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 Pérez and Jean-Pierre Le Cadre

Context Two applications:

Signal processing: target tracking

Image analysis: people tracking

observer

target

bearing

Challenges Generic tracking issues

(re)initialization Clutter (false alarms or background) Occlusions

Specific multiple-object issues Complexity Identity confusion when crossing Varying 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 variables Low-dimensional single-object hidden variable:

Bearings-only tracking:

People tracking:

: Fourier descriptors

: 2D translation vector of the center

Multiple-object hidden variable

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

Multi-object dynamics: independent single-object dynamics

Dynamics

time t-1

time t ??

velocity vector

position orfirst Fourier coef.

Bearings-only tracking: highly non-linear wrt the state variable

People tracking: obtained with a motion-based segmentation

How to assign the data to the states ?

Data

Data association Notation: association vector: if is issued from object Association assumptions:

one measurement can originate from one object or from the clutter (false alarms)

one object can produce zero or several measurements at one time Exhaustive enumeration ! NP-hard problem

Solution: probabilistic association:

false alarms

Irisa:Irisa:

Particle filtering

Current cloud at time t

Particle filtering

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

Particle filtering

Weighting according to data likelihood

Particle filtering

Resampling from the weighted particle set

Particle filtering

Prediction 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

t

Example: bearings-only tracking Measurement equation

Differences between the simulated bearings for each object

Example: bearings-only tracking Estimated association probabilities

Estimated trajectories

Example: people tracking Data likelihood

Example: people tracking

Conclusion and further work Generic multi-object tracker based on a mix of particle

filteringand Gibbs sampling

suitable for various applications

Drawbacks Costly Fixed number of objects For people tracking, the data association could be avoidedwhen there is no ambiguities

Perspectives Varying number of objects Initialization issue Test with real bearings measurements

The End

Questions?