杜文彬,毛征,梅伟军,贾文洋,韩嘉隆.基于DSP6678的KCF算法实现及优化系统设计[J].国外电子测量技术,2017,36(7):62-67
基于DSP6678的KCF算法实现及优化系统设计
Implementation of KCF tracking algorithm based on DSP6678 and optimization system design
  
DOI:
中文关键词:  KCF跟踪算法  程序优化  多核数字信号处理器  实时性跟踪
英文关键词:KCF tracking algorithm  program optimization  DSP  real time tracking
基金项目:
作者单位
杜文彬 北京工业大学信息与通信工程学院 北京 100124 
毛征 北京工业大学信息与通信工程学院 北京 100124 
梅伟军 北京工业大学信息与通信工程学院 北京 100124 
贾文洋 北京工业大学信息与通信工程学院 北京 100124 
韩嘉隆 北京工业大学信息与通信工程学院 北京 100124 
AuthorInstitution
Du Wenbin College of Information and Communication Engineering, Beijing University of Technology, Beijing 100124, China 
Mao Zheng College of Information and Communication Engineering, Beijing University of Technology, Beijing 100124, China 
Mei Weijun College of Information and Communication Engineering, Beijing University of Technology, Beijing 100124, China 
Jia Wenyang College of Information and Communication Engineering, Beijing University of Technology, Beijing 100124, China 
Han Jialong College of Information and Communication Engineering, Beijing University of Technology, Beijing 100124, China 
摘要点击次数: 1127
全文下载次数: 970
中文摘要:
      核相关滤波算法在目标旋转、部分遮挡等情况下具有很强的鲁棒性,但是其在实际工程上无法满足实时性跟踪的要求。借助DSP嵌入式平台,提出一种实时目标跟踪系统实现方案。首先,基于TMS32C6678处理器的硬件平台,对二维快速傅里叶变换(FFT2D)进行算法优化;其次,对KCF算法进行C语言级代码优化,包括循环展开、使用内联函数、使用关键字const等。实测结果表明,对640×480像素的图像,跟踪目标为32×64像素时,波门大小为64×128。在满足跟踪效果的同时,其跟踪速度最高达到25 ms/帧,符合实时性要求。
英文摘要:
      The kernel correlation filtering algorithm (KCF) is robust to target rotation and partial occlusion. However, it can not meet the requirements of real time tracking in the actual project. Based on DSP embedded platform, this paper proposes an implementation scheme of real time target tracking system. Firstly, based on the hardware resources of TMS320C6678 processor, the algorithm is optimized for two dimensional Fast Fourier Transform(FFT2D). Secondly, optimize the KCF algorithm by C language level code, including loop expansion, the use of inline functions, the use of const keywords and so on. The results show that when the input image’s size is 640×480 pixels and the tracking target’s size is 32×64 pixels,the size of bomen is 64×128. It meets the tracking effect, meanwhile the speed up to 25 ms per frame, which meets the real time requirements.
查看全文  查看/发表评论  下载PDF阅读器