利用快速傅里叶变换的双层搜索目标跟踪算法

Two-level searching tracking algorithm based on fast Fourier transform

针对视觉跟踪中目标表观变化、尺度及旋转变化问题,提出了利用快速傅里叶变换的双层搜索目标跟踪算法.算法在直角坐标系和对数极坐标系分别构建目标核岭回归模型并进行目标双层搜索,同时利用快速傅里叶变换将时域运算转换到频域运算提高跟踪效率.首先在直角坐标系中建立目标核岭回归模型并构建循环结构矩阵进行穷搜索得到目标中心位置;然后以目标中心位置为原点将跟踪区域变换到对数极坐标系再次建立目标核岭回归模型,并穷搜索得到目标在对数极坐标系的平移量;最后依据搜索结果确定目标状态并进行模型更新.实验结果表明,这种算法不仅对表观变化、尺度及旋转变化具有较强的鲁棒性,而且跟踪实时性较好.

In order to solve the problems of appearance change, scale and rotation change in the visual tracking, a two-level searching tracking algorithm based on Fast Fourier Transform（FFT）is proposed. It achieves two-level searching by establishing the object＇s kernel ridge regression model in the Cartesian coordinates and log-polar coordinates, respectively, and the efficiency can be improved by transforming the operation into the frequency domain based on FFT. First, the kernel ridge regression model is constructed in the Cartesian coordinate and the object＇s center position is obtained by the exhaustive search method based on the circular structure matrix. Then, it transforms the object area to the log-polar coordinates and searches the shift using the kernel ridge regression model in the log-polar coordinates. Finally, the object＇s state is calculated according to the searching results and the object＇s model is updated. Experimental results indicate that the proposed algorithm not only can obtain a distinct improvement in coping with the appearance change, scale and rotation change, but also have a high tracking efficiency.