基于多特征融合的核相关滤波目标跟踪算法

Kernelized correlation filter for target tracking with multi-feature fusion

为提高复杂背景下目标跟踪的精度和鲁棒性,提出一种多特征融合的核相关滤波目标跟踪算法。在海量无关图片训练集上得到深度模型,使用目标跟踪视频序列对其进行微调;从模型中提取多层深度线性特征并进行插值运算,同时提取图像序列的方向梯度直方图特征和颜色名特征;将得到的特征图通过核相关滤波计算相应的相关响应图;融合各个响应图,寻找最大响应值以确定目标位置。在OTB和VOT标准数据集上的实验结果表明,该模型具有较高的识别精度,能够在背景斑杂、光照变化、目标遮挡和变形等复杂环境下长期稳定地跟踪目标。

To improve the accuracy and robustness of target tracking task in complex background,a target tracking algorithm based on multi-feature fusion and kernelized correlation filters was proposed.A deep model was obtained based on a large number of unrelated images,and it was fine-tuned using tracking sequences.Hierarchical linear features were extracted from the model and were processed by interpolation operation,and histogrammic features of oriented gradient and color name were extracted.Correlation response maps were calculated using kernelized correlation filters.Each response map was fused to find the maximum response value to determine the target location.Experimental results on object tracking benchmark and visual object tracking datasets show that the proposed algorithm has higher tracking precision.In addition,the target can be tracked stably in complex environments such as cluttered background,illumination change,occlusion and deformation.