基于距离交并比回归的孪生网络目标跟踪算法

A target tracking algorithm for siamese network based distance intersection over union(DIOU)regression

为解决视频目标跟踪过程中目标框和预测框边界不重叠情况下无法优化的问题,提出了一种距离交并比(distance intersection over union,DIOU)回归的孪生网络目标跟踪算法.通过孪生网络和区域建议网络(region proposal network,RPN)保持实时效果,将距离交并比引入回归分支,结合重叠率和中心点距离构建损失度量,加快模型训练的收敛速度,为边界框提供更直接的回归方向.同时,使用Resnet50网络替换SiamRPN网络的特征提取网络,进一步提高目标特征表示的准确性.实验结果表明,DIOU回归损失在视频目标尺度变化、低分辨率、光照变化等干扰情况下,具有较强鲁棒性.

In order to solve the problem that the target bounding box and predicted bounding cannot be optimized when the boundary of them do not overlap in the process of visual target tracking,a siamese network target tracking algorithm based on distance intersection over union(DIOU)regression is proposed.The siamese network and regional proposal network(RPN)are used to maintain the real-time effect.The distance intersection over union is introduced into the regression branch,and the loss measurement is constructed in combination with the overlap rate and the center point distance,so as to provide a more direct regression direction for the bounding box and speed up the convergence of the model.At the same time,Resnet50 network is used to replace the feature extraction network in SiamRPN network to further improve the accuracy of target feature representation.Experimental results show that the algorithm has strong robustness under the interference of video target scale change,low resolution and illumination change.