中文摘要：

针对核相关滤波器（KCF）跟踪算法在目标跟踪中存在尺度变化、严重遮挡、相似目标干扰和出视野情况下跟踪失败等问题,提出了一种基于KCF的长期目标跟踪算法。该算法在分类器学习中加入空间正则化,利用基于样本区域空间位置信息的空间权重函数调节分类器系数,使分类器学习到更多负样本和未破坏的正样本,从而增强学习模型的判别力。然后,在检测区域利用Newton方法完成迭代处理,求取分类器最大响应位置及其目标尺度信息。最后,对最大响应位置的目标进行置信度比较,训练在线支持向量机（SVM）分类器,以便在跟踪失败的情况下,重新检测到目标而实现长期跟踪。采用OTB-2013评估基准50组视频序列验证了本文算法的有效性,并与30种其他跟踪方法进行了对比。结果表明：本文提出的算法跟踪精度为0.813,成功率为0.629,排名第一,相比传统KCF算法分别提高了9.86%和22.3%。在目标发生显著尺度变化、严重遮挡、相似目标干扰和出视野等复杂情况下,本文方法均具有较强的鲁棒性。

英文摘要：

As Kernelized Correlation Filters （KCF） tracking algorithm has poor performance in han- dling scale-variant, heavy occlusion, similar target interfere and out of view, this paper proposes a long-term tracking approach based on the KCF. Firstly, a spatial regularization component was intro- duced in the learning of a classifier , the classifier coefficients were penalized depending on the weight function of spatial location information in sample locations. By which the classifier could learn signifi- cantly larger set of negative training samples and uncorrupted positive samples, so that the discrimina- tive power of learned model was increased. Then, the Newton method was used to complete the itera- tion and obtain the maximizing response location and target score of the classifier in the detection area. Finally, to re-detect the target in the case of tracking failure and achieve a long-term tracking, the confidence of the target with the maximum response score was compared and an online Support Vector Machine （SVM） classifier was trained. To verify the feasibility of the proposed algorithm, fifty groups of OTB-2013 benchmark video sequences were tested and the obtained results were compared with thirty kinds of other tracking algorithms. Experimental results indicate that the precision andsuccess rate from the proposed method are respectively 0. 813 and 0. 629, ranking first. Compared with traditional KCF tracking algorithm, the proposed approach respectively improves by 9.86% and 22.3% in the precision and the success rate. Moreover, it is robust to significant scale changing, heavy occlusion, interfere with similar target, out of view and other complex scenes.