论文部分内容阅读
目标跟踪中的伺服系统需要极低的跟踪延时，由于粒子滤波跟踪算法固有的庞大计算量使得目标跟踪的精度大受影响。提出了一种粒子滤波跟踪算法在多核DSP系统中的快速实现方法。首先，利用DSP片上的包加速器来降低以太网相机的采集延时以及CPU占用率，CPU占用率从31%降低到10%；其次，通过手动操作高速缓存的刷新和实效，解决了多核同时共享图像数据带来的存储器一致性问题，多个核能通过高速缓存快速获取图像数据；最后，通过在多核核心上设置代理任务的方法，建立了一种多核并行计算的机制。粒子滤波算法中计算复杂度高的运算阶段被分配到多个核心上同时运算，实现了算法的低延时。实验结果显示8核加速比达到7倍以上，优于开放多处理标准OpenMP的并行优化效果。“,”The object tracking servo system requires a low delay from an object moving to starting of rotations while the inherent computational complexity of PF (Particle Filter) affects the tracking precision. In this paper, a multicore DSP parallel implementation strategy for particle filter object tracking was proposed. Firstly, the PA module on chip was used to reduce the GigE image capturing delay and the CPU occupancy. The CPU load was considerably reduced from 31% to 10%. Secondly, by manually FLUSH after writing and INVALID before reading, the memory consistency problem was addressed and cacheable shared image data can be accessed at high efficiency. Finally, a mechanism of parallel computing on multi-core processor was introduced by adding proxy task. The computational intensive stages of particle filter were dispatched to 8 cores to eliminate system delay. Experimental results show that the tracking response time was decreased and algorithmic speedup runs up to 7 and exceeds OpenMP.