An Integrated Feature Selection Strategy For Monocular...

32
An Integrated Feature Selection Strategy For Monocular SLAM Presenter: Kuan-Ting Yu Advisor: Li-Chen Fu NTU Intelligent Robot Lab Industrial Technology Research Institute September, 2011 1

Transcript of An Integrated Feature Selection Strategy For Monocular...

Page 1: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

An Integrated Feature Selection

Strategy For Monocular SLAM

Presenter: Kuan-Ting Yu

Advisor: Li-Chen Fu

NTU Intelligent Robot Lab

Industrial Technology Research Institute

September, 2011

1

Page 2: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Outline

Introduction

Feature Extraction

Bottom-up Feature Selection

Top-down Feature Selection

Experimental Results

Conclusions

2

Page 3: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Outline

Introduction

Feature Extraction

Bottom-up Feature Selection

Top-down Feature Selection

Experimental Results

Conclusions

3

Page 4: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Introduction

To perform desired tasks in indoor environments,

how to estimate robot’s position and map about

surroundings is a critical problem.

4

Page 5: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

SLAM

Simultaneous localization and mapping (SLAM)

Incrementally build a map of this environment while

simultaneously determining its location

Given:

1. Robot’s odometry

2. Observations of nearby features

Estimate:

1. Location of the robot

2. Map of features

5

Page 6: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Sensors

6

Page 7: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

vSLAM

Generic SLAM

Landmark extraction

Data association

State predict

State update &

landmark update

Vision-based SLAM

Interest points (features)

Interest points matching

7

Probabilistic framework

Page 8: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

vSLAM Challenge

Huge amount of features

Degrade performance

Cause mismatch

8

Page 9: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Solution Idea

Feature selection strategy

Remove useless features

Keep good features

9

Page 10: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

System Overview

1. Landmark extraction

2. Data association

3. State predict

4. State update &

landmark update

10

Page 11: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Outline

Introduction

Feature Extraction

Bottom-up Feature Selection

Top-down Feature Selection

Experimental Results

Conclusions

11

Page 12: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

What is Feature?

Features (interest points):

Easy to find their correspondences

Desired features:

Distinctive: outstanding, easily matched

Invariant: invariant to scale, viewpoint change

12

?

Page 13: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Detectors

Scale Invariant Feature Transform (SIFT) [Lowe, 2004] is

one of the best feature detectors,

Speeded Up Robust Features (SURF) [Bay, et al., 2006]

SURF has good performance as SIFT and faster

SIFT SURF

Images

Extraction time 863.998 ms 267.634 ms

13

but slow

Page 14: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Extraction Challenge

Hundreds of SURF points are extracted

Increase computational time

Need higher data storage

But, fewer features are desired in practice

Remove useless features

Keep repeatable features

A feature selection strategy is necessary

14

Page 15: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Selection

Human never process the whole image at once

Focus on some regions of interest (ROIs)

A natural solution — visual attention system

15

Bottom-up approach Top-down approach

Image-driven Knowledge-driven

Page 16: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Outline

Introduction

Feature Extraction

Bottom-up Feature Selection

Top-down Feature Selection

Experimental Results

Conclusions

16

Page 17: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Selection – Bottom-up Approach

Visual attention system

Saliency model

[Itti et al., 1998]1

Replace the original RGB

color space with CIE L*a*b

Mimic color opponencies

of human vision

17

[1] L. Itti, C. Koch, and E. Niebur, “A Model of Saliency-Based Visual Attention for Rapid Scene Analysis,” PAMI, 1998.

Page 18: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Selection – Bottom-up Approach

Results of modified saliency model

18

1( ( ))3

Sal N C I O

Page 19: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Outline

Introduction

Feature Extraction

Bottom-up Feature Selection

Top-down Feature Selection

Experimental Results

Conclusions

19

Page 20: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Selection - Top down approach

Sometimes, bottom-up ROIs are not enough

For example:

Integrates top-down approach to achieve flexible

feature selection

20

Page 21: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Selection - Top down approach

Human robot interaction (HRI) can be applied to

object learning

Communication with the robot

Pointing with intelligent devices

21

Page 22: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Selection - Top down approach

Robot redetects known objects process

Based on an object

recognition algorithm

RANSAC

(Reject inconsistent matches)

Compute Homography

22

'

'

1 1

i i

i i i

x x

s y H y

1 1,x y

1 1', 'x y

2 2,x y

2 2', 'x y

H

=

Page 23: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Solving the homography matrix, then we project the

four end-points to determine top-down ROI

23

Feature Selection - Top down approach

Page 24: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Feature Selection

Merge two ROIs to obtain versatile features

24

Page 25: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Outline

Introduction

Feature Extraction

Bottom-up Feature Selection

Top-down Feature Selection

Experimental Results

Conclusions

25

Page 26: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Experiment Setting

Pioneer 3DX

Logitech webcam V-UBH44

26

Room size: 6m*8m

Page 27: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

SLAM Map

27

Page 28: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Localization Error Comparison

28

TDBU: the proposed integrated bottom-up and top-down selection TD: top-down BU: bottom-up NO: without selection

Variance

Average

Page 29: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Time Comparison

29

TDBU: the proposed integrated bottom-up and top-down selection TD: top-down BU: bottom-up NO: without selection

Page 30: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Matching Comparison 1/2

30

Page 31: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Matching Comparison 2/2

31

Page 32: An Integrated Feature Selection Strategy For Monocular ...people.csail.mit.edu/peterkty/pub/KuanTing_CVGIP2011...An Integrated Feature Selection Strategy For Monocular SLAM Presenter:

Conclusions

We propose an integrated feature selection strategy

based on visual attention system for bearing-only

SLAM with EKF

Reduce computation time to 62%

Reduce localization error to 89%

Combining bottom-up and top-down approach to

construct ROIs allow us to

Select salient and useful features

Improve data association

32