[Paper Review] Multi-object filtering with stochastic populations

Emmanuel Delande, Jérémie Houssineau+1 more|arXiv (Cornell University)|Jan 19, 2015

Target Tracking and Data Fusion in Sensor Networks14 references8 citations

TL;DR

This paper introduces a fully probabilistic estimation framework for multi-target tracking using stochastic populations, enabling principled filtering through statistical moments and information-theoretic gain functions. The proposed Distinguishable and Independent Stochastic Populations (DISP) filter maintains track continuity for detected targets while propagating uncertainty over undetected targets, offering a robust solution to challenges like missed detections and false alarms in complex surveillance scenarios.

ABSTRACT

While the design of automated knowledge-based sensor scheduling is relevant to many multi-target detection and tracking problems, tracking algorithms are rarely built for this purpose and their outputs provide little flexibility for the design of sensor policies. In this paper, we present an estimation framework for stochastic populations in the context of multi-target estimation problems. Fully probabilistic in nature, it allows for the evaluation of the population of targets through statistical moments, as well as the assessment of sensor observations through information-theoretical gain functions. We present a principled solution derived from this framework addressing challenging multi-target scenarios involving missed detections and false alarms, the filter for Distinguishable and Independent Stochastic Populations, which propagates information on previously-detected targets as well as yet-to-be-detected targets while maintaining track continuity.

Motivation & Objective

• To address the lack of flexible, principled filtering frameworks for multi-target estimation that support automated sensor scheduling.
• To model uncertainty in target composition and state using statistical moments within a stochastic population framework.
• To maintain track continuity for previously detected targets while propagating information about undetected targets.
• To derive information-theoretic gain functions for sensor scheduling based on filter outputs.
• To provide a scalable, principled alternative to heuristic track-based filters like MHT and JPDA.

Proposed method

• The framework models a population of targets as a stochastic set, using two types of statistical moments: one for population composition and one for target state.
• It employs a common probabilistic description for indistinguishable subpopulations to avoid unnecessary permutations during data association.
• The DISP filter propagates posterior distributions over detected targets and prior knowledge about undetected targets using a state-space model with Markovian transitions.
• The time prediction step evolves the stochastic population forward using a Markov transition kernel, incorporating target birth and death processes.
• The data update step applies Bayes' rule using sensor observations, accounting for missed detections and false alarms via a likelihood model.
• Information gain functions are derived from the filter’s output to quantify the value of observations for specific targets or regions, enabling sensor scheduling.

Experimental results

Research questions

• RQ1 How can a fully probabilistic framework model uncertainty in both target composition and individual target states in multi-target tracking?
• RQ2 How can track continuity be preserved for detected targets while simultaneously propagating information about undetected targets?
• RQ3 What principled information-theoretic gain functions can be derived from the filter output to guide sensor scheduling?
• RQ4 How can the framework avoid combinatorial explosion in data association by exploiting indistinguishability of undetected targets?
• RQ5 To what extent does the DISP filter outperform heuristic filters in scenarios with missed detections and false alarms?

Key findings

• The DISP filter successfully maintains track continuity for previously detected targets while propagating uncertainty over undetected targets through a unified probabilistic framework.
• The framework enables the derivation of information-theoretic gain functions that quantify the value of sensor observations for specific targets or regions, supporting automated sensor scheduling.
• Statistical moments of the stochastic population allow for the evaluation of target state and population composition, extending regional statistics used in FISST-based filters.
• The filter avoids redundant permutations in data association by modeling indistinguishable subpopulations with a shared probabilistic description.
• A principled approximation of the DISP filter is proposed to reduce computational complexity while preserving key estimation properties.
• The theoretical derivation confirms that the set of valid hypotheses at time $ t $, $ extbf{H}_t $, is equivalent to the set of consistent track configurations $ \mathrm{Const}(\mathbb{I}^\bullet_t) $, ensuring correctness of the filtering process.