MultiSwarm Optimization and Application for Obect Tracking

1
Multi-Swarm Optimization and Application for
Obect Tracking

• ??? ???

2008.9.25
2
Outlines

• Overview for PSO
• A major drawback of PSO
• The basic ideas
• A Mutli-Swarm Approach
• Object Tracking by PSO
• Related papers
• Update of local/global information
• Re-initialization
• Simulation Experimental Result

3
Overview

• Introduced firstly by Russel Ebenhart (an
  Electrical Engineer) and James Kennedy (a Social
  Psychologist) in 1995
• It is a population-based search method, i.e. it
  moves from a set of points (particles positions)
  to another set of points with likely improvement
  in one iteration (move).
• The three values that effect the new search
  direction, namely, current motion, particle own
  memory, and swarm influence, are incorporated via
  a summation approach with three weight factors.

4
Initialization. Positions and velocities
5
Basic PSO Algorithm (1/5)

• Particle Description each particle has three
  features
• Position (this is the ith particle at time k,
  notice vector notation)
• Velocity (similar to search direction, used
  to update the position)
• Fitness or objective (determines which
  particle has the best value in the swarm and also
  determines the best position of each particle
  over time.

6
Basic PSO Algorithm (2/5)

• Initial Swarm
• No well established guidelines for swarm size.
• Initial particles are randomly distributed across
  the design space.
• where xmin and xmax are vectors of lower and
  upper limit values respectively, and rand is a
  uniformly distributed rand variable between 0 and
  1.
• Initial velocity is randomly generated.
• where is the time increment.

7
Basic PSO Algorithm (3/5)

• Velocity Update
• provides search directions
• Includes deterministic and probabilistic
  parameters.
• Combines effect of current motion, particle own
  memory,and swarm influence.

8
Basic PSO Algorithm (4/5)

• Position Update
• Position is updated by velocity vector.

9
Basic PSO Algorithm (5/5)

• Stopping Criteria
• Maximum change in best fitness smaller than
  specified tolerance for a specified number of
  moves (S).

10
Constraint Handling

• Side Constraints ( /- Vmax)
• Velocity vectors can drive particles to
  explosion.
• Upper and lower variable limits can be treated as
  regular constraints.
• Particles with violated side constraints could be
  reset to the nearest limit.
• Functional Constraints
• gi is the penalty function to violate the i-th
  constraint .

11
A major Drawback of PSO

• The swarm may prematurely converge.
• Modification of the Velocity Update Equation
• the inertia weight (from 1.4 to 0.4 over
  iterations)
• Constriction Factor
• Side constraints (/- Vmax)
• Adding new term to velocity update equation (move
  towards a particle in its neighborhood)
• 90 particles follow the leader, However, the
  leader is not good enough, 10 particles cross
  over with the leader.
• LBest Model vs. GBest Model
• Neighborhood topologies
• Multi-swarm (subpopulation)

12
Goble Neighbourhood (GBest Model)
Global
13
Neighbourhoods (LBest Model)
14
Neighborhood topologies

• There have been two basic topologies used in PSO
• Ring Topology (neighborhood of 3)
• Star Topology (global neighborhood)

15
Multi-swarm Two Swarms
Swarm 2
Swarm 1
GB1
Solution Space
16
The basic idea

• Cooperative Swarms
• All swarms search the global optimum.
• Swarm Diversity (to position different swarms on
  different peaks)
• How to determine the initial swarms ?
• Survival of the fittest
• A particle with the higher fitness value has
  higher probability to be selected for continuous
  flying in each swarm.
• To determine the new size of swarms which is
  based on the average fitness value of particles
  in each swarm.

17
A Multi-Swarm Approach

• Initial
• Based on some criteria to determine the particles
  in each swarm.
• Select
• A particle with the higher fitness value has
  higher probability to be selected
• Propagate
• Velocity and Position Update
• Observe
• Compute the fitness value and its weight for each
  particle.
• Estimate
• Compute the average fitness value for each swarm.
• Determine the new size in each swarm based on the
  principle Survival of the fittest
• Add new particles if the new size is larger than
  previous one.

18
Initial
?

• Set t 0,generate N particles from initial
  distribution.
• Set K2, determine two swarms(Swarm1 and Swarm2 )
  based on following methods.
• (1) Random selected
• (2) Side constraints /- Smax
• (3) Some criteria
• Problem-dependent constraints
• is obtained by computing the weight
  (fitness value), and normalize.

19
Select
?

• For each swarm j
• Select particles
  from according to

20
Propagate
?

• propagate each particles from to
  by velocity and position update
  in PSO.

21
Observe
?

• Evaluate the fitness function i1, ,
• Calculate the weight
  
• for each particle in the swarmj
• and construct the observation set.

22
Estimate
?

• Calculate the average weight for each swarm j,
  and normalize.
• Determine the new size in each swarm j

23
Related Papers

• 2006 Conference papers
• SWARMTRACK A Particle Swarm Approach to Visual
  Tracking (Span)
• Real-time Evolving Swarms for Rapid Pattern
  Detection and Tracking (USA)
• 2007 Conference papers
• Adaptive Object Tracking using Particle Swarm
  Optimization (USA)
• Real Time Object Tracking on Video Image Sequence
  Using Particle Swarm Optimization (Osaka
  Prefecture University, Japan) appeared in ICCAS
  2007 (Control, Automation and Systems)

24
Object Tracking - Assumption

• Extension for video image processing
• Assumption
• one object with high evaluation values
• the object dont change greatly between two
  consecutive images
• the maximum and minimum evaluation value is
  known.
• Condition
• gk is updated with only the position that
  satisfies the following condition in the next
  freeze-frame picture.
• Tracking algorithm is based on update of
  local/global information and re-initialization.

25
Object Tracking Illustration (1/5)
26
Object Tracking Illustration (2/5)
27
Object Tracking Illustration (3/5)
28
Object Tracking Illustration (4/5)
29
Object Tracking Illustration (5/5)
30
Object Tracking Algorithm (1/3)
31
Object Tracking Algorithm (2/3)
32
Object Tracking Algorithm (3/3)
33
Simulation (1/3)
34
Simulation (2/3)
35
Simulation (3/3)
36
Experimental Result
37
Future work

• Object detection and estimation
• Objects may get occluded and disappear from the
  image
• How to determine a suitable fitness function ?
• Dynamic environment
• Backgrounds
• Lighting
• Camera moving
• Real-time performance
• Object prediction
• Moving vector (Velocity)
• Multi-objects tracking
• A multi-swarm approach