基于DM642的KLT跟踪算法的实现及优化

摘要

Kanade-Lucas-Tomasi (KLT)算法是基于图像特征点的跟踪算法,由目标对象特征点提取,特征点跟踪两部分组成.本文首先阐述了KLT算法的基本原理,分析了影响算法执行速度的主要原因.分析表明KLT算法的操作主要集中在乘加运算和循环,图像卷积运算和循环占用的执行时间比较长.针对TMS320DM642 DSP的硬件平台特点,提出了算法优化的若干策略.通过配置编译环境,合理安排数据类型,消除存储器相关性,使用内联函数以及分解多层循环等方法,对算法的实现进行了优化.实验结果表明,优化后代码执行速度是优化前的3倍多.%KLT is a tracking algorithm based on image feature points,which is composed of two parts,namely the feature point extraction and the feature point tracking. In this paper, the basic principle of the KLT algorithm is proposed, and the main factors which influence the speed of the KLT algorithm are analyzed. It is found that the multiplication-addition and the loop operations cost the most processing time in the KLT algorithm. The image convolution operation and the implementation of loops take much more time. A sense strategies of the algorithm optimization are proposed considering the hardware platform of the TMS320DM642. The algorithm is implemented optimizely, by configing the compile environment, arranging the data types reasonably, eliminating the memory correlation, using the intrinsics and decompositing the number of loops. Experimental results show that the execution speed of the optimized code is three times faster than that without optimization.