基于动态位置编码和注意力增强的目标跟踪算法

Target tracking algorithm based on dynamic position encoding and attention enhancement

为了充分利用模板和搜索区域之间的位置信息以及提高融合特征的表征能力,提出使用动态位置编码和多域注意力特征增强的方法.在注意力模块内部嵌入带有卷积操作的位置编码模块,随注意力计算更新位置编码,提高自身空间结构信息的利用率.引入多域注意力增强模块,在空间维度上使用不同空洞率和步长的平行卷积进行采样,以应对不同大小的目标物,并聚合通道注意力增强后的特征.在解码器中加入空间域注意力增强模块,为预测头提供更精确的分类回归特征.本算法在GOT-10K数据集上的平均重叠度(AO)为73.9%;在TrackingNet、UAV123和OTB100数据集上分别取得了82.7%、69.3%和70.9%的曲线下面积(AUC).与主流算法的对比结果表明,融合了动态位置编码和通道、空间注意力增强的跟踪模型可以有效提升模板和搜索区域间的信息交互,提高跟踪的精度.

A method based on dynamic position encoding and multi-domain attention feature enhancement was proposed to fully exploit the positional information between the template and search region and harness the feature representation capabilities.Firstly,a position encoding module with convolutional operations was embedded within the attention module.Position encoding was updated with attention calculations to enhance the utilization of spatial structural information.Next,a multi-domain attention enhancement module was introduced.Sampling was conducted in the spatial dimension using parallel convolutions with different dilation rates and strides to cope with targets of different sizes and aggregate the enhanced channel attention features.Finally,a spatial domain attention enhancement module was incorporated into the decoder to provide accurate classification and regression features for the prediction head.The proposed algorithm achieved an average overlap(AO)of 73.9%on the GOT-10K dataset.It attained area under the curve(AUC)scores of 82.7%,69.3%,and 70.9%on the TrackingNet,UAV123,and OTB100 datasets,respectively.Comparative results with state-of-the-art algorithms demonstrated that the tracking model,which integrated dynamic position encoding as well as channel and spatial attention enhancement,effectively enhanced the interaction of information between the template and search region,leading to improved tracking accuracy.