Orb Slam

This article has been accepted for inclusion in a future issue of this journal.
Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON ROBOTICS
ORB-SLAM: A Versatile and Accurate Monocular

SLAM System
Raul Mur-Artal, J. M. M. Montiel, Member, IEEE, and Juan D. Tardos, Member, IEEE
AbstractThis paper presents ORB-SLAM, a feature-based

monocular simultaneous localization and mapping (SLAM) system
that operates in real time, in small and large indoor and outdoor
environments. The system is robust to severe motion clutter, allows
wide baseline loop closing and relocalization, and includes full automatic initialization. Building on excellent algorithms of recent
years, we designed from scratch a novel system that uses the same
features for all SLAM tasks: tracking, mapping, relocalization, and
loop closing. A survival of the fittest strategy that selects the points
and keyframes of the reconstruction leads to excellent robustness
and generates a compact and trackable map that only grows if
the scene content changes, allowing lifelong operation. We present
an exhaustive evaluation in 27 sequences from the most popular
datasets. ORB-SLAM achieves unprecedented performance with
respect to other state-of-the-art monocular SLAM approaches. For
the benefit of the community, we make the source code public.
Index TermsLifelong mapping, localization, monocular vision,
recognition, simultaneous localization and mapping (SLAM).
I. INTRODUCTION
UNDLE adjustment (BA) is known to provide accurate
estimates of camera localizations as well as a sparse geometrical reconstruction [1], [2], given that a strong network of
matches and good initial guesses are provided. For a long time,
this approach was considered unaffordable for real-time applications such as visual simultaneous localization and mapping
(visual SLAM). Visual SLAM has the goal of estimating the
camera trajectory while reconstructing the environment. Now,
we know that to achieve accurate results at nonprohibitive computational cost, a real-time SLAM algorithm has to provide BA
with the following.
1) Corresponding observations of scene features (map
points) among a subset of selected frames (keyframes).
2) As complexity grows with the number of keyframes, their
selection should avoid unnecessary redundancy.
3) A strong network configuration of keyframes and points
to produce accurate results, that is, a well spread set of
keyframes observing points with significant parallax and
with plenty of loop closure matches.
Manuscript received April 28, 2015; accepted July 27, 2015. This paper was
recommended for publication by Associate Editor D. Scaramuzza and Editor
D. Fox upon evaluation of the reviewers comments. This work was supported
by the Direccion General de Investigacion of Spain under Project DPI201232168, the Ministerio de Educacion Scholarship FPU13/04175, and Gobierno
de Aragon Scholarship B121/13.
The authors are with Instituto de Investigacion en Ingeniera de Aragon (I3A),
Universidad de Zaragoza, 50018 Zaragoza, Spain (e-mail: [email protected];
[email protected]; [email protected]).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TRO.2015.2463671
4) An initial estimation of the keyframe poses and point

locations for the nonlinear optimization.
5) A local map in exploration where optimization is focused
to achieve scalability.
6) The ability to perform fast global optimizations (e.g., pose
graph) to close loops in real time.
The first real-time application of BA was the visual odometry
work of Mouragon et al. [3], followed by the ground-breaking
SLAM work of Klein and Murray [4], known as parallel tracking and mapping (PTAM). This algorithm, while limited to
small-scale operation, provides simple but effective methods
for keyframe selection, feature matching, point triangulation,
camera localization for every frame, and relocalization after
tracking failure. Unfortunately, several factors severely limit its
application: lack of loop closing and adequate handling of occlusions, low invariance to viewpoint of the relocalization, and
the need of human intervention for map bootstrapping.
In this study, we build on the main ideas of PTAM, the place
recognition work of Galvez-Lopez and Tardos [5], the scaleaware loop closing of Strasdat et al. [6], and the use of covisibility information for large-scale operation [7], [8], to design
from scratch ORB-SLAM, i.e., a novel monocular SLAM system whose main contributions are as follows.
1) Use of the same features for all tasks: tracking, mapping,
relocalization, and loop closing. This makes our system
more efficient, simple, and reliable. We use ORB features
[9], which allow real-time performance without GPUs,
providing good invariance to changes in viewpoint and
illumination.
2) Real-time operation in large environments. Thanks to the
use of a covisibility graph, tracking and mapping are focused in a local covisible area, independent of global map
size.
3) Real-time loop closing based on the optimization of a pose
graph that we call the Essential Graph. It is built from
a spanning tree maintained by the system, loop closure
links, and strong edges from the covisibility graph.
4) Real-time camera relocalization with significant invariance to viewpoint and illumination. This allows recovery
from tracking failure and also enhances map reuse.
5) A new automatic and robust initialization procedure based
on model selection that permits to create an initial map of
planar and nonplanar scenes.
6) A survival of the fittest approach to map point and
keyframe selection that is generous in the spawning but
very restrictive in the culling. This policy improves tracking robustness and enhances lifelong operation because
redundant keyframes are discarded.
1552-3098 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2
We present an extensive evaluation in popular public datasets

from indoor and outdoor environments, including hand-held,
car, and robot sequences. Notably, we achieve better camera
localization accuracy than the state of the art in direct methods [10], which optimize directly over pixel intensities instead
of feature reprojection errors. We include a discussion in Section IX-B on the possible causes that can make feature-based
methods more accurate than direct methods.
The loop closing and relocalization methods here presented
are based on our previous work [11]. A preliminary version of
the system was presented in [12]. In the current paper, we add
the initialization method, the Essential Graph, and perfect all
methods involved. We also describe in detail all building blocks
and perform an exhaustive experimental validation.
To the best of our knowledge, this is the most complete and
reliable solution to monocular SLAM, and for the benefit of the
community, we make the source code public. Demonstration
videos and the code can be found in our project webpage.1
II. RELATED WORK
A. Place Recognition
The survey by Williams et al. [13] compared several approaches for place recognition and concluded that techniques
based on appearance, that is, image-to-image matching, scale
better in large environments than map-to-map or image-tomap methods. Within appearance-based methods, bags of words
techniques [14], such as the probabilistic approach FAB-MAP
[15], are to the fore because of their high efficiency. DBoW2 [5]
used for the first time bags of binary words obtained from BRIEF
descriptors [16] along with the very efficient FAST feature detector [17]. This reduced in more than one order of magnitude
the time needed for feature extraction, compared with SURF
[18] and SIFT [19] features that were used in bags of words
approaches so far. Although the system demonstrated to be very
efficient and robust, the use of BRIEF, neither rotation nor scale
invariant, limited the system to in-plane trajectories and loop detection from similar viewpoints. In our previous work [11], we
proposed a bag of words place recognizer built on DBoW2 with
ORB [9]. ORB are binary features invariant to rotation and scale
(in a certain range), resulting in a very fast recognizer with good
invariance to viewpoint. We demonstrated the high recall and
robustness of the recognizer in four different datasets, requiring
less than 39 ms (including feature extraction) to retrieve a loop
candidate from a 10 K image database. In this study, we use an
improved version of that place recognizer, using covisibility information and returning several hypotheses when querying the
database instead of just the best match.
B. Map Initialization
Monocular SLAM requires a procedure to create an initial
map because depth cannot be recovered from a single image.
One way to solve the problem is to initially track a known
structure [20]. In the context of filtering approaches, points can
1 http://webdiis.unizar.es/raulmur/orbslam
be initialized with high uncertainty in depth using an inverse

depth parameterization [21], which hopefully will later converge
to their real positions. The recent semidense work of Engel
et al. [10] follows a similar approach initializing the depth of
the pixels to a random value with high variance.
Initialization methods from two views either assume locally
scene planarity [4], [22] and recover the relative camera pose
from a homography using the method of Faugeras and Lustman
[23], or compute an essential matrix [24], [25] that models planar and general scenes, using the five-point algorithm of Nister
[26], which requires to deal with multiple solutions. Both reconstruction methods are not well constrained under low parallax
and suffer from a twofold ambiguity solution if all points of a
planar scene are closer to one of the camera centers [27]. On the
other hand, if a nonplanar scene is seen with parallax, a unique
fundamental matrix can be computed with the eight-point algorithm [2], and the relative camera pose can be recovered without
ambiguity.
In Section IV, we present a new automatic approach based
on model selection between a homography for planar scenes
and a fundamental matrix for nonplanar scenes. A statistical
approach to model selection was proposed by Torr et al. [28].
Under a similar rationale, we have developed a heuristic initialization algorithm that takes into account the risk of selecting
a fundamental matrix in close to degenerate cases (i.e., planar,
nearly planar, and low parallax), favoring the selection of the
homography. In the planar case, for the sake of safe operation,
we refrain from initializing if the solution has a twofold ambiguity, as a corrupted solution could be selected. We delay the
initialization until the method produces a unique solution with
significant parallax.
C. Monocular Simultaneous Localization and Mapping
Monocular SLAM was initially solved by filtering [20], [21],
[29], [30]. In that approach, every frame is processed by the filter to jointly estimate the map feature locations and the camera
pose. It has the drawbacks of wasting computation in processing
consecutive frames with little new information and the accumulation of linearization errors. On the other hand, keyframe-based
approaches [3], [4] estimate the map using only selected frames
(keyframes) allowing to perform more costly but accurate BA
optimizations, as mapping is not tied to frame rate. Strasdat
et al. [31] demonstrated that keyframe-based techniques are
more accurate than filtering for the same computational cost.
The most representative keyframe-based SLAM system is
probably PTAM by Klein and Murray [4]. It was the first work
to introduce the idea of splitting camera tracking and mapping
in parallel threads and demonstrated to be successful for realtime augmented reality applications in small environments. The
original version was later improved with edge features, a rotation estimation step during tracking, and a better relocalization
method [32]. The map points of PTAM correspond to FAST corners matched by patch correlation. This makes the points only
useful for tracking but not for place recognition. In fact, PTAM
does not detect large loops, and the relocalization is based on
the correlation of low-resolution thumbnails of the keyframes,
yielding a low invariance to viewpoint.
MUR-ARTAL et al.: ORB-SLAM: A VERSATILE AND ACCURATE MONOCULAR SLAM SYSTEM
Strasdat et al. [6] presented a large-scale monocular SLAM

system with a front-end based on optical flow implemented on a
GPU, followed by FAST feature matching and motion-only BA,
and a back-end based on sliding-window BA. Loop closures
were solved with a pose graph optimization with similarity constraints [7 degrees of freedom (DoF)], which was able to correct
the scale drift appearing in monocular SLAM. From this work,
we take the idea of loop closing with 7-DoF pose graph optimization and apply it to the Essential Graph defined in Section III-D.
Strasdat et al. [7] used the front-end of PTAM, but performed
the tracking only in a local map retrieved from a covisibility
graph. They proposed a double-window optimization back-end
that continuously performs BA in the inner window and pose
graph in a limited-size outer window. However, loop closing is
only effective if the size of the outer window is large enough
to include the whole loop. In our system, we take advantage
of the excellent ideas of using a local map based on covisibility and building the pose graph from the covisibility graph,
but apply them in a totally redesigned front-end and back-end.
Another difference is that, instead of using specific features for
loop detection (SURF), we perform the place recognition on the
same tracked and mapped features, obtaining robust frame-rate
relocalization and loop detection.
Pirker et al. [33] proposed CD-SLAM, i.e., a very complete
system including loop closing, relocalization, large-scale operation, and efforts to work on dynamic environments. However,
map initialization is not mentioned. The lack of a public implementation does not allow us to perform a comparison of
accuracy, robustness, or large-scale capabilities.
The visual odometry of Song et al. [34] uses ORB features
for tracking and a temporal sliding window BA back-end. In
comparison, our system is more general as they do not have
global relocalization, loop closing, and do not reuse the map.
They are also using the known distance from the camera to the
ground to limit monocular scale drift.
Lim et al. [25], work published after we submitted our preliminary version of this work [12], use also the same features for
tracking, mapping, and loop detection. However, the choice of
BRIEF limits the system to in-plane trajectories. Their system
only tracks points from the last keyframe; therefore, the map is
not reused if revisited (similar to visual odometry) and has the
problem of growing unbounded. We compare qualitatively our
results with this approach in Section VIII-E.
The recent work of Engel et al. [10], known as LSD-SLAM, is
able to build large-scale semidense maps, using direct methods
(i.e., optimization directly over image pixel intensities) instead
of BA over features. Their results are very impressive as the
system is able to operate in real time, without GPU acceleration, building a semidense map, with more potential applications for robotics than the sparse output generated by featurebased SLAM. Nevertheless, they still need features for loop detection, and their camera localization accuracy is significantly
lower than in our system and PTAM, as we show experimentally in Section VIII-B. This surprising result is discussed in
Section IX-B.
In a halfway between direct and feature-based methods is the
semidirect visual odometry SVO of Forster et al. [22]. With-
out requiring to extract features in every frame, they are able

to operate at high frame rates obtaining impressive results in
quadracopters. However, no loop detection is performed, and
the current implementation is mainly thought for downward
looking cameras.
Finally, we want to discuss about keyframe selection. All
visual SLAM works in the literature agree that running BA with
all the points and all the frames is not feasible. The work of
Strasdat et al. [31] showed that the most cost-effective approach
is to keep as much points as possible, while keeping only
nonredundant keyframes. The PTAM approach was to insert
keyframes very cautiously to avoid an excessive growth of the
computational complexity. This restrictive keyframe insertion
policy makes the tracking fail in hard exploration conditions.
Our survival of the fittest strategy achieves unprecedented robustness in difficult scenarios by inserting keyframes as quickly
as possible, and removing later the redundant ones, to avoid the
extra cost.
III. SYSTEM OVERVIEW
A. Feature Choice
One of the main design ideas in our system is that the same
features used by the mapping and tracking are used for place
recognition to perform frame-rate relocalization and loop detection. This makes our system efficient and avoids the need to interpolate the depth of the recognition features from near SLAM
features as in previous works [6], [7]. We require features that
need for extraction much less than 33 ms per image, which
excludes the popular SIFT ( 300 ms) [19], SURF ( 300 ms)
[18], or the recent A-KAZE ( 100 ms) [35]. To obtain general
place recognition capabilities, we require rotation invariance,
which excludes BRIEF [16] and LDB [36].
We chose ORB [9], which are oriented multiscale FAST corners with a 256-bit descriptor associated. They are extremely
fast to compute and match, while they have good invariance to
viewpoint. This allows us to match them with wide baselines,
boosting the accuracy of BA. We already shown the good performance of ORB for place recognition in [11]. While our current
implementation makes use of ORB, the techniques proposed are
not restricted to these features.
B. Three Threads: Tracking, Local Mapping, and
Loop Closing
Our system, see an overview in Fig. 1, incorporates three
threads that run in parallel: tracking, local mapping, and loop
closing. The tracking is in charge of localizing the camera with
every frame and deciding when to insert a new keyframe. We
perform first an initial feature matching with the previous frame
and optimize the pose using motion-only BA. If the tracking is
lost (e.g., due to occlusions or abrupt movements), the place
recognition module is used to perform a global relocalization.
Once there is an initial estimation of the camera pose and feature
matchings, a local visible map is retrieved using the covisibility
graph of keyframes that is maintained by the system [see
Fig. 2(a) and (b)]. Then, matches with the local map points are
4
Fig. 1. ORB-SLAM system overview, showing all the steps performed by the
tracking, local mapping, and loop closing threads. The main components of the
place recognition module and the map are also shown.
searched by reprojection, and camera pose is optimized again

with all matches. Finally, the tracking thread decides if a new
keyframe is inserted. All the tracking steps are explained in detail in Section V. The novel procedure to create an initial map
is presented in Section IV.
The local mapping processes new keyframes and performs local BA to achieve an optimal reconstruction in the surroundings
of the camera pose. New correspondences for unmatched ORB
in the new keyframe are searched in connected keyframes in
the covisibility graph to triangulate new points. Some time after
creation, based on the information gathered during the tracking, an exigent point culling policy is applied in order to retain
only high quality points. The local mapping is also in charge
of culling redundant keyframes. We explain in detail all local
mapping steps in Section VI.
The loop closing searches for loops with every new keyframe.
If a loop is detected, we compute a similarity transformation
that informs about the drift accumulated in the loop. Then, both
sides of the loop are aligned and duplicated points are fused.
Finally, a pose graph optimization over similarity constraints [6]
is performed to achieve global consistency. The main novelty is
that we perform the optimization over the Essential Graph, i.e.,
a sparser subgraph of the covisibility graph which is explained
in Section III-D. The loop detection and correction steps are
explained in detail in Section VII.
We use the LevenbergMarquardt algorithm implemented in
g2o [37] to carry out all optimizations. In the Appendix, we
describe the error terms, cost functions, and variables involved
in each optimization.
C. Map Points, Keyframes, and Their Selection
Each map point pi stores the following:
1) its 3-D position Xw ,i in the world coordinate system;
2) the viewing direction ni , which is the mean unit vector of all its viewing directions (the rays that join
Fig. 2. Reconstruction and graphs in the sequence f r3 long

of f ice household from the TUM RGB-D Benchmark [38]. (a) Keyframes
(blue), current camera (green), map points (black, red), current local map points
(red). (b) Covisibility graph. (c) Spanning tree (green) and loop closure (red).
(d) Essential graph.
the point with the optical center of the keyframes that

observe it);
3) a representative ORB descriptor Di , which is the associated ORB descriptor whose hamming distance is minimum with respect to all other associated descriptors in the
keyframes in which the point is observed;
4) the maximum dm ax and minimum dm in distances at which
the point can be observed, according to the scale invariance limits of the ORB features.
Each keyframe Ki stores the following:
1) the camera pose Tiw , which is a rigid body transformation that transforms points from the world to the camera
coordinate system;
2) the camera intrinsics, including focal length and principal
point;
3) all the ORB features extracted in the frame, associated or
not with a map point, whose coordinates are undistorted
if a distortion model is provided.
Map points and keyframes are created with a generous policy,
while a later very exigent culling mechanism is in charge of
detecting redundant keyframes and wrongly matched or not
trackable map points. This permits a flexible map expansion
during exploration, which boost tracking robustness under hard
conditions (e.g., rotations, fast movements), while its size is
bounded in continual revisits to the same environment, i.e.,
lifelong operation. Additionally, our maps contain very few
outliers compared with PTAM, at the expense of containing
less points. Culling procedures of map points and keyframes
are explained in Sections VI-B and VI-E, respectively.
D. Covisibility Graph and Essential Graph

Covisibility information between keyframes is very useful in
several tasks of our system and is represented as an undirected
weighted graph as in [7]. Each node is a keyframe, and an edge
between two keyframes exists if they share observations of the
same map points (at least 15), being the weight of the edge
the number of common map points.
In order to correct a loop, we perform a pose graph optimization [6] that distributes the loop closing error along the
graph. In order not to include all the edges provided by the covisibility graph, which can be very dense, we propose to build
an Essential Graph that retains all the nodes (keyframes), but
less edges, still preserving a strong network that yields accurate
results. The system builds incrementally a spanning tree from
the initial keyframe, which provides a connected subgraph of
the covisibility graph with minimal number of edges. When a
new keyframe is inserted, it is included in the tree linked to
the keyframe which shares most point observations, and when
a keyframe is erased by the culling policy, the system updates
the links affected by that keyframe. The Essential Graph contains the spanning tree, the subset of edges from the covisibility
graph with high covisibility (m in = 100), and the loop closure
edges, resulting in a strong network of cameras. Fig. 2 shows an
example of a covisibility graph, spanning tree, and associated
essential graph. As shown in the experiments of Section VIII-E,
when performing the pose graph optimization, the solution is
so accurate that an additional full BA optimization barely improves the solution. The efficiency of the essential graph and
the influence of the m in is shown at the end of Section VIII-E.
E. Bags of Words Place Recognition
The system has embedded a bags of words place recognition module, based on DBoW22 [5], to perform loop detection
and relocalization. Visual words are just a discretization of the
descriptor space, which is known as the visual vocabulary. The
vocabulary is created offline with the ORB descriptors extracted
from a large set of images. If the images are general enough, the
same vocabulary can be used for different environments getting
a good performance, as shown in our previous work [11]. The
system builds incrementally a database that contains an invert
index, which stores for each visual word in the vocabulary, in
which keyframes it has been seen, so that querying the database
can be done very efficiently. The database is also updated when
a keyframe is deleted by the culling procedure.
Because there exists visual overlap between keyframes, when
querying the database, there will not exist a unique keyframe
with a high score. The original DBoW2 took this overlapping
into account, adding up the score of images that are close in time.
This has the limitation of not including keyframes viewing the
same place but inserted at a different time. Instead, we group
those keyframes that are connected in the covisibility graph.
In addition, our database returns all keyframe matches whose
scores are higher than the 75% of the best score.
2 https://github.com/dorian3d/DBoW2
An additional benefit of the bags of words representation for

feature matching was reported in [5]. When we want to compute
the correspondences between two sets of ORB features, we can
constraint the brute force matching only to those features that
belong to the same node in the vocabulary tree at a certain level
(we select the second out of six), speeding up the search. We use
this trick when searching matches for triangulating new points,
and at loop detection and relocalization. We also refine the
correspondences with an orientation consistency test (see [11]
for details) that discards outliers ensuring a coherent rotation
for all correspondences.
IV. AUTOMATIC MAP INITIALIZATION
The goal of the map initialization is to compute the relative
pose between two frames to triangulate an initial set of map
points. This method should be independent of the scene (planar or general) and should not require human intervention to
select a good two-view configuration, i.e., a configuration with
significant parallax. We propose to compute in parallel two geometrical models: a homography assuming a planar scene and
a fundamental matrix assuming a nonplanar scene. We then use
a heuristic to select a model and try to recover the relative pose
with a specific method for the selected model. Our method only
initializes when it is certain that the two-view configuration is
safe, detecting low-parallax cases and the well-known twofold
planar ambiguity [27], avoiding to initialize a corrupted map.
The steps of our algorithm are as follows.
1) Find initial correspondences: Extract ORB features (only
at the finest scale) in the current frame Fc and search for
matches xc xr in the reference frame Fr . If not enough
matches are found, reset the reference frame.
2) Parallel computation of the two models: Compute in parallel threads a homography Hcr and a fundamental matrix
Fcr as
xc = Hcr xr ,
xTc Fcr xr = 0
(1)
with the normalized DLT and eight-point algorithms, respectively, as explained in [2] inside a RANSAC scheme.
To make homogeneous the procedure for both models,
the number of iterations is prefixed and the same for both
models, along with the points to be used at each iteration:
eight for the fundamental matrix, and four of them for
the homography. At each iteration, we compute a score
SM for each model M (H for the homography, F for the
fundamental matrix)

M d2cr (xic , xir , M )
SM =
i
M (d2 ) =

+ M (d2r c xic , xir , M )

d2 , if d2 < TM
0,
if d2 TM
(2)
where d2cr and d2r c are the symmetric transfer errors [2]
from one frame to the other. TM is the outlier rejection
threshold based on the 2 test at 95% (TH = 5.99,
TF = 3.84, assuming a standard deviation of 1 pixel in
6
the measurement error). is defined equal to TH so that

both models score equally for the same d in their inlier
region, again to make the process homogeneous.
We keep the homography and fundamental matrix with
the highest score. If no model could be found (not enough
inliers), we restart the process again from step 1.
3) Model selection: If the scene is planar, nearly planar or
there is low parallax, it can be explained by a homography.
However, a fundamental matrix can also be found, but the
problem is not well constrained [2], and any attempt to
recover the motion from the fundamental matrix would
yield wrong results. We should select the homography as
the reconstruction method will correctly initialize from a
plane or it will detect the low parallax case and refuse the
initialization. On the other hand, a nonplanar scene with
enough parallax can only be explained by the fundamental
matrix, but a homography can also be found explaining a
subset of the matches if they lie on a plane or they have
low parallax (they are far away). In this case, we should
select the fundamental matrix. We have found that a robust
heuristic is to compute
RH =
SH
SH + SF
(3)
and select the homography if RH > 0.45, which adequately captures the planar and low parallax cases. Otherwise, we select the fundamental matrix.
4) Motion and structure from motion recovery: Once a model
is selected, we retrieve the motion hypotheses associated.
In the case of the homography, we retrieve eight motion
hypotheses using the method of Faugeras and Lustman
[23]. The method proposes cheriality tests to select the
valid solution. However, these tests fail if there is low
parallax as points easily go in front or back of the cameras, which could yield the selection of a wrong solution.
We propose to directly triangulate the eight solutions and
check if there is one solution with most points seen with
parallax, in front of both cameras and with low reprojection error. If there is not a clear winner solution, we do
not initialize and continue from step 1. This technique to
disambiguate the solutions makes our initialization robust
under low parallax and the twofold ambiguity configuration and could be considered the key of the robustness of
our method.
In the case of the fundamental matrix, we convert it in an
essential matrix using the calibration matrix K as
Er c = KT Fr c K
(4)
and then retrieve four motion hypotheses with the singular

value decomposition method explained in [2]. We triangulate the four solutions and select the reconstruction as
done for the homography.
5) Bundle adjustment: Finally, we perform a full BA (see the
Appendix for details) to refine the initial reconstruction.
An example of a challenging initialization in the outdoor
NewCollege robot sequence [39] is shown in Fig. 3. It can be
seen how PTAM and LSD-SLAM have initialized all points
Fig. 3. Top: PTAM, middle: LSD-SLAM, bottom: ORB-SLAM, some time

after initialization in the NewCollege sequence [39]. PTAM and LSD-SLAM
initialize a corrupted planar solution, while our method has automatically initialized from the fundamental matrix when it has detected enough parallax.
Depending on which keyframes are manually selected, PTAM is also able to
initialize well.
in a plane, while our method has waited until there is enough

parallax, initializing correctly from the fundamental matrix.
V. TRACKING
In this section, we describe the steps of the tracking thread
that are performed with every frame from the camera. The camera pose optimizations, mentioned in several steps, consist in
motion-only BA, which is described in the Appendix.
A. ORB Extraction
We extract FAST corners at eight-scale levels with a scale factor of 1.2. For image resolutions from 512 384 to 752 480
pixels we found suitable to extract 1000 corners, for higher resolutions, as the 1241 376 in the KITTI dataset [40], we extract
2000 corners. In order to ensure an homogeneous distribution,
we divide each scale level in a grid, trying to extract at least five
corners per cell. Then, we detect corners in each cell, adapting the detector threshold if not enough corners are found. The
amount of corners retained per cell is also adapted if some cells
contains no corners (textureless or low contrast). The orientation
and ORB descriptor are then computed on the retained FAST
corners. The ORB descriptor is used in all feature matching, in
contrast with the search by patch correlation in PTAM.
B. Initial Pose Estimation From Previous Frame

If tracking was successful for last frame, we use a constant
velocity motion model to predict the camera pose and perform
a guided search of the map points observed in the last frame. If
not enough matches were found (i.e., motion model is clearly
violated), we use a wider search of the map points around their
position in the last frame. The pose is then optimized with the
found correspondences.
C. Initial Pose Estimation via Global Relocalization
If the tracking is lost, we convert the frame into bag of words
and query the recognition database for keyframe candidates for
global relocalization. We compute correspondences with ORB
associated with map points in each keyframe, as explained in
Section III-E. We then perform alternatively RANSAC iterations
for each keyframe and try to find a camera pose using the PnP
algorithm [41]. If we find a camera pose with enough inliers, we
optimize the pose and perform a guided search of more matches
with the map points of the candidate keyframe. Finally, the
camera pose is again optimized, and if supported with enough
inliers, tracking procedure continues.
D. Track Local Map
Once we have an estimation of the camera pose and an initial set of feature matches, we can project the map into the
frame and search more map point correspondences. To bound
the complexity in large maps, we only project a local map. This
local map contains the set of keyframes K1 , which share map
points with the current frame, and a set K2 with neighbors to the
keyframes K1 in the covisibility graph. The local map also has
a reference keyframe Kref K1 , which shares most map points
with current frame. Now, each map point seen in K1 and K2 is
searched in the current frame as follows.
1) Compute the map point projection x in the current frame.
Discard if it lays out of the image bounds.
2) Compute the angle between the current viewing ray v
and the map point mean viewing direction n. Discard if
v n < cos(60 ).
3) Compute the distance d from map point to camera center.
Discard if it is out of the scale invariance region of the
map point d
/ [dmin , dmax ].
4) Compute the scale in the frame by the ratio d/dmin .
5) Compare the representative descriptor D of the map point
with the still unmatched ORB features in the frame, at the
predicted scale, and near x, and associate the map point
with the best match.
The camera pose is finally optimized with all the map points
found in the frame.
E. New Keyframe Decision
The last step is to decide if the current frame is spawned as
a new keyframe. As there is a mechanism in the local mapping
to cull redundant keyframes, we will try to insert keyframes as
fast as possible, because that makes the tracking more robust to
challenging camera movements, typically rotations. To insert a

new keyframe, all the following conditions must be met.
1) More than 20 frames must have passed from the last global
relocalization.
2) Local mapping is idle, or more than 20 frames have passed
from last keyframe insertion.
3) Current frame tracks at least 50 points.
4) Current frame tracks less than 90% points than Kref .
Instead of using a distance criterion to other keyframes as
PTAM, we impose a minimum visual change (condition 4).
Condition 1 ensures a good relocalization and condition 3 a
good tracking. If a keyframe is inserted when the local mapping
is busy (second part of condition 2), a signal is sent to stop local
BA so that it can process as soon as possible the new keyframe.
VI. LOCAL MAPPING
In this section, we describe the steps performed by the local
mapping with every new keyframe Ki .
A. Keyframe Insertion
First, we update the covisibility graph, adding a new node
for Ki and updating the edges resulting from the shared map
points with other keyframes. We then update the spanning tree
linking Ki with the keyframe with most points in common. We
then compute the bags of words representation of the keyframe,
which will help in the data association for triangulating new
points.
B. Recent Map Points Culling
Map points, in order to be retained in the map, must pass a
restrictive test during the first three keyframes after creation,
which ensures that they are trackable and not wrongly triangulated, i.e., due to spurious data association. A point must fulfill
these two conditions.
1) The tracking must find the point in more than the 25% of
the frames in which it is predicted to be visible.
2) If more than one keyframe have passed from map point
creation, it must be observed from at least three keyframes.
Once a map point have passed this test, it can only be removed
if at any time it is observed from less than three keyframes.
This can happen when keyframes are culled and when local
BA discards outlier observations. This policy makes our map
contain very few outliers.
C. New Map Point Creation
New map points are created by triangulating ORB from connected keyframes Kc in the covisibility graph. For each unmatched ORB in Ki , we search a match with other unmatched
point in other keyframe. This matching is done as explained
in Section III-E and discard those matches that do not fulfill
the epipolar constraint. ORB pairs are triangulated, and to accept the new points, positive depth in both cameras, parallax,
reprojection error, and scale consistency is checked. Initially,
a map point is observed from two keyframes, but it could be
8
matched in others; therefore, it is projected in the rest of connected keyframes, and correspondences are searched as detailed
in Section V-D.
D. Local Bundle Adjustment
The local BA optimizes the currently processed keyframe
Ki , all the keyframes connected to it in the covisibility graph
Kc , and all the map points seen by those keyframes. All other
keyframes that see those points but are not connected to the
currently processed keyframe are included in the optimization
but remain fixed. Observations that are marked as outliers are
discarded at the middle and at the end of the optimization. See
the Appendix for more details about this optimization.
E. Local Keyframe Culling
In order to maintain a compact reconstruction, the local
mapping tries to detect redundant keyframes and delete them.
This is beneficial as BA complexity grows with the number of
keyframes, but also because it enables lifelong operation in the
same environment as the number of keyframes will not grow
unbounded, unless the visual content in the scene changes. We
discard all the keyframes in Kc whose 90% of the map points
have been seen in at least other three keyframes in the same or
finer scale. The scale condition ensures that map points maintain
keyframes from which they are measured with most accuracy.
This policy was inspired by the one proposed in the work of Tan
et al. [24], where keyframes were discarded after a process of
change detection.
VII. LOOP CLOSING
The loop closing thread takes Ki , the last keyframe processed
by the local mapping, and tries to detect and close loops. The
steps are next described.
A. Loop Candidates Detection
First, we compute the similarity between the bag of words
vector of Ki and all its neighbors in the covisibility graph
(m in = 30) and retain the lowest score sm in . Then, we query
the recognition database and discard all those keyframes whose
score is lower than sm in . This is a similar operation to gain
robustness as the normalizing score in DBoW2, which is computed from the previous image, but here we use covisibility
information. In addition, all those keyframes directly connected
to Ki are discarded from the results. To accept a loop candidate, we must detect consecutively three loop candidates that
are consistent (keyframes connected in the covisibility graph).
There can be several loop candidates if there are several places
with similar appearance to Ki .
B. Compute the Similarity Transformation
In monocular SLAM, there are seven DoFs in which the
map can drift: three translations, three rotations, and a scale
factor [6]. Therefore, to close a loop, we need to compute a
similarity transformation from the current keyframe Ki to the
loop keyframe Kl that informs us about the error accumulated
in the loop. The computation of this similarity will serve also as

geometrical validation of the loop.
We first compute correspondences between ORB associated
with map points in the current keyframe and the loop candidate
keyframes, following the procedure explained in Section III-E.
At this point, we have 3-D-to-3-D correspondences for each
loop candidate. We alternatively perform RANSAC iterations
with each candidate, trying to find a similarity transformation
using the method of Horn [42]. If we find a similarity Sil with
enough inliers, we optimize it (see the Appendix) and perform
a guided search of more correspondences. We optimize it again,
and if Sil is supported by enough inliers, the loop with Kl is
accepted.
C. Loop Fusion
The first step in the loop correction is to fuse duplicated map
points and insert new edges in the covisibility graph that will
attach the loop closure. First, the current keyframe pose Tiw is
corrected with the similarity transformation Sil , and this correction is propagated to all the neighbors of Ki , concatenating
transformations, so that both sides of the loop get aligned. All
map points seen by the loop keyframe and its neighbors are
projected into Ki , and its neighbors and matches are searched
in a narrow area around the projection, as done in Section V-D.
All those map points matched and those that were inliers in the
computation of Sil are fused. All keyframes involved in the fusion will update their edges in the covisibility graph effectively
creating edges that attach the loop closure.
D. Essential Graph Optimization
To effectively close the loop, we perform a pose graph optimization over the Essential Graph, described in Section III-D,
that distributes the loop closing error along the graph. The optimization is performed over similarity transformations to correct
the scale drift [6]. The error terms and cost function are detailed
in the Appendix. After the optimization, each map point is transformed according to the correction of one of the keyframes that
observes it.
VIII. EXPERIMENTS
We have performed an extensive experimental validation of
our system in the large robot sequence of NewCollege [39],
evaluating the general performance of the system, in 16 handheld indoor sequences of the TUM RGB-D benchmark [38],
evaluating the localization accuracy, relocalization, and lifelong
capabilities, and in 10 car outdoor sequences from the KITTI
dataset [40], evaluating real-time large scale operation, localization accuracy, and efficiency of the pose graph optimization.
Our system runs in real time and processes the images exactly at the frame rate they were acquired. We have carried
out all experiments with an Intel Core i7-4700MQ (four cores
@ 2.40 GHz) and 8 GB RAM. ORB-SLAM has three main
threads, that run in parallel with other tasks from ROS and the
operating system, which introduces some randomness in the results. For this reason, in some experiments, we report the median
from several runs.
Fig. 4. Example of loop detected in the NewCollege sequence. We draw the

inlier correspondences supporting the similarity transformation found.
Fig. 6. ORB-SLAM reconstruction of the full sequence of NewCollege. The

bigger loop on the right is traversed in opposite directions and not visual loop
closures were found; therefore, they do not perfectly align.
TABLE I
TRACKING AND MAPPING TIMES IN NEWCOLLEGE
Thread
Operation
TRACKING
ORB extraction
Initial Pose Est.
Track Local Map
Total
KeyFrame Insertion
Map Point Culling
Map Point Creation
Local BA
KeyFrame Culling
Total
LOCAL MAPPING
Fig. 5. Map before and after a loop closure in the NewCollege sequence. The
loop closure match is drawn in blue, the trajectory in green, and the local map
for the tracking at that moment in red. The local map is extended along both
sides of the loop after it is closed.
A. System Performance in the NewCollege Dataset

The NewCollege dataset [39] contains a 2.2-km sequence
from a robot traversing a campus and adjacent parks. The sequence is recorded by a stereo camera at 20 frames/s and a
resolution 512 382. It contains several loops and fast rotations that makes the sequence quite challenging for monocular
vision. To the best of our knowledge, there is no other monocular system in the literature able to process this whole sequence.
For example Strasdat et al. [7], despite being able to close loops
and work in large-scale environments, only showed monocular
results for a small part of this sequence.
As an example of our loop closing procedure, we show in
Fig. 4 the detection of a loop with the inliers that support the
similarity transformation. Fig. 5 shows the reconstruction before and after the loop closure. In red, the local map is shown,
which after the loop closure extends along both sides of the loop
Median (ms)
Mean (ms)
Std (ms)
11.10
3.38
14.84
30.57
10.29
0.10
66.79
296.08
8.07
383.59
11.42
3.45
16.01
31.60
11.88
3.18
72.96
360.41
15.79
464.27
1.61
0.99
9.98
10.39
5.03
6.70
31.48
171.11
18.98
217.89
closure. The whole map after processing the full sequence at its
real frame rate is shown in Fig. 6. The big loop on the right
does not perfectly align because it was traversed in opposite
directions, and the place recognizer was not able to find loop
closures.
We have extracted statistics of the times spent by each thread
in this experiment. Table I shows the results for the tracking
and the local mapping. Tracking works at frame rates around
2530 Hz, being the most demanding task to track the local
map. If needed, this time could be reduced limiting the number
of keyframes that are included in the local map. In the local
mapping thread, the most demanding task is local BA. The local
BA time varies if the robot is exploring or in a well-mapped area,
because during exploration, BA is interrupted if tracking inserts
a new keyframe, as explained in Section V-E. In case of not
needing new keyframes, local BA performs a generous number
of prefixed iterations.
Table II shows the results for each of the six loop closures
found. It can be seen how the loop detection increases sublinearly with the number of keyframes. This is due to the efficient
querying of the database that only compares the subset of images with words in common, which demonstrates the potential
of bag of words for place recognition. Our Essential Graph includes edges around five times the number of keyframes, which
is a quite sparse graph.
10
TABLE II
LOOP CLOSING TIMES IN NEWCOLLEGE
Loop Detection (ms)
Loop
1
2
3
4
5
6
Loop Correction (s)
KeyFrames
Essential Graph Edges
Candidates Detection
Similarity Transformation
Fusion
Essential Graph Optimization
Total (s)
287
1082
1279
2648
3150
4496
1347
5950
7128
12547
16033
21797
4.71
4.14
9.82
12.37
14.71
13.52
20.77
17.98
31.29
30.36
41.28
48.68
0.20
0.39
0.95
0.97
1.73
0.97
0.26
1.06
1.26
2.30
2.80
3.62
0.51
1.52
2.27
3.33
4.60
4.69
B. Localization Accuracy in the TUM RGB-D Benchmark

The TUM RGB-D benchmark [38] is an excellent dataset
to evaluate the accuracy of camera localization as it provides
several sequences with accurate ground truth obtained with an
external motion capture system. We have discarded all those sequences that we consider that are not suitable for pure monocular
SLAM systems, as they contain strong rotations, no texture, or
no motion.
For comparison, we have also executed the novel, direct,
semidense LSD-SLAM [10] and PTAM [4] in the benchmark.
We compare also with the trajectories generated by RGBDSLAM [43], which are provided for some of the sequences
in the benchmark website. In order to compare ORB-SLAM,
LSD-SLAM, and PTAM with the ground truth, we align the
keyframe trajectories using a similarity transformation, as scale
is unknown, and measure the absolute trajectory error [38]. In
the case of RGBD-SLAM, we align the trajectories with a rigid
body transformation, but also a similarity to check if the scale
was well recovered. LSD-SLAM initializes from random depth
values and takes time to converge; therefore, we have discarded
the first ten keyframes when comparing with the ground truth.
For PTAM, we manually selected two frames from which we
get a good initialization. Table III shows the median results over
five executions in each of the 16 sequences selected.
It can be seen that ORB-SLAM is able to process all the sequences, except for fr3_nostructure_texture_far (fr3_nstr_tex_
far). This is a planar scene that because the camera trajectory with respect to the plane has two possible interpretations,
i.e., the twofold ambiguity described in [27]. Our initialization
method detects the ambiguity and for safety refuses to initialize.
PTAM initializes selecting sometimes the true solution and others the corrupted one, in which case the error is unacceptable.
We have not noticed two different reconstructions from LSDSLAM, but the error in this sequence is very high. In the rest of
the sequences, PTAM and LSD-SLAM exhibit less robustness
than our method, loosing track in eight and three sequences,
respectively.
In terms of accuracy, ORB-SLAM and PTAM are similar in
open trajectories, while ORB-SLAM achieves higher accuracy
when detecting large loops as in the sequence fr3_nostructure_
texture_near_withloop (fr3_nstr_tex_near). The most surprising
results is that both PTAM and ORB-SLAM are clearly more accurate than LSD-SLAM and RGBD-SLAM. One of the possible
causes can be that they reduce the map optimization to a posegraph optimization where sensor measurements are discarded,
TABLE III
KEYFRAME LOCALIZATION ERROR COMPARISON IN THE
TUM RGB-D BENCHMARK [38]
Absolute KeyFrame Trajectory RMSE (cm)
ORB-SLAM
PTAM
fr1_xyz
0.90
1.15
fr2_xyz
0.30
0.20
fr1_floor
2.99
X
fr1_desk
1.69
X
fr2_360 _kidnap
3.81
2.63
fr2_desk
0.88
X
fr3_long _office
3.45
X
fr3_nstr_tex_far
ambiguity detected 4.92 / 34.74
fr3_nstr_ tex_near
1.39
2.74
fr3_str_tex_far
0.77
0.93
fr3_str_ tex_near
1.58
1.04
fr2_desk_person
0.63
X
fr3_sit_xyz
0.79
0.83
fr3_sit_halfsph
1.34
X
fr3_walk_xyz
1.24
X
fr3_walk_halfsph
1.74
X
LSD-SLAM RGBD- SLAM

9.00
2.15
38.07
10.65
X
4.57
38.53
18.31
7.54
7.95
X
31.73
7.73
5.87
12.44
X
1.34 (1.34)
2.61 (1.42)
3.51 (3.51)
2.58 (2.52)
393.3 (100.5)
9.50 (3.94)
6.97 (2.00)
Results for ORB-SLAM, PTAM, and LSD-SLAM are the median over five executions
in each sequence. The trajectories have been aligned with 7 DoFs with the ground truth.
Trajectories for RGBD-SLAM are taken from the benchmark website, only available
for fr1 and fr2 sequences, and have been aligned with 6 DoFs and 7 DoFs (results
between brackets). X means that the tracking is lost at some point and a significant
portion of the sequence is not processed by the system.
while we perform BA and jointly optimize cameras and map

over sensor measurements, which is the gold standard algorithm
to solve structure from motion [2]. We further discuss this result
in Section IX-B. Another interesting result is that LSD-SLAM
seems to be less robust to dynamic objects than our system as
seen in fr2_desk_with_person and fr3_walking_xyz.
We have noticed that RGBD-SLAM has a bias in the scale in
fr2 sequences, as aligning the trajectories with 7 DoFs significantly reduces the error. Finally, it should be noted that Engel
et al. [10] reported that PTAM has less accuracy than LSDSLAM in fr2_xyz with an RMSE of 24.28 cm. However, the
paper does not give enough details on how those results were
obtained, and we have been unable to reproduce them.
C. Relocalization in the TUM RGB-D Benchmark
We perform two relocalization experiments in the TUM RGBD benchmark. In the first experiment, we build a map with the
first 30 s of the sequence fr2_xyz and perform global relocalization with every successive frame and evaluate the accuracy
Fig. 7. Relocalization experiment in fr2_xyz. Map is initially created during

the first 30 s of the sequence (KFs). The goal is to relocalize subsequent frames.
Successful relocalizations (R) of our system and PTAM are shown. The ground
truth (GT) is only shown for the frames to relocalize.
11
Fig. 8. Example of challenging relocalizations (severe scale change, dynamic

objects) that our system successfully found in the relocalization experiments.
TABLE IV
RESULTS FOR THE RELOCALIZATION EXPERIMENTS
Initial Map
System
KFs
RMSE (cm)
Relocalization
Recall (%)
RMSE (cm)
Max. Error (cm)
fr2_xyz. 2769 frames to relocalize

PTAM
ORB-SLAM
37
24
0.19
0.19
34.9
78.4
0.26
0.38
1.52
1.67
fr3_walking_xyz. 859 frames to relocalize

PTAM
ORB-SLAM
34
31
0.83
0.82
0.0
77.9
1.32
4.95
of the recovered poses. We perform the same experiment with

PTAM for comparison. Fig. 7 shows the keyframes used to create the initial map, the poses of the relocalized frames, and the
ground truth for those frames. It can be seen that PTAM is only
able to relocalize frames, which are near to the keyframes due
to the little invariance of its relocalization method. Table IV
shows the recall and the error with respect to the ground truth.
ORB-SLAM accurately relocalizes more than the double of
frames than PTAM. In the second experiment, we create an initial map with sequence fr3_sitting_xyz and try to relocalize all
frames from fr3_walking_xyz. This is a challenging experiment
as there are big occlusions due to people moving in the scene.
Here, PTAM finds no relocalizations, while our system relocalizes 78% of the frames, as can be seen in Table IV. Fig. 8 shows
some examples of challenging relocalizations performed by our
system in these experiments.
D. Lifelong Experiment in the TUM RGB-D Benchmark
Previous relocalization experiments have shown that our system is able to localize in a map from very different viewpoints
and robustly under moderate dynamic changes. This property in
Fig. 9. Lifelong experiment in a static environment where the camera is

always looking at the same place from different viewpoints. PTAM is always
inserting keyframes, while ORB-SLAM is able to prune redundant keyframes
and maintains a bounded-size map.
conjunction with our keyframe culling procedure allows us to

operate lifelong in the same environment under different viewpoints and some dynamic changes.
In the case of a completely static scenario, our system is
able to maintain the number of keyframes bounded even if the
camera is looking at the scene from different viewpoints. We
demonstrate it in a custom sequence where the camera is looking
at the same desk during 93 s but performing a trajectory so that
the viewpoint is always changing. We compare the evolution of
the number of keyframes in our map and those generated by
PTAM in Fig. 9. It can be seen how PTAM is always inserting
keyframes, while our mechanism to prune redundant keyframes
makes its number to saturate.
While the lifelong operation in a static scenario should be a
requirement of any SLAM system, more interesting is the case
where dynamic changes occur. We analyze the behavior of our
system in such a scenario by running consecutively the dynamic
sequences from fr3: sitting_xyz, sitting_halfsphere, sitting_rpy,
walking_xyz, walking_halfspehere, and walking_rpy. All the
sequences focus the camera on the same desk but perform
different trajectories, while people are moving and change
some objects like chairs. Fig. 10(a) shows the evolution of the
12
total number of keyframes in the map, and Fig. 10(b) shows for
each keyframe its frame of creation and destruction, showing
how long the keyframes have survived in the map. It can be seen
that during the first two sequences, the map size grows as all the
views of the scene are being seen for the first time. In Fig. 10(b),
we can see that several keyframes created during these two
first sequences are maintained in the map during the whole
experiment. During the sequences sitting_rpy and walking_xyz,
the map does not grow, because the map created so far explains
well the scene. In contrast, during the last two sequences, more
keyframes are inserted showing that there are some novelties
in the scene that were not yet represented, due probably to
dynamic changes. Finally, Fig. 10(c) shows a histogram of the
keyframes according to the time they have survived with respect
to the remaining time of the sequence from its moment of
creation. It can be seen that most of the keyframes are destroyed
by the culling procedure soon after creation, and only a small
subset survive until the end of the experiment. On one hand,
this shows that our system has a generous keyframe spawning
policy, which is very useful when performing abrupt motions
in exploration. On the other hand, the system is eventually able
to select a small representative subset of those keyframes.
In these lifelong experiments, we have shown that our map
grows with the content of the scene but not with the time and is
able to store the dynamic changes of the scene, which could be
useful to perform some scene understanding by accumulating
experience in an environment.
E. Large-Scale and Large Loop Closing in the KITTI Dataset
The odometry benchmark from the KITTI dataset [40] contains 11 sequences from a car driven around a residential area
with accurate ground truth from GPS and a Velodyne laser scanner. This is a very challenging dataset for monocular vision due
to fast rotations, areas with lot of foliage, which make more difficult data association, and relatively high car speed, being the
sequences recorded at 10 frames/s. We play the sequences at the
real frame rate they were recorded, and ORB-SLAM is able to
process all the sequences by the exception of sequence 01, which
is a highway with few trackable close objects. Sequences 00,
02, 05, 06, 07, and 09 contain loops that were correctly detected
and closed by our system. Sequence 09 contains a loop that can
be detected only in a few frames at the end of the sequence, and
our system not always detects it (the results provided are for the
executions in which it was detected).
Qualitative comparisons of our trajectories and the ground
truth are shown in Figs. 11 and 12. As in the TUM RGB-D
benchmark, we have aligned the keyframe trajectories of our
system and the ground truth with a similarity transformation.
We can compare qualitatively our results from Figs. 11 and 12
with the results provided for sequences 00, 05, 06, 07, and 08 by
the recent monocular SLAM approach of Lim et al. [25, Fig. 10].
ORB-SLAM produces clearly more accurate trajectories for all
those sequences by the exception of sequence 08 in which they
seem to suffer less drift.
Table V shows the median RMSE error of the keyframe trajectory over five executions in each sequence. We also provide the
Fig. 10. Lifelong experiment in a dynamic environment from the TUM

RGB-D Benchmark. (a) Evolution of the number of keyframes in the map.
(b) Keyframe creation and destruction. Each horizontal line corresponds to a
keyframe, from its creation frame until its destruction. (c) Histogram of the
survival time of all spawned keyframes with respect to the remaining time of
the experiment.
dimensions of the maps to put in context the errors. The results

demonstrate that our system is very accurate, being the trajectory error typically around the 1% of its dimensions, sometimes
less as in sequence 03 with an error of the 0.3% or higher as in
sequence 08 with the 5%. In sequence 08, there are no loops,
and drift cannot be corrected, which makes clear the need of
loop closures to achieve accurate reconstructions.
In this experiment, we have also checked how much the reconstruction can be improved by performing 20 iterations of full
BA (see the Appendix for details) at the end of each sequence.
We have noticed that some iterations of full BA slightly improves
the accuracy in the trajectories with loops, but it has negligible
effect in open trajectories, which means that the output of our
system is already very accurate. In any case, if the most accurate results are needed, our algorithm provides a set of matches,
which define a strong camera network, and an initial guess, so
that full BA converge in few iterations.
13
Fig. 11. Sequences 00, 05, and 07 from the odometry benchmark of the KITTI dataset. (Left) Points and keyframe trajectory. (Center) trajectory and ground
truth. (Right) Trajectory after 20 iterations of full BA. The output of our system is quite accurate, while it can be slightly improved with some iterations of BA.
Finally, we wanted to show the efficacy of our loop closing

approach and the influence of the m in used to include edges
in the essential graph. We have selected the sequence 09 (a
very long sequence with a loop closure at the end), and in the
same execution, we have evaluated different loop closing strategies. In Table VI, we show the keyframe trajectory RMSE and
the time spent in the optimization in different cases: without
loop closing, if we directly apply a full BA (20 or 100 iterations), if we apply only pose graph optimization (ten iterations
with different number of edges), and if we apply pose graph
optimization and full BA afterwards. Fig. 13 shows the output
trajectory of the different methods. The results clearly show that
before loop closure, the solution is so far from the optimal that
BA has convergence problems. Even after 100 iterations, still
the error is very high. On the other hand, essential graph optimization shows fast convergence and more accurate results. It
can be seen that the choice of m in has not significant effect
in accuracy, but decreasing the number of edges, the time can
be significantly reduced. Performing an additional BA after the
pose graph optimization slightly improves the accuracy while
increasing substantially the time.
IX. CONCLUSION AND DISCUSSION

A. Conclusions
In this study, we have presented a new monocular SLAM
system with a detailed description of its building blocks and
an exhaustive evaluation in public datasets. Our system has
demonstrated that it can process sequences from indoor and
outdoor scenes and from car, robot, and hand-held motions. The
accuracy of the system is typically below 1 cm in small indoor
scenarios and of a few meters in large outdoor scenarios (once
we have aligned the scale with the ground truth).
Currently, PTAM by Klein and Murray [4] is considered the
most accurate SLAM method from monocular video in real time.
It is not coincidence that the back-end of PTAM is BA, which
is well known to be the gold standard method for the offline
Structure From Motion problem [2]. One of the main successes
of PTAM, and the earlier work of Mouragnon [3], was to bring
that knowledge into the robotics SLAM community and demonstrate its real-time performance. The main contribution of our
work is to expand the versatility of PTAM to environments that
are intractable for that system. To achieve this, we have designed
14
Fig. 12. ORB-SLAM keyframe trajectories in sequences 02, 03, 04 ,06, 08, 09, and 10 from the odometry benchmark of the KITTI dataset. Sequence 08 does
not contains loops and drift (especially scale) is not corrected. (a) Sequence 02. (b) Sequence 03. (c) Sequence 04. (d) Sequence 06. (e) Sequence 08 (f) Sequence
09. (g) Sequence 10.
from scratch a new monocular SLAM system with some new

ideas and algorithms, but also incorporating excellent works
developed in the past few years, such as the loop detection of
Galvez-Lopez and Tardos [5], the loop closing procedure and
covisibility graph of Strasdat et al. [6], [7], the optimization
framework g2o by Kuemmerle et al. [37], and ORB features by
Rubble et al. [9]. To the best of our knowledge, no other system
has demonstrated to work in as many different scenarios and
with such accuracy. Therefore, our system is currently the most
reliable and complete solution for monocular SLAM. Our novel
policy to spawn and cull keyframes permits to create keyframes
every few frames, which are eventually removed when considered redundant. This flexible map expansion is really useful in
poorly conditioned exploration trajectories, i.e., close to pure
rotations or fast movements. When operating repeatedly in the
same environment, the map only grows if the visual content of
the scene changes, storing a history of its different visual appearances. Interesting results for long-term mapping could be
extracted analyzing this history.
Finally, we have also demonstrated that ORB features have
enough recognition power to enable place recognition from severe viewpoint change. Moreover, they are so fast to extract and
match (without the need of multithreading or GPU acceleration)
that enable real-time accurate tracking and mapping.
B. Sparse/Feature-Based Versus Dense/Direct Methods
Recent real-time monocular SLAM algorithms such as
DTAM [44] and LSD-SLAM [10] are able to perform dense
or semidense reconstructions of the environment, while the
camera is localized by optimizing directly over image pixel
intensities. These direct approaches do not need feature
extraction and thus avoid the corresponding artifacts. They
are also more robust to blur, low-texture environments and
TABLE V
RESULTS OF OUR SYSTEM IN THE KITTI DATASET
ORB-SLAM
+ Global BA (20 its.)
Sequence
Dimension (mm)
KFs
RMSE (m)
RMSE (m)
Time BA (s)
KITTI 00
KITTI 01
KITTI 02
KITTI 03
KITTI 04
KITTI 05
KITTI 06
KITTI 07
KITTI 08
KITTI 09
KITTI 10
564 496
1157 1827
599 946
471 199
0.5 394
479 426
23 457
191 209
808 391
465 568
671 177
1391
X
1801
250
108
820
373
351
1473
653
411
6.68
X
21.75
1.59
1.79
8.23
14.68
3.36
46.58
7.62
8.68
5.33
X
21.28
1.51
1.62
4.85
12.34
2.26
46.68
6.62
8.80
24.83
X
30.07
4.88
1.58
15.20
7.78
6.28
25.60
11.33
7.64
high-frequency texture like asphalt [45]. Their denser reconstructions, as compared with the sparse point map of our system
or PTAM, could be more useful for other tasks than just camera
localization.
However, direct methods have their own limitations. First,
these methods assume a surface reflectance model that in real
scenes produces its own artifacts. The photometric consistency
limits the baseline of the matches, typically narrower than those
that features allow. This has a great impact in reconstruction
accuracy, which requires wide baseline observations to reduce
depth uncertainty. Direct methods, if not correctly modeled,
are quite affected by rolling-shutter, autogain, and autoexposure
artifacts (as in the TUM RGB-D Benchmark). Finally, because
direct methods are, in general, very computationally demanding,
the map is just incrementally expanded as in DTAM, or map
optimization is reduced to a pose graph, discarding all sensor
measurements as in LSD-SLAM.
TABLE VI
COMPARISON OF LOOP CLOSING STRATEGIES IN KITTI 09
Method
BA (20)
BA (100)
EG (200)
EG (100)
EG (50)
EG (15)
EG (100) + BA (20)
Time (s)
Pose Graph Edges
RMSE (m)
14.64
72.16
0.38
0.48
0.59
0.94
13.40
890
1979
3583
6663
1979
48.77
49.90
18.82
8.84
8.36
8.95
8.88
7.22
The first row shows results without loop closing. The number between
brackets for BA means number of LevenbergMarquardt (LM) iterations
while, for EG (essential graph), it is m i n to build the essential graph.
All EG optimizations perform ten LM iterations.
15
which a dense and accurate map of the scene can be built. A

first effort in this line is presented in [47].
APPENDIX
NONLINEAR OPTIMIZATIONS
1) Bundle Adjustment [1]: Map point 3-D locations Xw ,j
R3 and keyframe poses Tiw SE(3), where w stands
for the world reference, are optimized minimizing the
reprojection error with respect to the matched keypoints
xi,j R2 . The error term for the observation of a map
point j in a keyframe i is
ei,j = xi,j i (Tiw , Xw ,j )
where i is the projection function
xi,j
+ ci,u
fi,u
zi,j
i (Tiw , Xw ,j ) =
yi,j
fi,v
+ ci,v
zi,j
T

xi,j yi,j zi,j
= Riw Xw ,j + tiw
(5)
(6)
where Riw SO(3) and tiw R3 are, respectively, the

rotation and translation parts of Tiw , and (fi,u , fi,v ) and
(ci,u , ci,v ) are the focal length and principle point associated with camera i. The cost function to be minimized
is

C=
h (eTi,j 1
(7)
i,j ei,j )
i,j
Fig. 13. Comparison of different loop closing strategies in KITTI 09. (a)
Without Loop Closing. (b) BA (20). (c) EG (100). (d) EG (100) + BA (20).
In contrast, feature-based methods are able to match features

with a wide baseline, thanks to their good invariance to viewpoint and illumination changes. BA jointly optimizes camera
poses and points over sensor measurements. In the context of
structure and motion estimation, Torr and Zisserman [46] already pointed the benefits of feature-based against direct methods. In this study, we provide experimental evidence (see Section VIII-B) of the superior accuracy of feature-based methods
in real-time SLAM. We consider that the future of monocular
SLAM should incorporate the best of both approaches.
C. Future Work
The accuracy of our system can still be improved incorporating points at infinity in the tracking. These points, which are not
seen with sufficient parallax and our system does not include in
the map, are very informative of the rotation of the camera [21].
Another open way is to upgrade the sparse map of our system to a denser and more useful reconstruction. Thanks to our
keyframe selection, keyframes comprise a compact summary
of the environment with a very high pose accuracy and rich
information of covisibility. Therefore, the ORB-SLAM sparse
map can be an excellent initial guess and skeleton, on top of
where h is the Huber robust cost function, and i,j =

2
I22 is the covariance matrix associated with the scale
i,j
at which the keypoint was detected. In case of full BA (used
in the map initialization explained in Section IV and in the
experiments in Section VIII-E), we optimize all points and
keyframes, by the exception of the first keyframe which
remain fixed as the origin. In local BA (see Section VI-D),
all points included in the local area are optimized, while
a subset of keyframes is fixed. In pose optimization, or
motion-only BA (see Section V), all points are fixed, and
only the camera pose is optimized.
2) Pose Graph Optimization over Sim(3) Constraints [6]:
Given a pose graph of binary edges (see Section VII-D),
we define the error in an edge as
ei,j = logSim (3) (Sij Sj w S1
iw )
(8)
where Sij is the relative Sim(3) transformation between

both keyframes computed from the SE(3) poses just before
the pose graph optimization and setting the scale factor
to 1. In the case of the loop closure edge, this relative
transformation is computed with the method of Horn [42].
The logSim3 [48] transforms to the tangent space so that
the error is a vector in R7 . The goal is to optimize the
Sim(3) keyframe poses minimizing the cost function as

C=
(eTi,j i,j ei,j )
(9)
i,j
16
where i,j is the information matrix of the edge, which,

as in [48], we set to the identity. We fix the loop closure
keyframe to fix the 7 degrees of gauge freedom. Although
this method is a rough approximation of a full BA, we
demonstrate experimentally in Section VIII-E that it has
significantly faster and better convergence than BA.
3) Relative Sim(3) Optimization: Given a set of n matches
i j (keypoints and their associated 3-D map points)
between keyframe 1 and keyframe 2, we want to optimize
the relative Sim(3) transformation S12 (see Section VIIB) that minimizes the reprojection error in both images
as
e1 = x1,i 1 (S12 , X2,j )
e2 = x2,j 2 (S1
12 , X1,i )
and the cost function to minimize is

T
1
C=
h (eT1 1
1,i e1 ) + h (e2 2,j e2 )
(10)
(11)
where 1,i and 2,i are the covariance matrices associated with the scale in which keypoints in images 1 and 2
were detected. In this optimization, the points are fixed.
REFERENCES
[1] B. Triggs, P. F. McLauchlan, R. I. Hartley, and A. W. Fitzgibbon, Bundle adjustment a modern synthesis, in Vision Algorithms: Theory and
Practice. New York, NY, USA: Springer, 2000, pp. 298372.
[2] R. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision,
2nd ed., Cambridge, U.K.: Cambridge Univ. Press, 2004.
[3] E. Mouragnon, M. Lhuillier, M. Dhome, F. Dekeyser, and P. Sayd, Real
time localization and 3D reconstruction, in Proc. IEEE Comput. Soc.
Conf. Comput. Vision Pattern Recog., 2006, vol. 1, pp. 363370.
[4] G. Klein and D. Murray, Parallel tracking and mapping for small AR
workspaces, in Proc. IEEE ACM Int. Symp. Mixed Augmented Reality,
Nara, Japan, Nov. 2007, pp. 225234.
[5] D. Galvez-Lopez and J. D. Tardos, Bags of binary words for fast place
recognition in image sequences, IEEE Trans. Robot., vol. 28, no. 5,
pp. 11881197, Oct. 2012.
[6] H. Strasdat, J. M. M. Montiel, and A. J. Davison, Scale drift-aware
large scale monocular SLAM, presented at the Proc. Robot.: Sci. Syst.,
Zaragoza, Spain, Jun. 2010.
[7] H. Strasdat, A. J. Davison, J. M. M. Montiel, and K. Konolige, Double
window optimisation for constant time visual SLAM, in Proc. IEEE Int.
Conf. Comput. Vision, Barcelona, Spain, Nov. 2011, pp. 23522359.
[8] C. Mei, G. Sibley, and P. Newman, Closing loops without places, in
Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Taipei, Taiwan, Oct. 2010,
pp. 37383744.
[9] E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, ORB: An efficient
alternative to SIFT or SURF, in Proc. IEEE Int. Conf. Comput. Vision,
Barcelona, Spain, Nov. 2011, pp. 25642571.
[10] J. Engel, T. Schops, and D. Cremers, LSD-SLAM: Large-scale direct
monocular SLAM, in Proc. Eur. Conf. Comput. Vision, Zurich, Switzerland, Sep. 2014, pp. 834849.
[11] R. Mur-Artal and J. D. Tardos, Fast relocalisation and loop closing in
keyframe-based SLAM, in Proc. IEEE Int. Conf. Robot. Autom., Hong
Kong, Jun. 2014, pp. 846853.
[12] R. Mur-Artal and J. D. Tardos, ORB-SLAM: Tracking and mapping
recognizable features, presented at the MVIGRO Workshop Robot. Sci.
Syst., Berkeley, CA, USA, Jul. 2014.
[13] B. Williams, M. Cummins, J. Neira, P. Newman, I. Reid, and J. D. Tardos,
A comparison of loop closing techniques in monocular SLAM, Robot.
Auton. Syst., vol. 57, no. 12, pp. 11881197, 2009.
[14] D. Nister and H. Stewenius, Scalable recognition with a vocabulary tree,
in Proc. IEEE Comput. Soc. Conf. Comput. Vision Pattern Recog., New
York, NY, USA, Jun. 2006, vol. 2, pp. 21612168.
[15] M. Cummins and P. Newman, Appearance-only SLAM at large scale

with FAB-MAP 2.0, Int. J. Robot. Res., vol. 30, no. 9, pp. 11001123,
2011.
[16] M. Calonder, V. Lepetit, C. Strecha, and P. Fua, BRIEF: Binary robust
independent elementary features, in Proc. Eur. Conf. Comput. Vision,
Hersonissos, Greece, Sep. 2010, pp. 778792.
[17] E. Rosten and T. Drummond, Machine learning for high-speed corner
detection, in Proc. Eur. Conf. Comput. Vision, Graz, Austria, May 2006,
pp. 430443.
[18] H. Bay, T. Tuytelaars, and L. Van Gool, SURF: Speeded up robust features, in Proc. Eur. Conf. Comput. Vision, Graz, Austria, May 2006,
pp. 404417.
[19] D. G. Lowe, Distinctive image features from scale-invariant keypoints,
Int. J. Comput. Vision, vol. 60, no. 2, pp. 91110, 2004.
[20] A. J. Davison, I. D. Reid, N. D. Molton, and O. Stasse, MonoSLAM:
Real-time single camera SLAM, IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 29, no. 6, pp. 10521067, Jun. 2007.
[21] J. Civera, A. J. Davison, and J. M. M. Montiel, Inverse depth parametrization for monocular SLAM, IEEE Trans. Robot., vol. 24, no. 5,
pp. 932945, Oct. 2008.
[22] C. Forster, M. Pizzoli, and D. Scaramuzza, SVO: Fast semi-direct monocular visual odometry, in Proc. IEEE Int. Conf. Robot. Autom., Hong Kong,
Jun. 2014, pp. 1522.
[23] O. D. Faugeras and F. Lustman, Motion and structure from motion in a
piecewise planar environment, Int. J. Pattern Recog. Artif. Intell., vol. 2,
no. 03, pp. 485508, 1988.
[24] W. Tan, H. Liu, Z. Dong, G. Zhang, and H. Bao, Robust monocular SLAM
in dynamic environments, in Proc. IEEE Int. Symp. Mixed Augmented
Reality, Adelaide, Australia, Oct. 2013, pp. 209218.
[25] H. Lim, J. Lim, and H. J. Kim, Real-time 6-DOF monocular visual
SLAM in a large-scale environment, in Proc. IEEE Int. Conf. Robot.
Autom., Hong Kong, Jun. 2014, pp. 15321539.
[26] D. Nister, An efficient solution to the five-point relative pose problem,
IEEE Trans. Pattern Anal. Mach. Intell., vol. 26, no. 6, pp. 756770, Jun.
2004.
[27] H. Longuet-Higgins, The reconstruction of a plane surface from two
perspective projections, Proc. Royal Soc. London Ser. B, Biol, Sci.,
vol. 227, no. 1249, pp. 399410, 1986.
[28] P. H. Torr, A. W. Fitzgibbon, and A. Zisserman, The problem of degeneracy in structure and motion recovery from uncalibrated image sequences,
Int. J. Comput. Vision, vol. 32, no. 1, pp. 2744, 1999.
[29] A. Chiuso, P. Favaro, H. Jin, and S. Soatto, Structure from motion causally
integrated over time, IEEE Trans. Pattern Anal. Mach. Intell., vol. 24,
no. 4, pp. 523535, Apr. 2002.
[30] E. Eade and T. Drummond, Scalable monocular SLAM, in Proc. IEEE
Comput. Soc. Conf. Comput. Vision Pattern Recog., New York, NY, USA,
Jun. 2006, vol. 1, pp. 469476.
[31] H. Strasdat, J. M. M. Montiel, and A. J. Davison, Visual SLAM: Why
filter? Image Vision Comput., vol. 30, no. 2, pp. 6577, 2012.
[32] G. Klein and D. Murray, Improving the agility of keyframe-based
SLAM, in Proc. Eur. Conf. Comput. Vision, Marseille, France, Oct. 2008,
pp. 802815.
[33] K. Pirker, M. Ruther, and H. Bischof, CD SLAM-continuous localization
and mapping in a dynamic world, in Proc. IEEE/RSJ Int. Conf. Intell.
Robots Syst., San Francisco, CA, USA, Sep. 2011, pp. 39903997.
[34] S. Song, M. Chandraker, and C. C. Guest, Parallel, real-time monocular visual odometry, in Proc. IEEE Int. Conf. Robot. Autom., 2013,
pp. 46984705.
[35] P. F. Alcantarilla, J. Nuevo, and A. Bartoli, Fast explicit diffusion for
accelerated features in nonlinear scale spaces, presented at the Brit. Mach.
Vision Conf., Bristol, U.K., 2013.
[36] X. Yang and K.-T. Cheng, LDB: An ultra-fast feature for scalable augmented reality on mobile devices, in Proc. IEEE Int. Symp. Mixed Augmented Reality, 2012, pp. 4957.
[37] R. Kuemmerle, G. Grisetti, H. Strasdat, K. Konolige, and W. Burgard,
g2o: A general framework for graph optimization, in Proc. IEEE Int.
Conf. Robot. Autom., Shanghai, China, May 2011, pp. 36073613.
[38] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, A benchmark for the evaluation of RGB-D SLAM systems, in Proc. IEEE/RSJ Int.
Conf. Intell. Robots Syst., Vilamoura, Portugal, Oct. 2012, pp. 573580.
[39] M. Smith, I. Baldwin, W. Churchill, R. Paul, and P. Newman, The new
college vision and laser data set, Int. J. Robot. Res., vol. 28, no. 5,
pp. 595599, 2009.
[40] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, Vision meets robotics:
The KITTI dataset, Int. J. Robot. Res., vol. 32, no. 11, pp. 12311237,
2013.
[41] V. Lepetit, F. Moreno-Noguer, and P. Fua, EPnP: An accurate O(n)

solution to the PnP problem, Int. J. Comput. Vision, vol. 81, no. 2,
pp. 155166, 2009.
[42] B. K. P. Horn, Closed-form solution of absolute orientation using unit
quaternions, J. Opt. Soc. Amer. A, vol. 4, no. 4, pp. 629642, 1987.
[43] F. Endres, J. Hess, J. Sturm, D. Cremers, and W. Burgard, 3-D mapping
with an RGB-D camera, IEEE Trans. Robot., vol. 30, no. 1, pp. 177187,
Feb. 2014.
[44] R. A. Newcombe, S. J. Lovegrove, and A. J. Davison, DTAM: Dense
tracking and mapping in real-time, in Proc. IEEE Int. Conf. Comput.
Vision, Barcelona, Spain, Nov. 2011, pp. 23202327.
[45] S. Lovegrove, A. J. Davison, and J. Ibanez-Guzman, Accurate visual
odometry from a rear parking camera, in Proc. IEEE Intell. Vehicles
Symp., 2011, pp. 788793.
[46] P. H. Torr and A. Zisserman, Feature based methods for structure and
motion estimation, in Vision Algorithms: Theory and Practice. New York,
NY, USA: Springer, 2000, pp. 278294.
[47] R. Mur-Artal and J. D. Tardos, Probabilistic semi-dense mapping from
highly accurate feature-based monocular SLAM, presented at the Proc.
Robot.: Sci. Syst., Rome, Italy, Jul. 2015.
[48] H. Strasdat, Local accuracy and global consistency for efficient visual
SLAM, Ph.D. dissertation, Imperial College London, London, U.K., Oct.
2012.
Mur-Artal was born in Zaragoza, Spain, in

Raul
1989. He received the Industrial Engineering degree
(in industrial automation and robotics) in 2012 and
the M.S. degree in systems and computer engineering in 2013 from University of Zaragoza, Zaragoza,
where he is currently working toward the Ph.D. degree with the I3A Robotics, Perception and RealTime Group.
His research interests include visual localization
and long-term mapping.
17
J. M . M. Montiel (M15) was born in Arnedo, Spain,

in 1967. He received the M.S. and Ph.D. degrees in
electrical engineering from Universidad de Zaragoza,
Zaragoza, Spain, in 1992 and 1996, respectively.
He is currently a Full Professor with the Departamento de Informatica e Ingeniera de Sistemas,
Universidad de Zaragoza, where he is in charge of
perception and computer vision research grants and
courses. His interests include real-time vision localization and semantic mapping for rigid and nonrigid
environments, and the transference of this technology
to robotic and nonrobotic application domains.
Dr. Montiel is a Member of the I3A Robotics, Perception, and Real-Time
Group, Universidad de Zaragoza. He has been awarded several Spanish MEC
grants to fund research with the University of Oxford, U.K., and with Imperial
College London, U.K.
Juan D. Tardos (M05) was born in Huesca, Spain,

in 1961. He received the M.S. and Ph.D. degrees in
electrical engineering from University of Zaragoza,
Zaragoza, Spain, in 1985 and 1991, respectively.
He is a Full Professor with the Departamento de
Informatica e Ingeniera de Sistemas, University of
Zaragoza, where he is in charge of courses in robotics,
computer vision, and artificial intelligence. His research interests include simultaneous localization and
mapping, perception, and mobile robotics.
Dr. Tardos is a member of the I3A Robotics, Perception, and Real-Time Group, University of Zaragoza.

Orb Slam

Uploaded by

Copyright:

Available Formats

Orb Slam

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Orb Slam

Uploaded by

Copyright:

Available Formats

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

ORB-SLAM: A Versatile and Accurate Monocular

AbstractThis paper presents ORB-SLAM, a feature-based

4) An initial estimation of the keyframe poses and point

IEEE TRANSACTIONS ON ROBOTICS

We present an extensive evaluation in popular public datasets

be initialized with high uncertainty in depth using an inverse

Strasdat et al. [6] presented a large-scale monocular SLAM

out requiring to extract features in every frame, they are able

IEEE TRANSACTIONS ON ROBOTICS

searched by reprojection, and camera pose is optimized again

Fig. 2. Reconstruction and graphs in the sequence f r3 long

the point with the optical center of the keyframes that

D. Covisibility Graph and Essential Graph

An additional benefit of the bags of words representation for

IEEE TRANSACTIONS ON ROBOTICS

the measurement error). is defined equal to TH so that

and then retrieve four motion hypotheses with the singular

Fig. 3. Top: PTAM, middle: LSD-SLAM, bottom: ORB-SLAM, some time

in a plane, while our method has waited until there is enough

B. Initial Pose Estimation From Previous Frame

challenging camera movements, typically rotations. To insert a

IEEE TRANSACTIONS ON ROBOTICS

in the loop. The computation of this similarity will serve also as

Fig. 4. Example of loop detected in the NewCollege sequence. We draw the

Fig. 6. ORB-SLAM reconstruction of the full sequence of NewCollege. The

A. System Performance in the NewCollege Dataset

IEEE TRANSACTIONS ON ROBOTICS

Loop Correction (s)

Essential Graph Edges

Essential Graph Optimization

B. Localization Accuracy in the TUM RGB-D Benchmark

LSD-SLAM RGBD- SLAM

while we perform BA and jointly optimize cameras and map

Fig. 7. Relocalization experiment in fr2_xyz. Map is initially created during

Fig. 8. Example of challenging relocalizations (severe scale change, dynamic

Max. Error (cm)

fr2_xyz. 2769 frames to relocalize

fr3_walking_xyz. 859 frames to relocalize

of the recovered poses. We perform the same experiment with

Fig. 9. Lifelong experiment in a static environment where the camera is

conjunction with our keyframe culling procedure allows us to

IEEE TRANSACTIONS ON ROBOTICS

Fig. 10. Lifelong experiment in a dynamic environment from the TUM

dimensions of the maps to put in context the errors. The results

Finally, we wanted to show the efficacy of our loop closing

IX. CONCLUSION AND DISCUSSION

IEEE TRANSACTIONS ON ROBOTICS

from scratch a new monocular SLAM system with some new

+ Global BA (20 its.)

Pose Graph Edges

which a dense and accurate map of the scene can be built. A

where Riw SO(3) and tiw R3 are, respectively, the

In contrast, feature-based methods are able to match features

where h is the Huber robust cost function, and i,j =

where Sij is the relative Sim(3) transformation between

IEEE TRANSACTIONS ON ROBOTICS

where i,j is the information matrix of the edge, which,