Please enable JavaScript.
Coggle requires JavaScript to display documents.
Components of Visual SLAM (Image Mesurements (Image alignment (Dense 3D…
Components of Visual SLAM
Image Mesurements
Point Features
sparse map
consisting of 3D points
descriptors
match score
select salient point with best score
using algorithms such as
SIFT
BRISK
ORB
map points
matched with points in frames
defines current pose of camera
Image alignment
Dense 3D map
e.g.
volumetric grid
synthesize image for given pose
whole image
align with incoming frame using pixel differences
defines current camera pose
dense matching
SLAM engine
Alternatives
Tracking and mapping
Tracking and mapping separate
tracking effectively visual odometry
Global optimization
Pose graph optimization
7D error vector
Local optimization
Local optimization over
map points
local key frame poses
Bundle Adjustment
solve with Levenberg-Marquardt algorithm
Pros and cons
Pros
allows dense mapping as well as feature based
Higher accuracy over large areas
Cons
Computationally demanding
Used on
wheeled robot
robot with plenty of power
Probabilistic filtering
tightly coupled tracking and mapping
extended Kalman Filter (EKF), Gaussian
Particle filtering (PF)
Multimodal PDF
Pros and Cons
Pros
Rigorous framework incorporating uncertainty
Cons
Good for local mapping with loopy motion
Performance degrades over large areas
large state vectors become impractical
used on
drones
platforms with limited power
Types of V-SLAM
Probabilistic filtering
EKF V-SLAM
MonoSLAM
Robust feature matching and relocalization
Tracking and mapping
PTAM
DTAM
feature-based
ORB-SLAM
descriptor
ORB
local optimization
global pose graph optimization
dense matching
LSD-SLAM
dense mapping
local optimization
global optimization