Deep Learning-Powered Visual SLAM Aimed at Assisting Visually
Impaired Navigation
Marziyeh Bamdad
1,2 a
, Hans-Peter Hutter
1 b
and Alireza Darvishy
1 c
1
Institute of Computer Science, ZHAW School of Engineering, Obere Kirchgasse 2, 8401 Winterthur, Switzerland
2
Department of Informatics, University of Zurich, 8050 Zurich, Switzerland
Keywords:
Visual SLAM, Deep Learning-Based SLAM, Visual Impairment Navigation, Assistive Technology.
Abstract:
Despite advancements in SLAM technologies, robust operation under challenging condition such as low-
texture, motion-blur, or challenging lighting remains an open challenge. Such conditions are common in
applications such as assistive navigation for the visually impaired. These challenges undermine localization
accuracy and tracking stability, reducing navigation reliability and safety. To overcome these limitations, we
present SELM-SLAM3, a deep learning-enhanced visual SLAM framework that integrates SuperPoint and
LightGlue for robust feature extraction and matching. We evaluated our framework using TUM RGB-D, ICL-
NUIM, and TartanAir datasets, which feature diverse and challenging scenarios. SELM-SLAM3 outperforms
conventional ORB-SLAM3 by an average of 87.84% and exceeds state-of-the-art RGB-D SLAM systems
by 36.77%. Our framework demonstrates enhanced performance under challenging conditions, such as low-
texture scenes and fast motion, providing a reliable platform for developing navigation aids for the visually
impaired.
1 INTRODUCTION
Visual simultaneous localization and mapping (vi-
sual SLAM) plays a key role in computer vision and
robotics, enabling robots and autonomous vehicles to
map their surroundings while tracking their position.
Visual SLAM is applicable in various fields, including
enhancing navigation for blind and visually impaired
(BVI) individuals.
The field of SLAM for BVI navigation has un-
dergone significant advancements, from studies lever-
aging well-established SLAM techniques to devel-
oping new SLAM solutions tailored to the require-
ments of the visually impaired (Bamdad et al.,
2024). Some studies directly incorporated well-
known SLAM frameworks, such as ORB-SLAM and
RTAB-Map, without significant modifications or jus-
tification for their suitability. Several researchers
have utilized spatial tracking frameworks from exist-
ing platforms such as Google ARCore SLAM and
Intel RealSense SLAM. Some have developed cus-
tomized solutions tailored specifically for visually im-
a
https://orcid.org/0000-0003-2837-7881
b
https://orcid.org/0000-0002-1709-546X
c
https://orcid.org/0000-0002-7402-5206
paired navigations (Bamdad et al., 2024).
Despite these advancements, SLAM technologies
have certain limitations. The effectiveness of SLAM
systems often depends on distinct environmental fea-
tures that pose challenges in challenging environ-
ments (Zhang and Ye, 2017), (Jin et al., 2021). More-
over, frequent localization losses in areas lacking suf-
ficient feature points, such as blank corridors or plain
walls, compromise navigation reliability (Hou et al.,
2022). Achieving accurate and reliable localization,
especially for aiding the navigation of the visually im-
paired in challenging environments, remains an open
problem. Traditional SLAM algorithms often fail un-
der these conditions. Although recent deep-learning
SLAM models show promise, issues, such as compu-
tational complexity and data dependency, limit their
application in assistive technologies for BVI. The in-
tegration of learning-based modules into traditional
frameworks has improved efficiency but still requires
refinement for diverse environments.
SLAM methods can be classified as direct, opti-
cal flow, and feature-based. Although direct meth-
ods are robust in low-texture scenes, they struggle
in gradient-limited environments. Optical flow ap-
proaches handle environmental changes but are vul-
nerable to rapid movements or lighting fluctuations.
758
Bamdad, M., Hutter, H.-P. and Darvishy, A.
Deep Learning-Powered Visual SLAM Aimed at Assisting Visually Impaired Navigation.
DOI: 10.5220/0013338200003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages
758-765
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

Feature-based SLAM methodologies, on the other
hand, excel by identifying and tracking unique fea-
tures within an environment, thereby providing high
precision and efficiency.
Building on this understanding and motivated
by the goal of enhancing navigation aids for vi-
sually impaired individuals, we developed SELM-
SLAM3 to address the challenges faced by such ap-
plications. It is developed based on ORB-SLAM3,
a state-of-the-art feature-based visual SLAM sys-
tem. It integrates SuperPoint for robust feature ex-
traction and LightGlue for precise feature match-
ing, enhancing feature quality and matching pre-
cision. This contributes to accurate pose estima-
tion and improves localization and navigation relia-
bility, particularly under texture-limited or dynamic
lighting conditions. Evaluations of the TUM RGB-
D (Sturm et al., 2012), ICL-NUIM (Handa et al.,
2014), and TartanAir (Wang et al., 2020) datasets
demonstrated SELM-SLAM3’s superior tracking ac-
curacy and robustness compared to ORB-SLAM3
and state-of-the-art systems, consistently achieving
lower absolute trajectory errors (ATE). These re-
sults highlight the potential of improving naviga-
tion aids for BVI users in challenging scenarios.
The SELM-SLAM3 implementation is available at
https://github.com/banafshebamdad/SELM-SLAM3.
The key contributions of this study are as follows:
• Enhanced feature detection and matching: By
employing SuperPoint for feature extraction and
LightGlue for feature matching, SELM-SLAM3
enhances the quantity and quality of the detected
features and the precision of feature matching.
• Robust performance under adverse conditions:
The system demonstrated superior tracking capa-
bility, particularly in challenging conditions, such
as texture-poor scenarios.
• Adaptability and efficiency: The system’s adapt-
ability and efficiency in feature matching are sig-
nificantly improved by the capability of LightGlue
to dynamically adjust the matching process based
on the complexity of each frame.
• Superior comparative performance: Compared
with state-of-the-art systems, SELM-SLAM3
demonstrates superior performance in terms of
matching accuracy and pose estimation, high-
lighting its effectiveness.
2 RELATED WORK
2.1 SLAM-Based Navigation Solutions
for BVI
Visual SLAM technologies have been widely ex-
plored for the development of navigation aids for
visually impaired individuals. These solutions can
be categorized into three main approaches (Bamdad
et al., 2024). The first category leverages established
SLAM frameworks such as ORB-SLAM in their solu-
tions. For instance, (Son and Weiland, 2022) adopted
ORB-SLAM2 for crosswalk navigation and (Plikynas
et al., 2022) leveraged ORB-SLAM3 for indoor nav-
igation. However, these solutions are not specifically
designed to perform effectively in demanding envi-
ronments, such as low-texture areas or poor lighting,
which are typical navigation challenges for the visu-
ally impaired.
The second category includes solutions that uti-
lize spatial tracking frameworks from platforms such
as ARCore, ZED cameras, and Apple’s ARKit. For
example, (Zhang et al., 2019) took advantage of an
ARCore-supported smartphone to track the pose and
build a map of the surroundings in real-time. Al-
though these platforms provide ready-to-use SLAM
capabilities, they offer limited flexibility for optimiza-
tion and enhancement, making it difficult to adapt
them to the specific requirements of visually impaired
navigation.
The third category comprises customized SLAM
solutions tailored specifically for visually impaired
navigations. For example, the VSLAMMPT (Jin
et al., 2020) was developed to assist visually impaired
individuals in navigating indoor environments. These
specialized solutions often focus on addressing the
specific aspects of BVI navigation while using tra-
ditional SLAM techniques for localization and map-
ping, with their inherent limitations under challenging
conditions.
2.2 From Classical SLAM to Deep
Learning Techniques
Visual SLAM technologies have evolved signifi-
cantly, transitioning from traditional algorithms such
as ORB-SLAM3 (Campos et al., 2021), LSD-SLAM
(Engel et al., 2015), DSV-SLAM (Mo et al., 2021),
VINS-mono (Qin et al., 2018), and OpenVSLAM
(Sumikura et al., 2019), which rely on handcrafted
features and matching techniques to incorporate ad-
vanced deep learning-based approaches. Traditional
SLAM systems have been foundational, but they of-
ten encounter issues in challenging environments.
Deep Learning-Powered Visual SLAM Aimed at Assisting Visually Impaired Navigation
759

Recently, deep learning models have introduced
enhanced robustness and accuracy in SLAM applica-
tions. End-to-end deep learning visual odometry and
SLAM systems, such as DPVO (Teed et al., 2024),
DROID-SLAM (Teed and Deng, 2021), iDF-SLAM
(Ming et al., 2022), TrianFlow (Zhao et al., 2020),
DeepVO (Wang et al., 2017), and TartanVO (Wang
et al., 2021) have demonstrated potential for over-
coming traditional limitations. However, these sys-
tems occasionally encounter challenges. Issues such
as complexity, high computational demands, and data
dependency can limit their effectiveness. This be-
comes especially problematic for applications like
navigation aids for the visually impaired, where prac-
tical constraints on sensor carriage and computational
resources exist.
Another new trend in visual SLAM is the inte-
gration of learning-based modules with the traditional
SLAM frameworks. Systems such as SuperPointVO
(Han et al., 2020), SuperPoint-SLAM (Deng et al.,
2019) and SuperSLAM3 (Mollica et al., 2023) have
adopted deep learning for feature extraction, while re-
taining classical feature matching. However, the re-
sults showed that traditional matching approaches do
not effectively match the learning-based feature de-
scriptors provided by the deep learning-based feature
extraction models. (Fujimoto and Matsunaga, 2023)
and (Zhu et al., 2023) went one step further in this
area by incorporating deep-learning-based feature ex-
traction and matching into traditional SLAM frame-
works. They incorporated SuperPoint and SuperGlue
as the feature extraction and feature matching mod-
ules, respectively. However, the adaptability of Su-
perGlue to varying conditions leaves room for further
optimization.
In this study, we introduced a novel approach
by integrating SuperPoint and LightGlue into the
front end of the ORB-SLAM3 framework. Light-
Glue offers several enhancements over SuperGlue,
a previously established state-of-the-art method for
sparse matching. It is optimized for both memory
and computation, which are crucial for the real-time
processing requirements of SLAM systems. More-
over, LightGlue’s ability to adapt to the difficulty of
matching tasks and provide faster inference for intu-
itively easy-to-match image pairs makes it suitable for
latency-sensitive applications such as SLAM.
3 IMPLEMENTATION
The main objective of SELM-SLAM3 is to enhance
the feature detection and matching accuracy under
challenging conditions, such as changing lighting,
motion blur, and poor textures. As illustrated in Fig-
ure 1, the system builds on ORB-SLAM3, a state-of-
the-art SLAM framework known for its accuracy and
processing speed, with modifications focused on its
front end. The deep learning-based SuperPoint model
is used to identify and describe keypoints, ensuring
the reliable detection of high-quality features in di-
verse environments. LightGlue complements this by
accurately matching the features between consecutive
frames and keyframes, which are crucial for consis-
tent and precise tracking.
Integrating SuperPoint and LightGlue into ORB-
SLAM3 requires addressing the differences in data
formats, structures, and methodologies. These ad-
justments enabled the seamless incorporation of ad-
vanced feature detection and matching algorithms
into the SLAM framework. Essential libraries, in-
cluding OpenCV, were employed to manage the key-
points, descriptors, feature normalization, and data
access. The pre-trained SuperPoint and LightGlue
models, provided in the ONNX format, were exe-
cuted on a GPU for efficiency, while the other system
components ran on a CPU. These models, sourced
from the AIDajiangtang GitHub repository (AIDa-
jiangtang, 2023), delivered satisfactory test perfor-
mance, eliminating the need for fine-tuning at this
stage.
3.1 Feature Extraction
This study demonstrates the transition from tradi-
tional feature extraction methods to enhanced deep
learning models in visual SLAM systems. ORB-
SLAM3 uses a multiscale strategy to extract features
and detect FAST corners across eight scale levels
with a scale factor of 1.2. To ensure uniform fea-
ture distribution, it subdivides each scale level into
grids, identifying a minimum of five corners per grid
cell, and dynamically adjusting thresholds to detect
adequate corners, even in low-texture areas. The
orientation and ORB descriptors are then computed
for the selected corners, as detailed in (Mur-Artal
et al., 2015). In contrast, SuperPoint extracts features
from full-resolution images, capturing a broader array
of features while preserving the fine details missed
by ORB-SLAM3’s pyramid approach. Trained by
scale invariance and pattern recognition, SuperPoint
addresses the limitations of traditional handcrafted
methods. SELM-SLAM3 incorporates SuperPoint
for feature extraction, transforms each frame into
grayscale, and employs the SuperPoint detector to
identify key features, which are then utilized through-
out the SLAM pipeline. By leveraging SuperPoint’s
ability to recognize intricate patterns and their scale
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
760

Figure 1: SELM-SLAM3 front-end, highlighting the integration of SuperPoint and LightGlue. The green boxes indicate the
modified modules in ORB-SLAM3; the blue box represents the new module implemented in SELM-SLAM3; and the gray
boxes depict the original modules of ORB-SLAM3 without modifications.
invariance, SELM-SLAM3 captures high-quality fea-
tures, even under challenging conditions such as low
textures or poor lighting.
3.2 Feature Matching
SELM-SLAM3 employs the LightGlue model to
match features between frames. LightGlue is an ad-
vanced neural network designed to enhance the sparse
local feature matching across images. Building on the
achievements of its predecessor, SuperGlue, Light-
Glue combines the power of attention and transformer
mechanisms into the matching problem. LightGlue
can dynamically adjust its computational intensity
based on the complexity of each image pair owing to
its ability to assess the prediction confidence. This
adaptability allows LightGlue to surpass SuperGlue
in terms of speed, precision, and training ease, mak-
ing it an excellent choice for various computer vi-
sion applications (Lindenberger et al., 2023). The
substitution of ORB-SLAM3 traditional matching al-
gorithms with LightGlue necessitates certain adjust-
ments to the SLAM pipeline. The guided match-
ing approach, which involves projecting features and
searching for matches within a constrained window,
was eliminated to fully leverage the capabilities of
LightGlue. Similarly, computationally intensive Bag
of Words (BoW)-based matching was replaced, to en-
hance the feature matching process.
3.3 Tracking
Tracking is responsible for localizing the camera and
deciding when to insert a new keyframe, and relies
heavily on effective feature extraction and matching.
Robust feature extraction ensures that sufficient key-
points are available for tracking, even in challenging
scenarios. Efficient feature matching associates key-
points across frames, enabling accurate camera pose
estimation.
In ORB-SLAM3, tracking follows a constant-
velocity motion model to predict the camera pose us-
ing a guided search for the map points observed in
the last frame. Features are projected onto the current
frame using their 3D position and initial pose estima-
tion with matches searched within a small window.
However, inaccuracies in the motion model can result
in insufficient matches, requiring the system to com-
pute bag-of-words (BoW) vectors for the current and
reference keyframes, although this approach is com-
putationally demanding.
SELM-SLAM3 addresses these limitations by
bypassing 3D point projection, focusing on direct
feature matching between the previous and current
frames, increasing the number of matches, and im-
proving tracking accuracy. Regardless of the initial
method, the Track Local Map module is applied to se-
cure more matches. ORB-SLAM3 uses reprojection
to match local map points, whereas SELM-SLAM3
uses LightGlue to match these points directly with the
current frame. Only map points within the field of
view of the current frame were considered, and their
projected coordinates were treated as inputs to Light-
Glue.
The final tracking step involves deciding whether
to insert a new keyframe based on criteria such
as the number of frames since the last keyframe,
and whether the current frame tracks less than 90%
of its reference keyframe. SELM-SLAM3 inserts
fewer keyframes than ORB-SLAM3, while providing
a denser map because of its ability to extract more fea-
tures per frame and achieve more accurate matches.
These advancements from integrating SuperPoint and
LightGlue significantly enhanced the tracking perfor-
mance and map density in SELM-SLAM3.
Deep Learning-Powered Visual SLAM Aimed at Assisting Visually Impaired Navigation
761

4 EXPEREMENTS AND RESULTS
We evaluated SELM-SLAM3 on TUM RGB-D
(Sturm et al., 2012) freiburg1 sequences, ICL-NUIM
(Handa et al., 2014), and TartanAir (Wang et al.,
2020) datasets by running both the baseline ORB-
SLAM3 and the enhanced SELM-SLAM3 systems
on selected sequences. The absolute trajectory error
(ATE) was calculated by evaluating the Euclidean dis-
tances between the estimated camera poses and the
ground truth (Sturm et al., 2012). Tests were con-
ducted on a system with an Intel i9-12950HX CPU,
64 GB RAM, and an NVIDIA RTX A2000 8 GB
GPU running Ubuntu 20.04 LTS. We also compared
our results with those reported by (Fujimoto and Mat-
sunaga, 2023), (Li et al., 2021), (Whelan et al., 2015),
selected for their shared use of RGB-D sensors, test-
ing on similar datasets, and focus on improving track-
ing and mapping accuracy, aligning with the primary
goal of SELM-SLAM3.
4.1 TUM RGB-D Dataset Evaluation
The TUM RGB-D dataset includes color and depth
images from diverse indoor environments, captured
at 30 Hz with a resolution of 640 × 480. ORB-
SLAM3 was configured with eight scale levels, a
scale factor of 1.2, and 1000 features. For both ORB-
SLAM3 and SELM-SLAM3, the loop-closing mod-
ule was disabled to ensure a fair and focused compar-
ison of the core odometry capabilities with the results
reported in (Fujimoto and Matsunaga, 2023), which
describes an RGB-D odometry system. Both sys-
tems were evaluated on selected sequences and the
absolute trajectory error (ATE) was calculated using
the TUM RGB-D online tool. Figure 2 shows a sig-
nificant improvement in the ATE for SELM-SLAM3
compared to ORB-SLAM3. ORB-SLAM3 failed on
the ”floor” sequence, highlighting SELM-SLAM3’s
robustness. Table 1 demonstrates consistent RMSE
improvements, with SELM-SLAM3 outperforming
ORB-SLAM3 across the sequences. Sample trajec-
tories for the Freiburg1 desk and plant sequences are
shown in Figures 3a and 3b.
We also compared our system’s results with (Fu-
jimoto and Matsunaga, 2023) and (Whelan et al.,
2015). The method in (Fujimoto and Matsunaga,
2023) employs SuperPoint for feature extraction and
SuperGlue for feature matching, enhancing the accu-
racy and robustness of RGB-D odometry in challeng-
ing environments. (Whelan et al., 2015) uses RGB-D
sensors for real-time surface reconstructions, leverag-
ing volumetric reconstruction and a GPU-based 3D
cyclical buffer for unbounded space mapping. This
Figure 2: Comparison of ATE (m) on TUM RGB-D
sequences illustrates the enhanced accuracy achieved by
SELM-SLAM3 and the improvement over ORB-SLAM3.
(a) TUM f1 desk. (b) TUM f1 plant.
(c) ICL-NUIM lr-kt0. (d) ICL-NUIM or-kt2.
Figure 3: Sample of estimated trajectory and ground-truth.
method improves pose estimation with dense geomet-
ric and photometric constraints, and features efficient
map updates with space deformation for loop clo-
sure. Table 2 highlights SELM-SLAM3’s superior
performance across most of the tested Freiburg1 se-
quences in the TUM RGB-D dataset. SELM-SLAM3
achieved the lowest ATE in all sequences compared
with the deep-feature-based approach in (Fujimoto
and Matsunaga, 2023) and outperformed (Whelan
et al., 2015) in six out of seven sequences.
4.2 ICL-NUIM Dataset Evaluation
The ICL-NUIM dataset was used to benchmark the
RGB-D visual odometry and SLAM algorithms. Its
synthetic scenes and extensive low-textured surfaces,
such as walls, ceilings, and floors, pose specific chal-
lenges for tracking and mapping owing to the lack of
distinct features. The dataset includes living room se-
quences for benchmarking camera trajectory and re-
construction (with 3D ground truth, depth maps, and
camera poses), and office room sequences for camera
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
762

Table 1: Comparison of absolute trajectory error (ATE) in meters (m) on the TUM RGB-D dataset against baseline.
Freiburg1
ORB-SLAM3 SELM-SLAM3
RMSE % Boost
RMSE Mean Median S.D. RMSE Mean Median S.D.
360 0.207 0.2 0.201 0.054 0.175 0.16 0.147 0.071 15.46
desk 0.02 0.017 0.016 0.009 0.019 0.017 0.014 0.01 5
desk2 0.982 0.948 0.927 0.256 0.04 0.035 0.033 0.019 95.93
room 0.971 0.913 0.929 0.329 0.138 0.129 0.119 0.05 85.79
rpy 0.053 0.047 0.045 0.023 0.021 0.017 0.015 0.011 60.38
plant 0.329 0.243 0.171 0.221 0.034 0.03 0.025 0.015 89.67
teddy 0.632 0.528 0.355 0.347 0.131 0.104 0.076 0.079 79.27
floor x x x x 0.04 0.034 0.026 0.022 100
Avg. 0.456 0.414 0.378 0.177 0.08 0.07 0.061 0.036 66.44
Table 2: Comparison of ATE (m) on TUM RGB-D dataset.
DL-based: Deep Learning-based (Fujimoto and Matsunaga,
2023); Volumetric: Volumetric-based (Whelan et al., 2015).
NA indicates that ATE values were not reported.
Seq. SELM-SLAM3 DL-based Volumetric
desk 0.019 0.091 0.040
desk2 0.04 0.087 0.074
room 0.138 0.327 0.081
rpy 0.021 NA 0.031
plant 0.034 0.076 0.050
teddy 0.131 0.555 NA
floor 0.04 0.348 NA
trajectory benchmarking. We tested SELM-SLAM3
across all sequences of the ICL-NUIM dataset and
compared the results with (Li et al., 2021), which
introduced a robust RGB-D SLAM system tailored
for structured environments. This system enhances
the tracking and mapping accuracy by leveraging ge-
ometric features such as points, lines, and planes,
employing a decoupling-refinement method for pose
estimation with Manhattan World assumptions, and
using an instance-wise meshing strategy for dense
map construction. The results, shown in Figure 4,
and Table 3, demonstrate that SELM-SLAM3 outper-
formed both ORB-SLAM3 and the RGB-D SLAM
system from (Li et al., 2021). Figure 5 further high-
lights SELM-SLAM3’s superior tracking capability,
successfully handled texture-poor scenarios in which
ORB-SLAM3 experienced tracking loss. Sample tra-
jectories are shown in Figures 3c and 3d.
4.3 TartanAir Dataset Evaluation
The TartanAir dataset, with its extensive size and di-
versity, is a challenging benchmark comprising 1037
sequences, each with 500-4000 frames, totaling over
one million frames. Collected in photorealistic simu-
lation environments with realistic lighting, it includes
diverse motion patterns (e.g., hands, cars, robots, and
Figure 4: Comparison of ATE on ICL-NUIM dataset.
(a) Input. (b) ORB-SLAM3. (c) SELM.
Figure 5: Challenging scene in lr-kt3.
MAVs) across 30 environments spanning urban, rural,
natural, domestic, public, and science-fiction settings.
TartanAir categorizes sequences into Easy, Medium,
and Hard difficulty levels to evaluate SLAM systems
under varying complexities.
We focused on the ”Hospital” sequence at the
”Hard” difficulty level, which is relevant to visually
impaired navigation. A specialized tool was devel-
oped to convert TartanAir data into a format com-
patible with SELM-SLAM3 and ORB-SLAM3, al-
lowing custom dataset sequences. Additionally, the
depth information was modified to meet the system
requirements, and trajectory transformations aligned
the North-East-Down (NED) frame of TartanAir with
the reference frames of SELM-SLAM3 and ORB-
SLAM3 (Figure 6). This ensured compatibility of the
pose data with initial poses adjusted from TartanAir’s
coordinates (7, -30, 3) to (0, 0, 0).
The evaluation involved testing both systems on
selected sequences with ORB-SLAM3 configured us-
Deep Learning-Powered Visual SLAM Aimed at Assisting Visually Impaired Navigation
763

Table 3: Comparison of ATE (m) on ICL-NUIM dataset.
SELM: SELM-SLAM3; Recent SLAM (Li et al., 2021).
Seq. SELM Recent SLAM ORB-SLAM3
lr-kt0 0.006 0.006 0.603
lr-kt1 0.015 0.015 0.293
lr-kt2 0.012 0.02 0.826
lr-kt3 0.013 0.012 1.12
of-kt0 0.008 0.041 0.639
of-kt1 0.016 0.02 0.964
of-kt2 0.007 0.011 0.768
of-kt3 0.01 0.014 0.757
Avg. 0.0108 0.0173 0.746
Figure 6: Transformation between NED frame of TartanAir
and pose coordinate system in SELM-SLAM3.
ing eight scale levels, a scale factor of 1.2, and feature
point settings of 750, 900, and 1000. The results for
the ”Hospital-Hard” sequences are shown in Figure 7
and Table 4. SELM-SLAM3 exhibited significantly
lower ATE values, demonstrating its efficacy in the
challenging TartanAir environment.
Figure 7: ATE on the TartanAir Hospital sequences.
SELM-SLAM3’s advanced feature extraction and
matching yielded more accurate and abundant fea-
tures, enhancing the tracking accuracy with ATE be-
low 2 cm for ICL-NUIM, below 4 cm for TartanAir,
and below 8 cm for TUM RGB-D. In contrast, ORB-
SLAM3 showed a significantly lower accuracy, with
instances of tracking loss and complete failure in one
sequence.
Table 4: Comparison of ATE (m) on TartanAir dataset.
Systems P037 P038 Avg.
ORB-SLAM3 1200F 0.389 3.308 1.848
ORB-SLAM3 1000F 4.428 3.294 3.861
ORB-SLAM3 900F 5.298 4.167 4.733
SELM-SLAM3 0.049 0.020 0.035
5 CONCLUSIONS
This study introduced SELM-SLAM3, an enhanced
visual SLAM framework, as a foundational step
toward solutions for visually impaired navigation.
By incorporating advanced deep-learning-based tech-
niques, SuperPoint for feature extraction and Light-
Glue for feature matching, SELM-SLAM3 signifi-
cantly improves pose estimation and tracking accu-
racy. SELM-SLAM3 outperformed ORB-SLAM3
with an average improvement of 87.84% and sur-
passed state-of-the-art RGB-D SLAM systems with
an average enhancement of 36.77% in pose estimation
accuracy, particularly under challenging conditions,
such as low texture, rapid motion, and changing light-
ing. Comprehensive evaluations of the TUM RGB-
D, ICL-NUIM, and TartanAir datasets confirmed their
effectiveness. Its adaptability to diverse environments
underscores SELM-SLAM3’s potential for real-world
applications, particularly in aiding BVI navigation,
where reliability and precision are critical. However,
current evaluations rely on standard datasets, which
may not fully reflect the real-world complexities. Fu-
ture studies should test the system on tailored datasets
to ensure broader applicability. Future work could
also focus on optimizing computational performance
to ensure real-time applicability.
REFERENCES
AIDajiangtang (2023). Aidajiangtang’s github repository.
https://github.com/AIDajiangtang. Accessed: on 15
Novemner 2023.
Bamdad, M., Scaramuzza, D., and Darvishy, A. (2024).
Slam for visually impaired people: A survey. IEEE
Access.
Campos, C., Elvira, R., Rodr
´
ıguez, J. J. G., Montiel,
J. M., and Tard
´
os, J. D. (2021). Orb-slam3: An ac-
curate open-source library for visual, visual–inertial,
and multimap slam. IEEE Transactions on Robotics,
37(6):1874–1890.
Deng, C., Qiu, K., Xiong, R., and Zhou, C. (2019). Compar-
ative study of deep learning based features in slam. In
2019 4th Asia-Pacific Conference on Intelligent Robot
Systems (ACIRS), pages 250–254. IEEE.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
764

Engel, J., St
¨
uckler, J., and Cremers, D. (2015). Large-scale
direct slam with stereo cameras. In 2015 IEEE/RSJ
international conference on intelligent robots and sys-
tems (IROS), pages 1935–1942. IEEE.
Fujimoto, S. and Matsunaga, N. (2023). Deep feature-based
rgb-d odometry using superpoint and superglue. Pro-
cedia Computer Science, 227:1127–1134.
Han, X., Tao, Y., Li, Z., Cen, R., and Xue, F. (2020). Super-
pointvo: A lightweight visual odometry based on cnn
feature extraction. In 2020 5th International Confer-
ence on Automation, Control and Robotics Engineer-
ing (CACRE), pages 685–691. IEEE.
Handa, A., Whelan, T., McDonald, J., and Davison, A. J.
(2014). A benchmark for rgb-d visual odometry, 3d
reconstruction and slam. In 2014 IEEE international
conference on Robotics and automation (ICRA), pages
1524–1531. IEEE.
Hou, X., Zhao, H., Wang, C., and Liu, H. (2022). Knowl-
edge driven indoor object-goal navigation aid for vi-
sually impaired people. Cognitive Computation and
Systems, 4(4):329–339.
Jin, L., Zhang, H., and Ye, C. (2021). A wearable robotic
device for assistive navigation and object manipula-
tion. In 2021 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), pages 765–
770. IEEE.
Jin, S., Ahmed, M. U., Kim, J. W., Kim, Y. H., and Rhee,
P. K. (2020). Combining obstacle avoidance and vi-
sual simultaneous localization and mapping for indoor
navigation. Symmetry, 12(1):119.
Li, Y., Yunus, R., Brasch, N., Navab, N., and Tombari, F.
(2021). Rgb-d slam with structural regularities. In
2021 IEEE international conference on Robotics and
automation (ICRA), pages 11581–11587. IEEE.
Lindenberger, P., Sarlin, P.-E., and Pollefeys, M. (2023).
Lightglue: Local feature matching at light speed. In
Proceedings of the IEEE/CVF International Confer-
ence on Computer Vision, pages 17627–17638.
Ming, Y., Ye, W., and Calway, A. (2022). idf-slam: End-to-
end rgb-d slam with neural implicit mapping and deep
feature tracking. arXiv preprint arXiv:2209.07919.
Mo, J., Islam, M. J., and Sattar, J. (2021). Fast direct stereo
visual slam. IEEE Robotics and Automation Letters,
7(2):778–785.
Mollica, G., Legittimo, M., Dionigi, A., Costante, G., and
Valigi, P. (2023). Integrating sparse learning-based
feature detectors into simultaneous localization and
mapping—a benchmark study. Sensors, 23(4):2286.
Mur-Artal, R., Montiel, J. M. M., and Tardos, J. D. (2015).
Orb-slam: a versatile and accurate monocular slam
system. IEEE transactions on robotics, 31(5):1147–
1163.
Plikynas, D., Indriulionis, A., Laukaitis, A., and
Sakalauskas, L. (2022). Indoor-guided navigation for
people who are blind: Crowdsourcing for route map-
ping and assistance. Applied Sciences, 12(1):523.
Qin, T., Li, P., and Shen, S. (2018). Vins-mono: A robust
and versatile monocular visual-inertial state estimator.
IEEE Transactions on Robotics, 34(4):1004–1020.
Son, H. and Weiland, J. (2022). Wearable system to guide
crosswalk navigation for people with visual impair-
ment. Frontiers in Electronics, 2:790081.
Sturm, J., Engelhard, N., Endres, F., Burgard, W., and Cre-
mers, D. (2012). A benchmark for the evaluation of
rgb-d slam systems. In 2012 IEEE/RSJ international
conference on intelligent robots and systems, pages
573–580. IEEE.
Sumikura, S., Shibuya, M., and Sakurada, K. (2019). Open-
vslam: A versatile visual slam framework. In Pro-
ceedings of the 27th ACM International Conference
on Multimedia, pages 2292–2295.
Teed, Z. and Deng, J. (2021). Droid-slam: Deep visual slam
for monocular, stereo, and rgb-d cameras. Advances
in neural information processing systems, 34:16558–
16569.
Teed, Z., Lipson, L., and Deng, J. (2024). Deep patch visual
odometry. Advances in Neural Information Process-
ing Systems, 36.
Wang, S., Clark, R., Wen, H., and Trigoni, N. (2017).
Deepvo: Towards end-to-end visual odometry with
deep recurrent convolutional neural networks. In 2017
IEEE international conference on robotics and au-
tomation (ICRA), pages 2043–2050. IEEE.
Wang, W., Hu, Y., and Scherer, S. (2021). Tartanvo: A gen-
eralizable learning-based vo. In Conference on Robot
Learning, pages 1761–1772. PMLR.
Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang,
C., Hu, Y., Kapoor, A., and Scherer, S. (2020). Tar-
tanair: A dataset to push the limits of visual slam. In
2020 IEEE/RSJ International Conference on Intelli-
gent Robots and Systems (IROS), pages 4909–4916.
IEEE.
Whelan, T., Kaess, M., Johannsson, H., Fallon, M.,
Leonard, J. J., and McDonald, J. (2015). Real-time
large-scale dense rgb-d slam with volumetric fusion.
The International Journal of Robotics Research, 34(4-
5):598–626.
Zhang, H. and Ye, C. (2017). An indoor wayfinding sys-
tem based on geometric features aided graph slam for
the visually impaired. IEEE Transactions on Neural
Systems and Rehabilitation Engineering, 25(9):1592–
1604.
Zhang, X., Yao, X., Zhu, Y., and Hu, F. (2019). An arcore
based user centric assistive navigation system for vi-
sually impaired people. Applied Sciences, 9(5):989.
Zhao, W., Liu, S., Shu, Y., and Liu, Y.-J. (2020). Towards
better generalization: Joint depth-pose learning with-
out posenet. In Proceedings of the IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition,
pages 9151–9161.
Zhu, B., Yu, A., Hou, B., Li, G., and Zhang, Y. (2023). A
novel visual slam based on multiple deep neural net-
works. Applied Sciences, 13(17):9630.
Deep Learning-Powered Visual SLAM Aimed at Assisting Visually Impaired Navigation
765