摘要
传统压缩跟踪算法使用固定学习率更新特征分布,导致跟踪易受遮挡影响且鲁棒性较低。为此,提出一种可自动调节特征分布学习率的压缩跟踪算法。利用压缩感知理论得到样本的压缩域特征并计算其在正负类中的特征分布,结合两帧之间特征分布重叠度和正类更新阈值自适应更新特征分布,通过样本分类实现目标跟踪。在此基础上,利用相邻两帧目标改进的SIFT特征求解目标尺度变化,使跟踪窗口随目标变化实时更新。实验结果表明,该算法可有效抵抗遮挡、光线、尺度等因素对跟踪的干扰,具有较高的准确性、鲁棒性以及实时性。
A Compression Tracking（CT） algorithm is proposed to automatically adjust the learning rate of feature distribution,which is based on the problem that the fixed learning rate is used to update the feature distribution of the tracking algorithm,which is easily affected by the occlusion and the robustness is low.Compressed domain feature samples are obtained by the compressive sensing theory,calculate the distribution characteristics of various compression characteristics in the positive class and negative class,use the distribution of overlap between the two frames combines with adaptive threshold update distribution.Target tracking is achieved by sample classification.At the same time,the algorithm makes use of the improved SIFT features of adjacent two frames to solve the target scale change,and realize the tracking window with the change of the target in real time.Experimental results show that the proposed algorithm can effectively resist the interference of tracking,such as occlusion,ray and scale.It has higher accuracy,robustness and real-time performance.
出处
《计算机工程》
CAS
CSCD
北大核心
2018年第2期264-270,共7页
Computer Engineering
关键词
特征分布
压缩特征
稀疏矩阵
巴氏系数
SIFT特征
仿射变换
feature distribution
compression feature
sparse matrix
Bhattacharyya coefficient
SIFT feature
affine transformation
