A particle filter to track multiple objects Carine Hue Patrick Pérez and Jean-Pierre Le Cadre.

A particle filter to track multiple objects

Carine Hue Patrick Prez and Jean-Pierre Le Cadre

ContextTwo applications:

Signal processing: target tracking

Image analysis: people tracking

observertargetbearing

ChallengesGeneric tracking issues(re)initialization Clutter (false alarms or background)Occlusions

Specific multiple-object issuesComplexityIdentity confusion when crossingVarying number of entities

Basic ingredients

Hidden variables (position, velocity, shape of the objects)

Data ( bearings, images at each time)

Estimation of posterior

Dynamics:

Likelihood:

Hidden variablesLow-dimensional single-object hidden variable:

Bearings-only tracking:

People tracking:

: Fourier descriptors

: 2D translation vector of the center

Multiple-object hidden variable

Dynamics

Single-object dynamics prior: first-order auto-regressive process

Multi-object dynamics: independent single-object dynamics

time t-1time t ??velocity vectorposition orfirst Fourier coef.

DataBearings-only tracking: highly non-linear wrt the state variablePeople tracking: obtained with a motion-based segmentation

How to assign the data to the states ?

Data associationNotation: association vector: if is issued from objectAssociation assumptions:one measurement can originate from one object or from the clutter (false alarms)one object can produce zero or several measurements at one timeExhaustive enumeration ! NP-hard problem

Solution: probabilistic association: false alarmsIrisa:

Particle filteringCurrent cloud at time t

Particle filteringPropagation by sampling from the dynamic prioror from an importance function based on the data

Particle filteringWeighting according to data likelihood

Particle filteringResampling from the weighted particle set

Particle filteringPrediction again . . .

Example: bearings-only tracking Particle cloud for one object

Multi-object data likelihood

Estimation of the association probabilities vectorwith a Gibbs sampler

take the average value of

Example: bearings-only trackingMeasurement equation

Differences between the simulated bearings for each object

Example: bearings-only trackingEstimated association probabilities

Estimated trajectories

Example: people trackingData likelihood

Example: people tracking

Conclusion and further workGeneric multi-object tracker based on a mix of particle filteringand Gibbs sampling suitable for various applications

DrawbacksCostlyFixed number of objectsFor people tracking, the data association could be avoidedwhen there is no ambiguities

PerspectivesVarying number of objectsInitialization issueTest with real bearings measurements

The End

Questions?

Eigenface examplesEigenface examples