基于并行多外观特征的孪生网络目标跟踪算法

Tracking algorithm of Siamese network based on parallel multiple appearance features

目标跟踪通常只能使用视频第1帧的外观信息,在线学习目标的外观特征,并预测后续帧中该目标的位置和大小.然而,跟踪过程中目标外观时刻变化,仅通过第1帧并不能准确描述后续目标的外观.针对上述问题,提出一种基于并行多外观特征的孪生网络目标跟踪算法.首先,引入包含目标近期外观信息的动态模板帧,同时提出3种方法:多外观特征、并行外观特征、并行多外观特征,利用动态模板帧进行目标跟踪.与简单地使用动态模板帧替换初始模板帧不同,所提出的方法可解决由动态模板帧中目标容易漂移导致的跟踪算法性能下降的问题.其次引入评价模块,使用基于信息熵的评价方法或基于IOU-Net的评价方法,对得到的多个预测结果分别进行打分,选择得分最高的预测结果作为最终的预测结果.最后提出更新模块,对评价模块得到的得分进行分析,当得分满足更新模块设立的更新条件时,用最终的预测结果更新动态模板帧,使用新的外观信息指导下一帧跟踪.实验结果显示,该算法在GOT-10k、OTB100等标准数据集上取得较好效果,验证了所提算法的有效性.

Target tracking usually only uses the appearance information of the first frame of the video to obtain the appearance characteristics of the target online,and predict the position and size of the target in subsequent frames.However,the appearance of the target changes all the time during the tracking process,and the appearance of subsequent targets is not be accurately described by the first frame alone.Focusing on the above problems,this paper proposes a Siamese network target tracking algorithm based on parallel multi-appearance features.First,dynamic template frames containing information about the recent appearance of the target are introduced.At the same time,three methods of multi-appearance,parallel appearance and parallel multi-appearance are proposed,which make dynamic template frames for target tracking.Second,either the evaluation strategy of information entropy or the evaluation method of the neural network in the evaluation module is applied to score the obtained multiple predictions separately,and the prediction result with the highest score is selected as the final prediction result.Finally,an update module is proposed to analyse the scores obtained from the evaluation module.If the score meets the update conditions set by the update module,the final prediction result is used to update the dynamic template frame,and the new appearance information is used to guide the tracking of the next frame.The experimental results show that the algorithm achieves good results on standard datasets such as GOT-10k,OTB100,which verif the effectiveness of the proposed algorithm.