An information-based approach to sensor resource allocation.

Abstract

This work addresses the problem of scheduling the resources of agile sensors. We advocate an information-based approach, where sensor tasking decisions are made based on the principle that actions should be chosen to maximize the information expected to be extracted from the scene. This approach provides a single metric able to automatically capture the complex tradeoffs involved when choosing between possible sensor allocations. We apply this principle to the problem of tracking multiple moving ground targets from an airborne sensor. The aim is to task the sensor to most efficiently estimate both the number of targets and the state of each target simultaneously. The state of a target includes kinematic quantities like position and velocity and also discrete variables such as target class and target mode (e.g., turning or stopped). In many experiments presented herein, target motion is taken from real recorded vehicle histories. The information-based approach to sensor management involves the development of three interrelated elements. First, we form the joint multitarget probability density (JMPD), which is the fundamental entity capturing knowledge about the number of targets and the states of the individual targets. Unlike traditional methods, the JMPD does not assume any independence, but instead explicitly models coupling in uncertainty between target states, between targets, and between target state and the number of targets. Furthermore, the JMPD is not assumed to be of some parametric form (e.g., Gaussian). Because of this generality, the JMPD must be estimated using sophisticated numerical techniques. Our representation of the JMPD is via a novel multitarget particle filter with an adaptive sampling scheme. Second, we use the estimate of the JMPD to perform (myopic) sensor resource allocation. The philosophy is to choose actions that are expected to maximize information extracted from the scene. This metric trades automatically between allocations that provide different types of information (e.g., actions that provide information about position versus actions that provide information about target class) without ad hoc assumptions as to the relative utility of each. Finally, we extend the information-based paradigm to non-myopic sensor scheduling. This extension is computationally challenging due to an exponential growth in action sequences with horizon time. We investigate two approximate methods to address this complexity. First, we directly approximate Bellman's equation by replacing the value-to-go function with an easily computed function of the ability to gain information in the future. Second, we apply reinforcement learning as a means of learning a non-myopic policy from a set of example episodes.