1. Vision-based SLAM Enhanced by Particle Swarm Optimization on the Euclidean Group
Vision seminar : Dec
Young Ki BAIK
Computer Vision Lab.

2. Single camera SLAM using ABC algorithm
Outline
Introduction
Related works
Problem statement
Proposed algorithm
PSO-based visual SLAM
Single camera SLAM using ABC algorithm
Demonstration
Conclusion

3. SLAM : Simultaneous Localization And Mapping
What is SLAM?
SLAM : Simultaneous Localization And Mapping

4. Laser rangefinders, Sonar sensors
Why visual SLAM?
To acquire observation data
Use many different type of sensor
Laser rangefinders, Sonar sensors
Too expensive : about 2000$
Scanning system : complex mechanics
Camera
Low price : about 30$
Acquire large and meaningful information
from one shot measure

5. How to solve SLAM problem?
Solved by filtering approaches
Extended Kalman Filter (EKF)
has scalability problem of the map
Rao-Blackwellised Particle Filter (RBPF)
handles nonlinear and non-Gaussian
reduces computation cost by decomposing sampling space

6. Previous works EKF-based visual SLAM RBPF-based visual SLAM
Andrew Davison (1998)
Stereo camera + odometry
Andrew Davison (2002)
Single camera without odometry
RBPF-based visual SLAM
Robert Sim (2005)
Stereo camera + odometry
Mark Pupilli (2005)
Single camera without odometry

7. RBPF-SLAM State equation Measurement equation (Process noise)
(User input or odometry)
(Process noise)
(Nonlinear stochastic difference equation)
Measurement equation
(Camera projection function)
(Measurement noise)

8. ? + Problem of RBPF-SLAM How to choose importance function? t+1 t
Odometry
Naive motion model
Constant position
Xt+1=Xt+N
Angle Change +
Distance Change
Left Encoder Distance
Right Encoder Distance
Constant velocity
Xt+1=Xt+∇t(Vt+N)

9. Problem of RBPF-SLAM Sampling by transition model t Landmark Particle
Robot t

10. Problem of RBPF-SLAM
Sampling by transition model t
t+1

11. Problem of RBPF-SLAM
Sampling by transition model t
t+1

12. Problem of RBPF-SLAM
Sampling by transition model
t+1
(Gaussian)

13. Problem of RBPF-SLAM
Sampling by transition model
t+1

14. ? Problem of RBPF-SLAM How to choose importance function?
Hand-held camera case ? t
t+1

15. RBPF-SLAM
Sampling by transition model t
t+1

16. Problem of RBPF-SLAM Particle impoverishment
Mismatch between proposal and likelihood distribution.
Likelihood
Proposal

17. Optimal Importance Function (OIF)
For better proposal distribution
Use observation for proposal distribution
Optimal importance function approach
(Doucet et al., 2000)
Observation incorporated proposal
Linearize the optimal importance function
Used in FastSLAM 2.0 (Montemerlo et al.)
The state of the art !!

18. Optimal Importance Function (OIF)
Sampling by optimal importance function
OIF t
t+1

19. Problem of OIF-based SLAM
Linearization Error
Smooth camera motion
Linearization Error
Abrupt camera motion
: Real camera state
: Estimated camera state by linearization
: Predicted camera state by a motion model

20. Problem statement OIF-based visual SLAM State of the art
Weak to abrupt camera motion
Novel visual SLAM
robust to abrupt camera motion

21. Target Proposed SLAM system 6-DOF SLAM Hand-held camera
Single or stereo camera
No odometry
RBPF-based SLAM
Robust to sudden changes
Real-time system

22. Our contribution We propose … Robust to abrupt camera motion!!
Novel particle filtering framework
combined with geometric PSO
Based on special Euclidean group SE (3)
Reformulating original PSO in consideration of SE (3)
Applying Quantum particles
to more actively explore the problem space
Robust to abrupt camera motion!!

23. Special Euclidean group SE (3)
Conventional
Geometric
State
6-D vector
by local coordinates
as a Lie group SE(3)
State Equation
Ignores geometry of the underlying space
Considers geometry of the curved space!

24. Special Euclidean group SE (3)
6D vector  Euclidean group SE(3)
Lie group  Group + Differentiable manifold
Lie algebra  Tangent space at the identity (se(3))
Origin
se(3)
Exp
Log
Exp: se(3)  SE(3)
Log: SE(3)  se(3)
Identity
SE(3)

25. Special Euclidean group SE (3)
6D vector  Euclidean group SE(3)
Sampling on Tangent space at the identity (se(3))
Reasonable to consider the geometry of motion
Sampling
Exp
se(3)
SE(3)

26. Main idea We use optimization method for better proposal distribution…
Particle Swarm Optimization
Prior
Propagate particles using motion prior
PSO Moves
Particles
with high likelihood

27. Particle Swarm Optimization
Developed in evolutionary computation community
Sampling-based optimization method
Uses the relationship between particles
PSO
OIF
Interaction
Linearization

28. Particle Swarm Optimization
Particle from motion prior

29. Particle Swarm Optimization
Initialization
(current optimum)
(individual best)

30. Particle Swarm Optimization
Particle from motion prior
(current optimum)
(individual best)

31. Particle Swarm Optimization
Particle from motion prior
(current optimum)
(individual best)
(Inertia)
(Coefficient)
(Random)

32. Particle Swarm Optimization
Velocity updating
(current optimum)
(individual best)

33. Particle Swarm Optimization
Moving
(current optimum)
(individual best)

34. Particle Swarm Optimization
Global and local best updating
(current optimum)
(individual best)

35. Particle Swarm Optimization
For all Particles

36. Geometric Particle Swarm Optimization
Tangent space at
Manifold
Random perturbation &
coefficient multiplication

37. Bumblebee stereo camera
Experiments
System environment
CPU : Intel Core-2 Quad 2.4 GHz process
Real-time with C++ implementation
Synthetic sequence
Real sequence
Virtual stereo camera
Bumblebee stereo camera
(BB-HICOL-60)
Quantitative analysis

38. Demonstration

39. Demonstration

40. Artificial Bee Colony Additional work !! Visual Odometry
Determining the position and orientation of a robot
by analyzing the associated camera images …
David Nister (2004)
Monocular or binocular camera
Yang Cheng et al. (2008)
Stereo camera

41. Propagate particles using motion prior
Artificial Bee Colony
Additional work !!
Propagate particles
via visual odometry
Propagate particles using motion prior
PSO Moves
PSO Moves
Particles
with high likelihood
Artificial Bee Colony

42. Conclusion Novel visual SLAM is presented !!
RBPF based on the special Euclidean group SE (3)
Geometric Particle Swarm Optimization
Robust to abrupt camera motion
Real-time system
Novel monocular SLAM will be presented !!
Geometric Artificial Bee Colony
Combined proposal ( VO + Naive motion model )

43. Q & A