基于归一化注意力机制的特征自适应融合目标跟踪算法

Target Tracking Algorithm Based on Normalized Attention Mechanism with Feature Adaptive Fusion

针对快速运动目标跟踪时图像的形变和低分辨率等问题,基于当前的孪生网络,提出一种基于归一化注意力机制的特征自适应融合目标跟踪算法。首先,通过轻量级的注意力机制抑制不太明显的权重,对注意力模块施加权重稀疏惩罚,并对主干网络最后4个特征层进行路径增强;其次,为捕捉在线跟踪过程中目标的外观变化,提升算法鲁棒性,提出了一种插件式的模板在线更新方法;最后,利用回归增强分类的方法完成对目标的跟踪。实验结果表明:该算法在OTB100,UAV123两个挑战性数据集上分别取得了63.3%和59.5%的较高成功率;同时,在外界光照变化、图像背景复杂、目标平面内旋转时,算法具有较强的鲁棒性。

To solve the difficult problems of image deformation and low resolution in fast moving target tracking.A feature-adaptive fusion target tracking algorithm based on a normalised attention mechanism is proposed based on current siamese networks.Firstly,the less obvious weights are suppressed by a lightweight attention mechanism that imposes a weight sparsity penalty on the attention module.While a path enhancement method is proposed,path strengthening is applied to the last four feature layers of the backbone.Secondly,a plug-in template online update method is proposed in order to capture changes in the appearance of the target during online tracking and improve the robustness of the algorithm.Finally,a regression-enhanced classification method is used to complete the tracking of the target.The experimental results show that the proposed algorithm achieves high success rates of 63.3%and 59.5%on two challenging datasets,OTB100,UAV123,respectively.Also,the algorithm has strong robustness when the illumination changes,the image background is complex and the target is rotated in-plane.