论文部分内容阅读
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.
Despite many years of research, target tracking is still a challenging issue due to the apparent changes in the target caused by internal and external factors such as scale changes, shape changes, severe occlusions, background disturbances, light changes and camera motions. In order to effectively handle Target apparent change, target tracking algorithm based on layered convolution feature and scale-adaptive kernel-related filter, the target tracking is decomposed into two steps: the prediction of target location and the estimation of scale.In the aspect of target location estimation, Based on the hand-designed feature-based target tracking algorithm, we use the correlation filter algorithm based on the hierarchical convolution feature to calculate the confidence map of the tracking results on different convolutional layers, and weighted the sum of the results obtained at each layer to obtain the target confidence In order to effectively capture the target scale change, we first use the scale pyramid to predict the scale applicable to the next frame, and update the target scale at the same time.In the standard test set ( OTB-50), the experimental results show that the proposed fusion stratified convolution feature and scale from The filter should be relevant target tracking algorithm to achieve better accuracy and robustness.