![](https://csdnimg.cn/release/download_crawler_static/33813083/bg4.jpg)
4
Table I: Summary of the most representative visual (top) and visual-inertial (bottom) systems, in chronological order.
SLAM
or VO
Pixels
used
Data
association
Estimation
Relocali-
zation
Loop
closing
Multi Maps
Mono
Stereo
Mono IMU
Stereo IMU
Fisheye
Accuracy
Robustness
Open source
Mono-SLAM
[13], [14]
SLAM
Shi
Tomasi
Correlation EKF - - - X - - - - Fair Fair [15]
1
PTAM
[16]–[18]
SLAM FAST
Pyramid
SSD
BA Thumbnail - - X - - - -
Very
Good
Fair [19]
LSD-SLAM
[20], [21]
SLAM Edgelets Direct PG -
FABMAP
PG
- X X - - - Good Fair [22]
SVO [23], [24] VO
FAST+
Hi.grad.
Direct
Local
BA
- - - X X - - X
Very
Good
Very
Good
[25]
2
ORB-SLAM2
[2], [3]
SLAM ORB Descriptor
Local
BA
DBoW2
DBoW2
PG+BA
- X X - - - Exc.
Very
Good
[26]
DSO [27]–[29] VO
High
grad.
Direct
Local
BA
- - - X X - - X Fair
Very
Good
[30]
DSM [31] SLAM
High
grad.
Direct
Local
BA
- - - X - - - -
Very
Good
Very
Good
[32]
MSCKF
[33]–[36]
VO
Shi
Tomasi
Cross
correlation
EKF - - - X - X X - Fair
Very
Good
[37]
3
OKVIS
[38], [39]
VO BRISK Descriptor
Local
BA
- - - - - X X X Good
Very
Good
[40]
ROVIO
[41], [42]
VO
Shi
Tomasi
Direct EKF - - - - - X X X Good
Very
Good
[43]
ORBSLAM-VI
[4]
SLAM ORB Descriptor
Local
BA
DBoW2
DBoW2
PG+BA
- X - X - -
Very
Good
Very
Good
-
VINS-Fusion
[7], [44]
VO
Shi
Tomasi
KLT
Local
BA
DBoW2
DBoW2
PG
X - X X X X Good Exc. [45]
VI-DSO [46] VO
High
grad.
Direct
Local
BA
- - - - - X - -
Very
Good
Exc. -
BASALT
[47]
VO FAST
KLT
(LSSD)
Local
BA
-
ORB
BA
- - - - X X
Very
Good
Exc. [48]
Kimera [8] VO
Shi
Tomasi
KLT
Local
BA
-
DBoW2
PG
- - - - X - Good Exc. [49]
ORB-SLAM3
(ours)
SLAM ORB Descriptor
Local
BA
DBoW2
DBoW2
PG+BA
X X X X X X Exc. Exc. [5]
1
Last source code provided by a different author. Original software is available at [50].
2
Source code available only for the first version, SVO 2.0 is not open source.
3
MSCKF is patented [51], only a re-implementation by a different author is available as open source.
B. Visual-Inertial SLAM
The combination of visual and inertial sensors provide
robustness to poor texture, motion blur and occlusions, and
in the case of monocular systems, make scale observable.
Research in tightly coupled approaches can be traced back
to MSCKF [33] where the EKF quadratic cost in the number
of features is avoided by feature marginalization. The initial
system was perfected in [34] and extended to stereo in [35],
[36]. The first tightly coupled visual odometry system based
on keyframes and bundle adjustment was OKVIS [38], [39]
that is also able to use monocular and stereo vision. While
these systems rely on features, ROVIO [41], [42] feeds an
EFK with photometric error using direct data association.
ORB-SLAM-VI [4] presented for the first time a visual-
inertial SLAM system able to reuse a map with short-term,
mid-term and long-term data association, using them in an
accurate local visual-inertial BA based on IMU preintegration
[60], [61]. However, its IMU initialization technique was
too slow, taking 15 seconds, which harmed robustness and
accuracy. Faster initialization techniques were proposed in
[62], [63], based on a closed-form solution to jointly retrieve
scale, gravity, accelerometer bias and initial velocity, as well
as visual features depth. Crucially, they ignore IMU noise
properties, and minimize the 3D error of points in space,
and not their reprojection errors, that is the gold standard in
feature-based computer vision. Our previous work [64] shows
that this results in large unpredictable errors.
VINS-Mono [7] is a very accurate and robust monocular-
inertial odometry system, with loop closing that uses DBoW2
and 4 DoF pose-graph optimization, and map-merging. Fea-
ture tracking is performed with Lucas-Kanade tracker, being
slightly more robust than descriptor matching. In VINS-Fusion
[44] it has been extended to stereo and stereo-inertial.
VI-DSO [46] extends DSO to visual-inertial odometry,
proposing a bundle adjustment that combines inertial obser-
vations with the photometric error of selected high gradient
pixels, what renders very good accuracy. As the information
from high gradient pixels is successfully exploited, the robust-
ness in scene regions with poor texture is also boosted. Their
initialization method relies on visual-inertial BA and takes 20-
30 seconds to converge within 1% scale error.
The recent BASALT [47] is a stereo-inertial odometry
system that extracts non-linear factors from visual-inertial
odometry to use them in BA, and closes loops matching ORB
features, achieving very good to excellent accuracy. Kimera [8]