Samuel J. Davey

A comparison of detection performance for several Track-Before-Detect algorithms

A typical sensor data processing sequence uses a detection algorithm prior to tracking to extract point measurements from the observed sensor data. Track before detect (TBD) is a paradigm which combines target detection and estimation by removing the detection algorithm and supplying the sensor data directly to the tracker. Various different approaches exist for tackling the TBD problem. This article compares the ability of several different approaches to detect low amplitude targets. The following algorithms are considered in this comparison: Bayesian estimation over a discrete grid, dynamic programming, particle filtering methods, and the histogram probabilistic multihypothesis tracker. Algorithms are compared on the basis of detection performance and computation resource requirements.

PDAF with multiple clutter regions and target models

This paper presents the theory of a new multiple model probabilistic data association filter (PDAF). The analysis is generalized for the case of multiple nonuniform clutter regions within the measurement data that updates each model of the filter. To reduce the possibility of clutter measurements forming established tracks, the solution includes a model for a visible target. That is, a target that gives sensor measurements that satisfy one of the target models. Other features included in the algorithm are the selection of a fixed number of nearest measurements and the addition of signal amplitude to the target state vector. The nonuniform clutter model developed here is applicable to tracking signal amplitude. Performance of this algorithm is illustrated using experimentally recorded over-the-horizon radar (OTHR) data.

Simultaneous Localization and Map Building Using the Probabilistic Multi-Hypothesis Tracker

This paper demonstrates how the data-association technique known as the probabilistic multi-hypothesis tracker (PMHT) can be applied to the feature-based simultaneous localization and map building (SLAM) problem. The main advantage of PMHT over other conventional data-association techniques is that it has low computational complexity, while still providing good performance. Low complexity is a particularly desirable feature for the SLAM problem where the estimators used may already be costly to implement. The paper also proposes an estimation approach based on generalized expectation-maximization iterations of the PMHT SLAM problem, which is able to achieve low computation complexity at the expense of local convergence. The performance of the PMHT SLAM algorithm is compared with other approaches, and its output is demonstrated on a benchmark data set recorded in Victoria Park, Sydney, Australia

A Comparison of Three Algorithms for Tracking Dim Targets

Tracking of dim, or low signal-to-noise ratio (SNR), targets is commonly achieved using track-before-detect (TBD) techniques. While traditional tracking algorithms operate on detections, typically formed by applying an intensity threshold to the sensor data, TBD algorithms operate directly on un-thresholded sensor data. Increasing the information available to the tracker in this way potentially allows tracking of lower SNR targets compared to trackers using detections. This paper compares three algorithms on simulated data containing a single dim target: a histogram PMHT algorithm and a particle filter, both of which could be classified as track-before-detect algorithms, and a PDA algorithm which operates on detections. The algorithms are compared in terms of detection performance, false track rate and RMS error.

Integrated track maintenance for the PMHT via the hysteresis model

Unlike other tracking algorithms the probabilistic multi-hypothesis tracker (PMHT) assumes that the true source of each measurement is an independent realisation of a random process. Given knowledge of the prior probability of this assignment variable, data association is performed independently for each measurement. When the assignment prior is unknown, it can be estimated provided that it is either time independent, or fixed over the batch. This paper presents a new extension of the PMHT, which incorporates a randomly evolving Bayesian hyperparameter for the assignment process. This extension is referred to as the PMHT with hysteresis. The state of the hyperparameter reflects each models contribution to the mixture, and thus can be used to quantify the significance of mixture components. The paper demonstrates how this can be used as a method for automated track maintenance in clutter. The performance benefit gained over the standard PMHT is demonstrated using simulations and real sensor data

Using Phase to Improve Track-Before-Detect

Track-Before-Detect (TkBD) is a paradigm that combines the target detection and estimation processes that are usually sequentially applied to sensor data in a conventional system. Existing literature uses only the envelope of complex data; this article presents an approach that also includes the phase information. The inclusion of phase is shown to both improve the discrimination of targets from noise and reduce the computation overhead, with improved performance demonstrated using three representative algorithms.

Histogram-PMHT Unfettered

The Histogram Probabilistic Multi-Hypothesis Tracker (H-PMHT) is a parametric mixture-fitting approach to track-before-detect. The original implementations of H-PMHT dealt with Gaussian shaped targets with fixed or known extent. More recent applications have addressed other special cases of the target shape. This article reviews these recent extensions and consolidates them into a new unified framework for targets with arbitrary appearance. The framework adopts a stochastic appearance model that describes the sensor response to each target and describes filters and smoothers for several example models. The article also demonstrates that H-PMHT can be interpreted as the decomposition of multi-target track-before-detect into decoupled single target track-before-detect using the notion of associated images.

A Markov model for initiating tracks with the probabilistic multi-hypothesis tracker

An important problem in multi-target tracking is track initiation and termination. The tracking algorithm aims to discriminate false detections caused by various sources of interference from valid detections caused by targets of interest. This is a problem of model order estimation. One approach to solving this problem with the Probabilistic Data Association Filter has been referred to as target visibility. This paper shows how the target visibility model can be incorporated into the Probabilistic Multi-Hypothesis Tracker to provide integrated initiation and termination.

Multi-sensor track-before-detect for complementary sensors

A typical sensor data processing sequence uses a detection algorithm prior to tracking to extract point measurements from the observed sensor data. Track-before-detect (TkBD) is a paradigm which combines target detection and estimation by removing the detection algorithm and supplying the sensor data directly to the tracker. It is a non-linear non-Gaussian estimation problem, typically requiring numerical methods. While, in general, TkBD schemes are a more computationally expensive method of tracking, removal of the detection threshold allows for the detection and tracking of targets at much lower signal-to-noise ratios. In this paper we develop TkBD algorithms for multiple asynchronous dissimilar sensors. The problem of radar and IRST fusion is presented as an example and the performance of TkBD is compared with a Multi-Hypothesis Tracker. In the example, the TkBD algorithm is able to utilise IRST data to produce high accuracy in azimuth and elevation well beyond the conventional operating range of the sensor.

Tracking system performance assessment

This paper defines an approach to evaluate automatic target tracking system performance against 15 metrics. These metrics are grouped in the categories of track establishment, track maintenance, track error and false tracks. Performance for each metric is expressed as a probability and these are weighted and combined to give an overall measure of performance. The data for this assessment process comes from comparing the tracking system output with tracks in a truth file. This comparison produces a results file for each tracking system and data set. The results file is combined over multiple data sets to form a single file for each tracking system which is the source data to assess tracking performance. This approach is embodied in a Tracker Assessment Tool (TAT) which is described with an example of using it to compare alternative Over-the-Horizon Radar (OTHR) tracking systems.