融合分层卷积特征和尺度自适应核相关滤波器的目标跟踪

Object Tracking Based on Hierarchical Convolution Feature and Scale Adaptive Kernel Correlation Filter

尽管经过多年的研究,尺度变化、形状变化、严重的遮挡、背景干扰、光照变化和相机运动等内外因素引起的目标表观变化,使得目标跟踪仍然是一个极具挑战的问题.为了有效地处理目标表观变化,基于分层卷积特征和尺度自适应核相关滤波器的目标跟踪算法,将目标跟踪分解为目标位置的预测和尺度的估计两个步骤.在目标位置估计方面,区别于传统的基于手工设计特征的目标跟踪算法,我们使用基于分层卷积特征的相关滤波器算法计算出不同卷积层上的跟踪结果置信图,对各个层上得到的结果进行加权求和得到目标置信图,估计出目标的最终位置.在目标的尺度估计方面,为了有效捕捉目标尺度变化,我们首先使用尺度金字塔对下一帧适用的尺度进行预测,同时对目标尺度进行更新.在标准测试集(OTB-50)上的实验结果表明,本文所提出的融合分层卷积特征和尺度自适应的相关滤波器的目标跟踪算法取得较好的精度和鲁棒性.

Despiteyearsofstudy,scale change,shape change,serious changes in object appearance caused by occlusion,background interference,illumination changes and camera motion and other internal and external factors,the object tracking is still a challenging problem.In order to effectively deal with values of the target table,the hierarchical convolution features and scale adaptive kernel correlation filter based target tracking algorithm for estimation and prediction of target tracking decomposition scale of target position in two steps.In the estimation of the target location,different from the traditional manual design features of the target tracking algorithm based on the calculated results we use confidence map different convolution layer on the tracking filter algorithm based on hierarchical convolution characteristics of each layer,the results of the weighted sum of the estimated target confidence map,the final position of the target.In terms of the scale of the target,in order to effectively capture the change of the target scale,we first use the scale of Pyramid on the next frame of the scale to predict,while the target scale to be updated.The experimental results on standard test set （OTB-50） show that the proposed algorithm based on hierarchical convolution feature and scale adaptive correlation filter has better accuracy and robustness.