基于孪生网络特征融合与阈值更新的跟踪算法

Tracking Algorithm Based on Twin Networks Feature Fusion and Threshold Update

针对常见的视频目标跟踪过程中由于背景相似干扰、遮挡和形变等一系列问题导致跟踪精度下降的问题,该文提出一种基于孪生网络特征融合与阈值更新的跟踪算法,并引入加权平均模块。该算法采用MobileNetV2作为特征提取网络,借助深度可分离卷积运算,减少模型的参数量,并由深至浅进行多层特征融合,使特征同时具备浅层与深层的信息;模板分支的特征送入注意力机制模块,提高对目标关键特征的关注度;将融合并通过注意力机制后的特征向量的互相关结果与原始提取的特征向量的互相关结果再进行加权平均获得响应图;响应最高位置的相似度与设定阈值比较,判断是否进行模板更新,进一步提高跟踪精度。实验结果表明,该算法在OTB100数据集上能够获得较好的跟踪效果,与现有的一些算法相比展现了更强的鲁棒性。

Aiming at the problem that the tracking accuracy is reduced due to a series of problems such as background similar interference,occlusion and deformation in common video target tracking process,we propose a tracking algorithm based on twin network feature fusion and threshold update and introduce a weighted average module.In this algorithm,MobileNetV2 is used as the feature extraction network.With the help of deep separable convolution operation,the number of model parameters is reduced,and multi-layer feature fusion is performed from deep to shallow,so that features have shallow and deep information at the same time.The features of the template branch are sent into the attention mechanism module to improve the attention to the key features of the target.The response graph is obtained by weighted average of the cross-correlation results of the fused and extracted feature vectors and the original extracted feature vectors through the attention mechanism.The similarity of the highest response position is compared with the set threshold value to determine whether to update the template and further improve the tracking accuracy.Experiment shows that the proposed algorithm can achieve better tracking performance on OTC100 data set,and shows stronger robustness compared with some existing algorithms.