基于多特征和尺度估计的核相关滤波跟踪算法

Kernel correlation filter tracking algorithmbased on feature fusion and scale estimation

针对传统核相关滤波器跟踪方法(KCF)在尺度估计不足和抗遮挡性低等问题上,本文提出了一种把梯度直方图和颜色直方图相结合,并利用尺度估计策略提升跟踪框适应性的核相关滤波跟踪算法.该方法首先通过建立核岭回归模型,使用二维核化相关位置滤波器,融合方向梯度直方图(HOG)特征和颜色直方图(CN)特征,采取根据响应大小的方式加权融合跟踪坐标,精确确定目标的中心位置;然后,利用滤波响应的峰值旁瓣比的高低来判定是否发生遮挡,当特征响应的旁瓣比低于设定的阈值时,暂停更新滤波模板;最后,利用光流法计算出视频帧间关键角点的位移来估计被跟踪目标的形变比例和尺寸,同时结合尺度集合进行跟踪框缩放.通过理论分析和在跟踪基准数据库OTB-2013中的50组视频序列进行仿真实验,对比了当下主流的相关滤波跟踪算法,在保证实时性的同时,较原核相关滤波算法跟踪的精度提升了14.5%,成功率提高了9.2%,并且在复杂场景下具备较强的抗遮挡性和鲁棒性.

Aiming at the problems that traditional kernel correlation filter tracking method(KCF)lacks scale estimation and cannot accurately track occluded objects,this paper proposes a kernel correlation filter tracking algorithm combining multiple features and scale estimation.This method first establishes a kernel ridge regression model,constructs a two-dimensional kernelization related position filter,fuses the directional gradient histogram(HOG)feature and the color histogram(CN)feature,weights the fusion tracking coordinates according to the response size,and accurately determines The center position of the target;then,the peak sidelobe ratio of the filter response is used to determine whether occlusion occurs.When the sidelobe ratio of the characteristic response is lower than the set threshold,the filter template is temporarily updated;finally,the video is calculated using the optical flow method The displacement of key corner points between frames is used to estimate the deformation ratio and size of the tracked target,and the tracking frame is scaled by combining the scale set.Through theoretical analysis and simulation experiments on 50 sets of video sequences in the tracking reference database OTB-2013,the current mainstream related filtering and tracking algorithms are compared.While ensuring real-time performance,the tracking accuracy is improved by 14.5%compared to the original kernel related filtering algorithm.The success rate is increased by 9.2%,and it has strong anti-occlusion and robustness.