摘要
目标跟踪作为计算机视觉研究的重要方向,具备良好的发展前景。针对传统Mean-shift跟踪算法在目标存在背景干扰或遇到遮挡时,目标跟踪框不准确以及目标丢失的现状进行了相关研究。结合目标特征建模优势,提出了基于Hu不变矩相似关联度的Mean-shift运动跟踪算法。该算法采用Hu不变矩特征描述子提取目标特征向量矩进行建模,并由目标位移矢量拟合预估Mean-shift初始迭代搜索的位置。通过Pearson相关系数度量目标特征矩与候选特征矩之间关联度,并结合8步态迭代搜索的方式定位运动目标,实现了目标跟踪。试验表明,当目标存在遮挡时,该算法依然能够进行有效跟踪定位,有效地改善了复杂条件下跟踪框BBox与真实值(GT)之间的交并比(IOU)值,减少了目标搜索时的迭代次数,具有较好的鲁棒性。
Target tracking, as an important direction of computer vision research, has good prospects for development. The current situation of the target tracking frame inaccuracy and target loss when the traditional Mean-shift tracking algorithm has background interference or encounters occlusion is studied. Combining the advantages of target feature modelling, a Mean-shift motion tracking algorithm based on Hu-invariant moment similarity correlation degree is proposed. The algorithm uses Hu-invariant moment feature descriptors to extract the target feature vector moments for modelling, and the position of the initial iterative search of Mean-shift is predicted by fitting the target displacement vector. Target tracking is achieved by measuring the correlation between target eigen moments and candidate eigen moments by Pearson correlation coefficients and locating the moving target by combining the 8-step iterative search. The experimental results show that the algorithm can track and locate the target effectively even when the target is occluded, effectively improving the intersection and intersection over union(IOU) between the BBox and the ground truth(GT) of the tracking frame under complex conditions, reducing the number of iterations in the target search, and having better robustness.
作者
郝行猛
肖娜
舒梅
HAO Xingmeng;XIAO Na;SHU Mei(Zhejiang Dahua Technology Co.,Ltd.,Hangzhou 310051,China)
出处
《自动化仪表》
CAS
2022年第11期24-28,共5页
Process Automation Instrumentation
