互相关跟踪算法的多核DSP快速实现  

Fast implementation of cross correlation tracking algorithm using multi-core DSP

作　　者：田立[1] 周付根[1] 孟偲[1] 白相志[1] 金挺[1] 

机构地区：[1]北京航空航天大学图像处理中心,北京100191

出　　处：《高技术通讯》2013年第12期1248-1253,共6页Chinese High Technology Letters

基　　金：863计划(2011AA0641)资助项目

摘　　要：针对光侦系统的目标快速跟踪问题,提出了一种用多核数字信号处理器(DSP)高效实现归一化互相关(NCC)跟踪算法的方案。该方案充分利用德州仪器(TI)最新的6678 DSP的资源和特性,大大提高了跟踪算法实现效率。首先通过增加CPU内核对高速缓存(CACHE)控制器的刷新和失效,解决了多核按照可缓存的方式共享图像数据时出现的存储器一致性问题;其次利用CACHE数据预取机制对相关系数NCC的计算方法做了改进,优先访问同一条CACHE线中的图像数据,使效率提升15%以上;最后提出一种多核间并行计算的机制,通过在从核上设置代理任务来处理主核的并行计算请求并反馈计算结果,8核加速比达到7倍以上,优于开放多处理(OpenMP)标准的效果。该方案在100FPS高帧频舰船目标跟踪系统中运行稳定且可以应用于其他需要快速跟踪的场合。To achieve an optical scout system＇ s fast object tracking, the fast implementation of a normalized cross correlation（NCC） tracking algorithm by using a multi-core digital signal processor（DSP） was studied. To achieve the fast processing,6678 DSP, the new generation multi-core DSP from TI waw utilized. Firstly,The CACHE coherency problem resulting from multiple cores accessing the same image on shared memory through its own CACHE control- ler was solved by adding FLUSH operation and INVALID operation. Secondly, the CACHE line pre-fetch mechanism was utilized to improve the NCC data processing sequence which makes the performance speedup exceeds 15%. Finally, a mechanism of parallel computing was introduced by adding proxy task on slave cores to communicate with master core and master core process task computing and dispatching. The DSP 8 core speed-up-ratio exceeded 7 times. The experimental result showed that the performance exceeded that of OpenMP（ Open Multiprocessing）. This implementation was used in 100FPS ship object tracking system and can be applied to more general cases.

关 键 词：互相关跟踪 多核数字信号处理器(DSP) 多核并行计算 高速缓存(CACHE) 一致性 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]