High Performance Low Complexity Multitarget Tracking Filter for a Array of Non-directional Sensors

TL;DR

This paper introduces the TT filter, a high-performance, low-complexity algorithm for tracking multiple targets with non-directional sensors, outperforming Kalman and particle filters in accuracy and efficiency.

Contribution

The paper presents the TT filter, a novel tracking algorithm that combines Hessian-based ML estimation and nonlinear transformations to improve accuracy and reduce computational complexity.

Findings

Superior accuracy over Kalman and particle filters

Lower computational complexity

Effective initial target acquisition method

Abstract

This paper develops an accurate, efficient filter (called the `TT filter') for tracking multiple targets using a spatially-distributed network of amplitude sensors that estimate distance but not direction. Several innovations are included in the algorithm that increase accuracy and reduce complexity. For initial target acquisition once tracking begins, a constrained Hessian search is used to find the maximum likelihood (ML) target vector, based on the measurement model and a Gaussian approximation of the prior. The Hessian at the ML vector is used to give an initial approximation of the negative log likelihood for the target vector distribution: corrections are applied if the Hessian is not positive definite due to the near-far problem. Further corrections are made by applying a transformation that matches the known nonlinearity introduced by distance-only sensors. A set of integration points is constructed using this information, which are used to estimate the mean and moments of the target vector distribution. Results show that the TT filter gives superior accuracy and lower complexity than previous alternatives such as Kalman-based or particle filters.