基于背景感知与快速尺寸判别的相关滤波跟踪算法

背景感知相关滤波(Background-aware correlation filters, BACF)算法有效地解决了相关滤波类跟踪算法中的边界效应问题，提升了训练样本集的质量和数量，能够精确估计目标的位置变化，从而提高了跟踪器的性能。然而为了检测尺度变化，BACF算法通过多次重复计算不同尺度的目标区域，严重影响了跟踪速度。本文在BACF算法的基础上，采用平移加尺度滤波的思想，设计独立的一维尺度滤波器，与BACF算法无缝结合。只需预测一次目标的位置变化，再利用尺度滤波器预测目标尺度变化。因为两个滤波器单独训练、局部优化，尺度滤波器计算量远小于BACF算法，所以本文算法在保证精准预测目标尺度变化的同时极大提升目标的跟踪速度。实验结果表明：与BACF算法相比，本文算法在不损失跟踪精度的基础上提高约75%的跟踪速度。

Object Tracking Based on Kernelized Correlation Filter with Background-Aware and Fast Discriminative Scale Space

To deal with the limitation of boundary effect and improve tracking speed in the traditional object tracking based on correlation filter, a fast discriminative scale space with background-aware correlation filters (BACF)algorithm is proposed. The BACF algorithm effectively solves the boundary effect caused by cyclic shift, greatly increases the quality and quantity of training samples, and thus improves the performance of the tracker. However, because of its disadvantage in scale detection strategy, it seriously affects its tracking speed. The we design a one-dimensional scale filter. The scale filter and translation filter are trained and optimized independently. The proposed algorithm greatly improves tracking speed while ensuring scale and translation estimation. The experimental results show that compared with BACF algorithm, the proposed algorithm can improve the tracking speed by about 75% without losing the tracking accuracy.