GPU加速的差分进化粒子滤波算法

Speeding up differential evolution particle filter algorithm by GPU

为了解决实时系统中粒子滤波的计算复杂性问题,提出了一种零bank冲突并行规约的差分进化粒子滤波方法。该方法首先分析了并行差分进化粒子滤波算法在GPU中的内存访问模式,根据粒子滤波器的均方根误差与内存访问bank(存储体)冲突度成正比的关系,提出了一种去除bank冲突的有填充寻址的差分进化粒子滤波算法,降低了计算复杂度。将该算法在NVIDIA GTX960 GPU中实现,与串行差分进化粒子滤波算法进行比较。实验表明,随着粒子数增加,计算量以指数增加,采用GPU加速的跟踪算法的执行时间明显减少,有效提高了跟踪精度,降低了计算时间。

This paper proposed a differential evolution particle filter method based on parallel protocol of zero bank conflict,in order to solve the computational complexity of particle filtering in real-time systems. Firstly it analysed the memory access mode of parallel differential evolution particle filter algorithm in GPU. Then according to the relationship that the root mean square error was proportional to the degree of bank conflicts in memory access,it proposed a differential evolution particle filter with filled addressing mode to remove bank conflict,which reduced the computational complexity. It implemented the algorithm in NVIDIA GTX960 GPU,and compared with the serial differential evolution particle filter algorithm. As the number of particles increases,the amount of calculation increases exponentially. Theoretical analysis and simulation results show that the tracking algorithm using GPU acceleration is significantly reduced the execution time. It improves the tracking accuracy and reduces the computation effectively.