基于交叉通道注意力的目标跟踪方法

Target Tracking Method Based on Cross-channel Attention

针对跟踪目标被遮挡或者目标周围存在敏感干扰物,从而导致前景背景分类错误和边界框预测错误的问题,提出了一种基于交叉通道注意力的无锚框的目标跟踪方法。首先使用交叉通道注意力对特征提取部分的最后三层输出进行通道增强,利用模板特征和搜索特征中目标的相似性,整合所有通道特征的相关性,从而选择性的对目标特征的通道的增强。之后使用加权求和的方式进行特征融合,使用浅层特征和深层特征融合提高分类精度和定位的准确度。最后使用位置注意力对分类特征图进行全局编码,再次增强分类特征图的特征,提高网络对目标的定位准确性。实验结果表明,提出的算法在OTB100数据集上取得了85.5%的准确率和64.1%的成功率,在UAV20L数据集上取得了70.5%的准确率和56.0%的成功率。

Aiming at the problem that the tracking target is blocked or there are sensitive interferences around the target,which leads to the classification error of foreground background and the prediction error of boundary box,an anchor frame-free target tracking method based on cross-channel attention is proposed.Firstly,cross-channel attention is used to enhance the last three layers of output of feature extraction,and the correlation of all channel features is integrated by using the similarity of target in template features and search features,so as to selectively enhance the channel of target features.After that,feature fusion is carried out by weighted summation,and shallow feature fusion and deep feature fusion are used to improve classification accuracy and location accuracy.Finally,location attention is used to enhance the features of the classification feature map again to improve the accuracy of target location.Experimental results show that the proposed algorithm achieves 85.5%precision rate and 64.1%success rate on OTB100 dataset,and 70.5%precision rate and 56.0%success rate on UAV20L dataset.