融合位置估计的自适应尺度相关滤波跟踪

Adaptive Scale Correlation Filtering Tracking Based on Fusion Position Estimation

针对相关滤波跟踪算法中不能适应目标多尺度变化的问题,提出了一种融合位置估计和尺度估计的自适应尺度相关滤波器.通过提取当前图像中不同尺度大小的目标模板,作为先验信息加到滤波器中学习,训练滤波器.对多尺度模板训练赋予新的权重定义,重新定义了多尺度模板对应的标签.提高了滤波器对目标尺度变化的敏感度.通过在CVPR 2015数据集验证,该方法的精准率为0.803、成功率为0.705.特别是在多尺度环境影响下,该算法结果优于其他算法,达到了预期跟踪效果.

This paper proposed an adaptive scale correlation filter combining multi-scale supervision of position estimation and scale estimation aiming at the problem that the correlation filtering and tracking algorithm can not adapt to the target scale variation.The information about the target templates of different scales extracted from the current frame picture is added to the filter as priori information for the filter to learn and train the filter as well.A new weight definition is given to the multi-scale template training,and the corresponding label of the multi-scale template is redefined.As a result,the sensitivity of the filter to the change of target scale is improved.Verified in CVPR 2015 dataset,the accuracy rate of the new method is 0.803 and the success rate is 0.705.Especially under the influence of multi-scale environment,the results of this algorithm are better than those of other algorithms,and the expected tracking effect can be achieved.