结合多注意力机制的自监督目标跟踪

Self-supervised object tracking based on multi-attention mechanism

为解决现有目标跟踪数据集不足及手工标注数据成本过大的问题,提出结合判别式相关滤波及多注意力机制的自监督目标跟踪方法。训练集选用原始未标记的视频图像,使用子空间注意力机制及通道注意力机制针对不同的输入目标对特征进行自适应调整,构建相关滤波输出响应图进行跟踪定位;通过前向跟踪和后向跟踪两个跟踪过程,以自监督的训练方式用最终响应结果与初始标签构建循环一致性损失。在OTB50和OTB100两个公开数据集的实验结果表明了所提方法的实时性和有效性。

To solve the problem of insufficient target tracking data set and excessive manual labeling data cost,a self-supervised target tracking method combining discriminant correlation filter and multi-attention mechanism was proposed.The original unlabeled video image was used as the training set,subspace attention mechanism and channel attention mechanism were used to adaptively adjust features for different tracking targets,and correlation filters were constructed at the end to output the response map for tracking and positioning.Through two tracking processes,forward tracking and backward tracking,the self-supervised training method used the final response result and the initial label to construct the cycle consistency loss.Experimental results on two public data sets OTB50 and OTB100 show the real-time and effectiveness of the proposed method.