中文摘要：

提出了一种基于频域核回归模型的尺度目标跟踪算法。该算法将时域相关滤波转换为频域的线性回归问题,构建一个包含似然项和先验项的代价损失函数,当似然项和先验项选取不同的分布函数,获得不同约束条件下的跟踪模型;通过核函数用来适应目标在跟踪过程的尺度变化。在初始帧中,通过手动标注目标初始状态,获得目标样本及其标记,利用核回归得到目标的频域模板;在跟踪过程中,利用循环卷积定理,将时域相关运算转换为频域乘积运算,快速计算候选样本的响应,得到目标在当前帧的估计;利用估计结果更新目标频域模板,同时在线更新核函数,适应目标尺度变化。实验分析表明,文中算法能够实时地跟踪目标,适应目标外观和尺度的变化,获得较好的跟踪效果。

英文摘要：

An algorithm of tracking scalable object is proposed based on frequency kernel regression. The algorithm converts the correlated filter in time domain into linear regression in frequency domain, and constitutes a cost function including likelihood and prior. A different distribution function can be chosen by the likelihood and prior to obtain the different models. The kernel function is used to adapt the scalable object in the tracking processing. The object states and labels are obtained in the initial frame manually, and the object frequency template is got by the kernel ridge regression model. The circulated convolution theorem is used to compute the responds in the tracking processing. The tracking model is updated after the estimated states in the current frame. The results show that the proposed algorithm can track the object real time, and adapt to the object＇s changes. Compared with the state-of-the-art algorithms, the proposed algo- rithm can obtain a good result when the object has large changes, even the object scale changes.