Ethan Eade & Tom Drummond <Scalable Monocular SLAM>

Scalable Monocular SLAM
Eade, E.   Drummond, T.  
Cambridge University;

This paper appears in: Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on
Publication Date: 17-22 June 2006
Volume: 1,  On page(s): 469- 476
ISSN: 1063-6919
ISBN: 0-7695-2597-0
Digital Object Identifier: 10.1109/CVPR.2006.263
Current Version Published: 2006-07-05

ScalableMonocularSLAM.pdf

 
Ethan Eade & Tom Drummond
Machine Intelligence Laboratory
the Division of Information Engineering at Cambridge University Engineering Department

ethan_eade_cvprtalk2006.pdf

monocular SLAM
particle filter + top-down search => real-time, large number  of landmarks

the first to apply this FastSLAM-type particle filter to single-camera SLAM

1. Introduction

SLAM = Simultaneous Localization and Mapping
: process of causally estimating both egomotion and structure in an online system

 SLAM using visual data in computer vision

SFM (= structure from motion): reconstructing scene geometry
+ causal or recursive estimation techniques

perspective-projection cameras

filtering methods to allow indirect observation models

Kalman filtering framework

Extended Kalman filter = EKF (-> to linearize the observation and dynamics models of the system)

causal estimation with recursive algorithms (cp. estimation depending only on observations up to the current time)
=> online operation (cp. SFM on global nonlinear optimization)

Davision's SLAM with a single camera
> EKF estimation framework
> top-down Bayesian estimation approach searching for landmarks in image regions constrained by estimate > uncertainty (instead of performing extensive bottom-up image processing and feature matching)
> Bayesian partial-initialization scheme for incorporating new landmarks
- cannot scale to large environment

EKF = the Extended Kalman filter
-  N*N covariace matrix for N landmarks
- updated with N*N computation cost

> SLAM system using a single camera as the only sensor
> frame-rate operation with many landmarks
> FastSLAM-style particle filter (the first use of such an approach in a monocular SLAM setting)
> top-down active search
> an efficient algorithm for discovering the depth of new landmarks that avoids linearization errors
> a novel method for using partially initialized landmarks to help constrain camera pose

FastSLAM
: based on the Rao-Blackwellized Particle Filter

2. Background

2.1 Scalable SLAM

> submap
bounded complexity -> bounded computation and space requirements

Montemerlo & Thrun
If the entire camera motion is known then the estimates of the positions of different landmarks become independent of each other.







Rao-Blackwellized Particle Filter



ZNCC = the Zero mean Normalized Cross-Correlation function epipolar constraint


epipolar constraint

http://en.wikipedia.org/wiki/Epipolar_geometry