Extensions to the probabilistic multi-hypothesis tracker for tracking, navigation and SLAM.

Abstract

Multi-target tracking is a problem that involves estimating target states from noisy data whilst simultaneously deciding which measurement was produced by each target. The Probabilistic Multi-Hypothesis Tracker (PMHT) is an algorithmthat solves the multi-target tracking problem. This thesis presents extensions to the PMHT to address problems that may arise in the use of real sensors and considers multi-target tracking techniques for use in other applications such as autonomous vehicles. It is generally assumed that a sensor collects a set of noisy position measurements at known times. In some situations, the time information may not be reliable and cause filtering issues. This thesis derives an extension to the PMHT that introduces an assignment index that identifies the true time at which a measurement was collected. This extension of the PMHT allows for tracking on measurements with time errors, such as time delays. A further extension allows the PMHT algorithm to simultaneously estimate the time error parameters whilst tracking targets. The above extension is applied to the problem of planning paths for multiple platforms to explore an unknown area. Given a set of locales to be visited and the platform initial positions, the path planning problem has the same mathematical form as a multi-target tracking problem, with locales as measurements and the platforms as targets. The extended PMHT algorithm uses hypothesised time-stamps to associate locales to platforms and times simultaneously. Autonomous vehicles are expected to use information from their sensors to navigate and map their environment. Simultaneous localisation and mapping (SLAM) is the name given to this task and is essentially a multi-target tracking problem. This thesis proposes the use of PMHT and landmark classification information received with measurements to improve the performance of SLAM.