Parallelization of a multiple model multitarget tracking algorithm with superlinear speedups

Abstract

Real-time optimal filtering for stochastic systems with multiresolutional measurements. Comparison of two-sensor tracking methods based on state vector fusion and measurement fusion. The interacting multiple model (IMM) estimator has been shown to be very effective when applied to air traffic surveillance problems. However, because of the additional filter modules necessary to cover the possible target maneuvers, the IMM estimator also imposes an increasing computational burden. Hence, in an effort to design a real-time multiple model multitarget tracking algorithm that is independent of the number of modules used in the state estimator, we propose a "coarse-grained" (dynamic) parallelization that is superior, in terms of computational performance, to a "fine-grained" (static) parallelization of the state estimator, while not sacrificing tracking accuracy. In addition to having the potential of realizing superlinear speedups, the proposed parallelization scales to larger multiprocessor systems and is robust, i.e., it adapts to diverse multitarget scenarios maintaining the same level of efficiency given any one of numerous factors influencing the problem size. We develop and demonstrate the dynamic parallelization on a shared-memory MIMD multiprocessor for a civilian air traffic surveillance problem using a measurement database based on two FAA air traffic control radars.

High-order interacting multiple-model estimation for hybrid systems with Markovian switching parameters

“…Atherton and Lin (1994) suggested and carefully examined the possibility of implementation of the IMM1 algorithm as a network of four transputers. Recently Popp et al (2000) developed a sharedmemory multiprocessor algorithm based on the IMM1 paradigm. It is still an open area for future investigations on how to implement the high-order IMMi algorithm, taking into account practical constraints referring to the processing time.…”

An Optimization-Based Parallel Particle Filter for Multitarget Tracking

“…In the multitarget particle filter, data association and estimation are done jointly. Thus, multitarget particle filter parallelization is very different from Kalman filter or Interacting Multiple Model (IMM) estimator [4] based parallel multitarget implementations [29] [28], where an Auction based assignment is commonly used to handle measurement-totrack association. For nonlinear systems, existing association techniques like assignment are not directly applicable in conjunction with particle filter based techniques [15].…”

“…We consider the problem in a master-slave (or, more appropriately, primary-secondary) topology, which is suitable for both multiprocessor architectures and network of workstations. Previously, scheduling algorithms for multitarget tracking were developed [29] [27] within the IMM-Assignment framework. In this thesis, we are concerned with the mapping of a multi target particle filter onto a set of single instruction, multiple data stream (SIMD) multiprocessors, wherein the processors may be homogeneous or heterogeneous.…”

<title>M-best SD assignment algorithm with application to multitarget tracking</title>

“…. , in -1) corresponding to subproblem D on the queue Q having minimum cost (or maximum likelihood), i.e., a(m) () (29) The complexity of our rn-best S D assignment algorithm is as follows. Per S D assignment problem, we perform one partitioning task for each of the rn-best assignments determined.…”

“…Consequently, the assignment of measurements to tracks and false alarms is often done in a variety of ways. For multitarget tracking problems, a literature survey shows numerous well-known approaches proposed over the years, e.g., (in order of decreasing complexity) Multiple Hypothesis Tracking (MHT),4"6'33 multidimensional S D (S 3)t assignment,12 '13'21'22'2527 Joint Probabilistic Data Association (JPDA),4 and two-dimensional (2 D) assignment (single scan processing) algo- 1,2,5,14,15,[28][29][30]35 Data association becomes especially difficult if the sensors are passive and measure LOS angles only for the targets. Measurements from multiple scans (S 3) have to be associated to determine the estimates of target states, leading to a combinatorial explosion of the problem.…”