48h x 30w poster template

Download Report

Transcript 48h x 30w poster template

Visual Object Tracking Using Particle Filters : A Survey
Satarupa Mukherjee
Department of Computing Science, University of Alberta, Edmonton, Canada
Submission ID: 23
[email protected]
Use of Particle Filter in Multi-target “Track Before Detect” Application
Introduction
• Tracking is one of the most popular methods in video processing.
• Most of the methods in tracking are object-dependent.
• In the last decade particle filter based tracking became very popular due to its
applicability to practical problems.
Concepts
1.
2.
Dynamic systems.
Analysis needs two vectors and two models.
State
vector
5.
Deals with targets that are influenced by the proximity and/or behavior of other targets.
Problem: The interaction.
Solution: A joint tracker is developed that includes a more sophisticated motion model
to maintain the identity of targets throughout an interaction.
Result : 20 ants tracked in a small arena.
Models
Measurement
vector
4.
Use of Particle Filter for Interacting Targets
System
model
Vectors
3.
Deals with “track before detect (TBD)” application in multi-target setting.
Problem: Targets may pop up and also disappear and the number of targets may vary
over time.
Solution: A recursive Bayesian TBD algorithm that can deal with the situation of a limited
number of closely spaced targets. A particle filter is used to perform this recursive TBD.
Result : A stronger primary target and a weaker but more agile secondary target have
been taken into account which resemble a moving launch platform.
Measurement
model
In Bayesian approach to dynamic state estimation, the probability density function of
the dynamic state is constructed based on all available information.
Optimal Bayesian solutions for non-linear tracking problems are either tractable or
intractable.
Particle filter is an approximation strategy for intractable optimal solution.
Frame 9043
Frame 9080
Frame 9083
Fig.2. Tracking results of three interacting ants
Condensation Algorithm
Basics
1. Principle of particle filtering -> tracking an object evolving over time, typically with a
non-Gaussian and multi-modal probability density function.
2. Concept of particle.
3. Particle filter algorithm is recursive in nature.
4. Operates in two phases : prediction and update.
Color Based Particle Filter
A technique for tracking deformable objects in image sequences with complex
background.
Problem: When several objects sharing the same color description are present in a frame,
the particles of a color filter are attracted by different objects.
Solution : Color histograms are used as object features.
Result : Tracking results on a soccer sequence video.
Frame 1
Frame 4
Frame 8
Frame 33
Frame 40
Fig.1. Results of tracking in a soccer video.
Tracking Multiple Objects with Particle Filter
Deals with estimation of state of an unknown number of moving objects .
Problem : Measurements may arrive both from the targets if they are detected and from
clutter.
Solution : The algorithm consists of two major steps of prediction and weighting of particle
filter along with computation of estimation of the vector of assignment probabilities.
Result : Applied to three targets which follow a near-constant-velocity model.
Used for tracking curves in visual clutter.
Problem: Track outlines and features of foreground objects, modeled as curves, as they
Frame 91
Frame 121
move in substantial clutter. This is challenging
as background clutter elements may mimic parts
of foreground features.
Solution: Application of probabilistic models of
Frame 265
Frame 221
object shape and motion for analyzing the video
stream.
Result : Tracking agile motion against cluttered
background.
Fig.3. Tracking agile motion in clutter
Future Scope
In color based particle filter, color histograms can be computed in different regions
of the target for taking the topological information into account.
For the case of particle filter for interacting objects, the approach can be validated
by tracking much more number of interacting objects.
For condensation algorithm, alternative observation models can be developed for
making greater use of image intensity.
Reference
1. Arulampalam S.,Simon Maskell, Neil Gordon and Tim Clapp A Tutorial on Particle Filters for Online
Nonlinear/Non-Gaussian Bayesian Tracking IEEE Transactions on signal processing,50(2):174-188,2002.
2. Acton S. and Ray N. Biomedical image Analysis: Tracking Morgan & Claypol Publisher,USA, 2006.
3. Jacek Czyz, Branko Ristic and Benoit Macq, A Color-Based Particle Filter for Joint Detection and Tracking
of Multiple Objects. ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing
– Proceedings II:art. no. 1415380, II217-II220,2005.
4. C Hue, J-P. LE CADRE, Tracking multiple objects with particle filtering. IEEE Transactions On Aerospace
And Electronic Systems,38(3):791-811,2002.
5. Y.Boers, J.N.Dreissen, Multitarget particle filter track before detect application. IEE Proceedings online no.
6. 20040841, 351-357,2004.
7. Khan Z., Balch T and Dellaert F. An MCMC-Based Particle Filter for Tracking Multiple Interacting Targets
8. Lecture Notes in Computer ScienceSpringer,3024:pp. 279-290, 2004.
9. Isard M. and Blake A. , Condensation-Conditional Density Propagation for Visual Tracking. International
Journal of Computer Vison,29(1):5-28,1998.