随机生成遮挡块对图像进行多区域遮挡,通过人工模拟被遮挡图像的方式扩充负样本数据集,提升模型在遮挡情况下对判别性特征的提取能力。从深度、高度与宽度三个维度挖掘特征图通道信息,并通过融合空间注意力,聚合特征图上每个位置的空间依赖性,增强特征表达能力,进一步提高跟踪的鲁棒性。实验结果表明,在OTB100、VOT2018、GOT-10K公开数据集上,本研究方法在复杂场景下能有效提升跟踪精度和鲁棒性。

Aiming at the problem that the target tracking algorithm based on Siamese networks is easy to lose targets when similar targets interfere and occlusion occurs,an anti-occlusion target tracking algorithm based on multi-attention fusion(AOTMAF)is proposed.In order to better simulate occluded images,a progressive random occlusion module is introduced to randomly generate occlusion blocks from easy to difficult to occlude the image in multiple areas,and the negative sample data set is expanded by manually simulating the occluded image,so as to improve the ability of the model to extract discriminant features under occlusion.The channel information of the feature map is mined from the three dimensions of depth,height and width,and the spatial dependence of each position on the feature map is aggregated by integrating spatial attention,so as to enhance the feature expression ability and further improve the robustness of tracking.The experimental results show that the proposed method can effectively improve tracking accuracy and robustness in complex scenarios on OTB100,VOT2018,GOT-10K public datasets.