【作者】 李明；

【导师】 崔丽群；

【作者基本信息】 辽宁工程技术大学 ， 软件工程（专业学位）， 2022， 硕士

【摘要】 核相关滤波算法跟踪速度快但精度低,在目标发生形变、模糊和快速运动等多种情况下会出现跟踪漂移或跟踪失败的问题,针对以上问题提出了改进KCF的目标跟踪算法。首先,为了提高目标外观建模的准确性,本文利用深度卷积特征对目标外观建模;其次,为了解决传统核相关滤波算法在目标发生形变或旋转的情况下产生跟踪误差的问题,本文利用高斯混合模型生成目标的样本空间,增加了目标样本的多样性,选取高斯混合模型中的目标样本训练并更新核相关滤波器;最后,为了解决传统核相关滤波算法在跟踪失败时无法找回正确的跟踪目标问题,本文在核相关滤波算法中加入了跟踪置信度监测器,来监测当前跟踪是否成功,将跟踪误差作为跟踪结果的置信度,当跟踪误差高于设定阈值,判定当前帧跟踪失败,启动目标重检测模式。目标重检测模式是以上一帧目标框为中心扩大当前帧目标寻找范围,范围内对相关滤波器用滑窗法得到目标可能位置的坐标分布。对该位置分布进行K-means聚类,找出分布规律得到M个位置分布域,分布域中以类簇心为中心进行相关滤波定位目标中心位置,得到M个目标中心候选点。通过计分器选出最佳目标位置,重新定位目标。本文算法在OTB2015和VOT2016数据集上,与Staple,MEEM,DSST,SRDCF,KCF,LSART,ECO和ASRCF算法对比,验证了本文算法在跟踪精度和鲁棒性方面超过了其他相关滤波对比算法。该论文有图25幅,表6个,参考文献52篇。更多还原

【Abstract】 The kernel correlation filtering algorithm has fast tracking speed but low accuracy.In many cases,such as target deformation,blur and fast motion,there will be tracking drift or tracking failure,aiming at the above problems,an improved KCF target tracking algorithm is proposed.Firstly,in order to improve the accuracy of target appearance modeling,the depth convolution feature is used to model the target appearance;Secondly,in order to solve the problem of tracking error in the case of target deformation or rotation in the traditional kernel correlation filter algorithm,in this paper,the Gaussian mixture model is used to generate the sample space of the target,which increases the diversity of the target samples,and the target samples in the Gaussian mixture model are selected to train and update the kernel correlation filter;Finally,in order to solve the problem that the traditional kernel correlation filtering algorithm cannot find the correct tracking target when the tracking fails,this paper adds a tracking confidence monitor to the kernel correlation filtering algorithm to monitor whether the current tracking is successful.The tracking error is taken as the confidence of the tracking result.When the tracking error is higher than the set threshold,it determines that the current frame tracking fails and starts the target redetect mode.Target redetection mode is that the above one frame target box expands the current frame target search range,within the scope of the relevant filter to obtain the coordinate distribution of the possible target position by the sliding window method.The location distribution was K-means clustered,to find out the distribution law to obtain M position distribution domain,in the distribution domain with the correlation filter to locate the target center position,to obtain M target center candidates.Select the best target location,reset the target,and solve the traditional core correlation filter algorithm fails to retrieve the correct tracking target when the tracking failure.This algorithm verifies the improvement in tracking accuracy on OTB2015 and VOT2016 in comparison with Staple,MEEM,DSST,SRDCF,KCF,LSART,ECO,ASRCF and VOT2016 datasets.It is verified that the algorithm in this paper outperforms other correlation filter comparison algorithms in tracking accuracy and robustness.This paper has 25 figures,6 tables and 52 references.更多还原