基于启发式候选区域推荐的目标孪生跟踪

Heuristic Candidate Region Proposal for Siamese Visual Tracking

在基于孪生网络的目标跟踪方法中,候选目标区域推荐的质量十分关键.目前的方法普遍采用锚点固定的推荐方式,该方式所生成的候选区域通常会数量庞大且质量不高.为此,本文提出一种基于概率图的启发式候选区域推荐的孪生跟踪模型.该网络模型由模板分支和检测分支构成,其中模板分支用于提取第一帧或前一帧目标图片的特征,而检测分支用于提取当前帧图片的特征.在提取当前帧特征后,采用概率图的方法来预测目标可能出现的位置,并利用可变锚点方式在该位置周边生成候选区域;之后推荐区域的特征再经过一个特征适应网络,与目标模板进一步的适应;最后,将模板分支与检测分支提取的特征进行互相关操作来完成目标跟踪.在VOT2016和OTB2015数据集上的测试结果表明,所提出的方法在候选目标推荐质量和跟踪精度方面都表现优异.

The proposal quality of candidate target region plays an important role in target tracking base on siamese network.Exciting trackers tend to use proposal method with fixed anchors,and the number and quality of candidate regions generated by this method are usually large and low respectively.To solve the problem,this paper proposes a tracking model using heuristics candidate region proposal base on probabilistic graph.The proposed tracking model consists of template branch and detection branch,where the template branch is used to extract the features of the first or previous target image,and the detection branch is used to extract the features of the current frame image.After extracting the features of current frame,the probability graph method is used to predict the possible position of the target,and some candidate regions are generated around the position using variable anchors.Then,the feature of each candidate region is further adapted to the target feature through a feature adaptation network.Finally,the feature from the template branch is correlated to the feature from the detection branch,so as to find the target.The experimental results on VOT2016 and OTB2015 dataset show that the proposed method performs well in anchor quality and tracking accuracy.