Kalman filter-based retrodiction plays an indispensable role in modern multitarget tracking and retrodiction (MTTR) algorithms. To this end, the Rauch–Tung–Striebel (RTS) smoother is a widely used Kalman filter-based target state smoother. With a large number of targets, MTTR algorithms, particularly with large window sizes, become very computationally intensive. If not addressed, these algorithms will not meet the requirements for tracking a large number of targets in real time. A natural approach is to parallelize these algorithms to render them useful, particularly in the context of emerging multicore platforms. However, this is nontrivial, as the governing mathematical framework of the RTS smoother, namely the dependencies between complex computations, prevents any form of parallelization. Although the MTTR component can naively be parallelized ignoring the smoothing component, the overall benefit, as we demonstrate in this article, is a fraction of the best possible benefits. In this article, by carefully reformulating the underlying mathematical framework that is necessary for retrodiction, we propose a novel, easily parallelizable RTS smoother. The proposed parallelized RTS smoother we outline in this article has best data reuse and enables the overall MTTR problem to be parallelized more efficiently. We demonstrate this on a state-of-the-art multicore processor platform using the shared-memory parallelism. Our results show that the parallel MTTR solution, which includes gating, assignment, tracking, and retrodiction, can offer nearly 150 times speed up against a fully sequential version. With excellent computational performance, our proposed RTS smoother enables very large window sizes with little or no impact on the overall performance.

中文翻译：

---

大规模多目标跟踪的并行追溯算法

基于卡尔曼滤波器的追溯在现代多目标跟踪和追溯（MTTR）算法中起着不可或缺的作用。为此，Rauch-Tung-Striebel（RTS）平滑器是一种广泛使用的基于Kalman滤波器的目标状态平滑器。在目标数量众多的情况下，MTTR算法（尤其是具有较大窗口大小的算法）的计算量很大。如果不解决，这些算法将无法满足实时跟踪大量目标的要求。一种自然的方法是并行化这些算法以使其有用，尤其是在新兴的多核平台的情况下。但是，这是不平凡的，因为RTS平滑器的主要数学框架（即复杂计算之间的依赖性）可以防止任何形式的并行化。尽管可以忽略平滑部分而天真地并行化MTTR组件，但是正如我们在本文中演示的那样，总的好处只是最好的好处的一小部分。在本文中，通过仔细地重新构造追溯所需的基础数学框架，我们提出了一种新颖，易于并行化的RTS平滑器。我们在本文中概述的拟议中的并行RTS平滑器具有最佳的数据重用性，并使整个MTTR问题得以更有效地并行化。我们使用共享内存并行机制在最新的多核处理器平台上对此进行了演示。我们的结果表明，包括门控，分配，跟踪和追溯在内的并行MTTR解决方案与完全顺序版本相比，可以提供近150倍的速度提高。凭借出色的计算性能，