摘要
目前,目标跟踪已成为计算机视觉领域的一个重要分支。近年来,基于核相关滤波器(Kernel Correlation Filter,KCF)跟踪算法在频域使用循环矩阵性质进行元素的点积运算,与以往的跟踪算法相比,在性能和速度上具有很大的优势。但是当目标尺寸发生变化以及目标受到严重遮挡时,KCF算法不能准确跟踪。因此,在KCF算法基础上做了改进,提出了一种尺度更新算法以及目标跟踪丢失后由粗到精的重定位算法,最后算法在8核DSP处理器TMS320C6678上成功实现了移植。通过多核并行处理,达到30帧/s的实时跟踪帧率。
Object tracking has become an important branch of computer vision currently. In recent years, because the tracking algorithm with kernel correlation filter uses the properties of circulant matrix, the main operation is element-wise product in the frequency domain. It achieved the great performance and speed than the previous tracking algorithm. However, when the target size changes and the target is seriously blocked, it couldn＇t track accurately. Based on the above reasons, a scale updating algorithm and aeoarse-to-fine target relocation algorithm are proposed to improve the KCF algorithm. The algorithm is transplanted on the eight-core DSP processor TMS320C6678 successfully. Through multi-core parallel processing,it achieved 30 frames/s real-time tracking frame rate.
出处
《电子技术应用》
2018年第2期36-38,43,共4页
Application of Electronic Technique
基金
成都市科技惠民项目资助(2015-HM01-00293-SF)
