基于空间加权对数似然比相关滤波与Deep Snake的目标轮廓跟踪

Object Contour Tracking Based on Correlation Filters with Spatially-Weighted Logarithm Likelihood Ratio and Deep Snake

近年来,目标跟踪中目标的状态表示已由粗糙的矩形框转化为精细的目标掩膜.然而,现有方法利用区域分割得到目标掩膜,速度慢并且掩膜精度受限于目标跟踪框.针对以上问题,本文提出基于空间加权对数似然比相关滤波与Deep Snake的目标轮廓跟踪方法 .该方法包括三个阶段:在第一阶段,利用提出的空间加权对数似然比相关滤波器估计目标的初始矩形框;在第二阶段,通过Deep Snake将初始矩形框变形为目标轮廓;在第三阶段,根据目标轮廓拟合出跟踪结果 .对提出的方法在OTB(Object Tracking Benchmark)-2015和VOT(Visual Object Tracking)-2018数据集上进行了实验验证,结果表明:与现有先进的目标跟踪方法相比,本文提出的跟踪方法具有较优的性能.

Recently, the state representation of the target in object tracking has been transformed from the coarse bounding-box to fine-grained segmentation map. However, the existing methods use pixel-based segmentation to obtain object mask, which is slow and the accuracy of mask is limited by the object bounding box of tracking. To solve the above problems, we propose an object contour tracking method based on correlation filters with spatially-weighted logarithm likelihood ratio and deep snake. The method consists of three stages: at the first stage, the initial bounding box of the object is estimated by the proposed correlation filters with spatially-weighted logarithm likelihood ratio;at the second stage, the initial bounding box is deformed into the object contour via deep snake;at the third stage, the tracking results are fitted with the object contour. Experimental results on OTB(Object Tracking Benchmark)-2015 and VOT(Visual Object Tracking)-2018 datasets show that the proposed method is superior to the state-of-the-art approaches.