PaperlayerView exact revision on arXiv

Source-linked AI summary

SLAM in Low-Light Environments: Project Report

Oleh Basystyi, Anna Stasyshyn, Oleksandr Kosovan, Yaroslav Prytula

arXiv:2607.17699v1cs.ROcs.CV

TL;DR

Low-light conditions challenge RGB-only SLAM, while existing evaluations largely emphasize well-lit environments. This report benchmarks six paradigms on five LaMARia sequences and finds that Kimera-VIO alone completes every track with low relative pose error, whereas other systems trade tracking robustness for accuracy or fail.

  • Problem

    Most SLAM benchmarks use well-lit indoor or daylight conditions, leaving performance under degraded low-light visual signals insufficiently characterized.

  • Method

    The report benchmarks six representative feature-based, direct, filter-based, and learning-based systems on five varied-illumination LaMARia sequences.

  • Results

    Kimera-VIO covers all five tracks with low relative pose error, while DPVO and DPV-SLAM avoid tracking failure at reduced accuracy and OpenVINS, DSO, and ORB-SLAM3 fail on harder cases.

  • Takeaways & Limitations

    The report links system performance to paradigm assumptions and identifies algorithmic additions for improving robustness in low-light conditions.

  • Takeaways & Limitations

    DPVO and DPV-SLAM are limited by monocular, inertial-free scale ambiguity and a training domain gap from well-exposed imagery.

Abstract

from arXiv · show

Simultaneous localization and mapping (SLAM) is one of the fundamental problems in robotics, as it enables autonomous operations in real-world scenarios. Under low illumination, reduced contrast, sensor noise, and motion blur degrade both feature extraction and feature matching, while compensating with LiDAR, depth, or thermal sensors raises cost, power draw, and integration complexity. Existing benchmarks remain dominated by well-lit indoor or daylight sequences, leaving open how far SLAM with standard RGB cameras can be pushed in the dark. We benchmark six systems spanning the feature-based, direct, filter-based, and learning-based paradigms - ORB-SLAM3, DSO, Kimera-VIO, OpenVINS, DPVO, and DPV-SLAM - on five LaMARia sequences of varying difficulty and illumination, reporting absolute and relative pose error alongside control-point recall. Kimera-VIO is the only system to track all five sequences to completion, combining the lowest relative pose error with steadily growing absolute error due to the absence of loop closure; DPVO and DPV-SLAM never lose tracking but incur absolute errors of roughly 100 m under low light; and the classical monocular pipelines (ORB-SLAM3, DSO) together with the filter-based OpenVINS fail outright or diverge on most of the harder and low-light sequences. The results suggest that RGB-only SLAM maintains stable low-light tracking only when both inertial fusion and global optimization are present. Closing the remaining gap will likely require low-light-specific learned front-ends or a return to complementary sensing.

1 Introduction

SLAM enables autonomous robots to estimate their position while constructing a map of the surrounding environment. For UGVs, continuous state estimation supports navigation, while mapping enables exploration, obstacle avoidance, and adaptive path planning.

  • SLAM Foundations: SLAM jointly estimates a robot’s location and generates a map representation of its environment for autonomous navigation.This state estimation problem includes both the agent’s pose and environmental landmarks.
  • SLAM Foundations: Filtering-based SLAM methods update joint landmark-and-pose states but face rapidly increasing computational costs and random-walk noise as the state vector grows.These effects lead to drift and motivated a shift toward optimization-based or keyframe-based systems.
  • UGV Navigation: UGVs require continuous state estimation to determine their current position throughout an operation, even with known starting coordinates and a predefined destination.Reliable localization is described as SLAM’s core operational requirement.
  • UGV Navigation: SLAM-generated maps support path planning that lets UGVs explore, recalculate routes, and adapt trajectories around unknown or moving obstacles.Third-party path-planning algorithms use the map to respond to unpredictable environments.
  • Mapping Requirements: Sparse maps suffice for camera-pose tracking but lack the structural detail required for physical-barrier detection, obstacle avoidance, and path replanning.Successful avoidance and replanning therefore require dense or semi-dense scene representations.

2 Motivation

UGVs must operate in low-light and degraded environments where standard visual SLAM is least reliable, while specialized sensors can impose prohibitive cost, power, and integration burdens. This motivates benchmarking RGB-only SLAM across paradigms and sequence conditions to assess localization, mapping, degradation, and algorithmic mitigation.

  • Operational motivation: UGV applications require 24-hour autonomy in power outages, dawn or dusk, and collapsed or smoke-filled structures with absent or inconsistent lighting.These scenarios require dense and semi-dense maps for path planning despite severely degraded visual signals.
  • Benchmark gap: Prior benchmarks mainly use well-lit indoor or daylight sequences and do not systematically combine ultra-low illumination, dust haze, and motion blur.The cited examples include TUM RGB-D, EuRoC MAV, and KITTI.
  • Sensing trade-offs: Adding LiDAR, depth cameras, or thermal imagers increases cost, power consumption, and integration complexity, potentially prohibiting lightweight UGV deployments.The passage frames specialized sensing as a costly way to address the low-light gap.
  • Research questions: The report asks whether RGB-only SLAM remains stable and accurate, how performance degrades across four paradigms, and how far algorithmic choices can close the low-light gap.RQ1 and RQ2 are addressed by benchmarking six systems on five LaMARia sequences, while RQ3 relates failure modes to algorithmic mitigations.

3 Background

SLAM estimates a robot trajectory and map from controls and observations by combining motion and observation constraints. Its modern architecture separates front-end tracking from back-end mapping, with visual methods and map optimization organized into complementary paradigms.

  • SLAM formulation: SLAM estimates the robot trajectory x_1:t and map m from controls u_1:t and observations z_1:t.The formal objective is to compute the posterior over trajectory and map.
  • SLAM formulation: The SLAM formulation factorizes into motion constraints between consecutive poses and observation constraints linking poses with the map.The objective is then to find the maximum a posteriori (MAP) estimate.
  • System architecture: Modern SLAM pipelines separate localization in the front-end from mapping in the back-end.The front-end abstracts raw sensor inputs for tracking and incremental localization, while the back-end generates a consistent map and optimizes the trajectory.
  • Front-end localization: Visual localization methods are classified as indirect feature-based methods that minimize reprojection error or direct appearance-based methods that minimize photometric error.Both estimate inter-frame motion, but they use different optimization constraints.
  • Front-end localization: Loop closure detects revisited areas and corrects accumulated drift through image representation, candidate selection, and geometric verification.Loop closure and relocalization are additional front-end modules supporting global consistency.
  • Back-end mapping: Back-end mapping uses filter-based or optimization-based approaches, with filters compacting past measurements and optimization solving keyframe constraint graphs.Filter-based methods accumulate noise over time, whereas optimization-based systems are the current de facto standard.

4 Methods overview

The methods overview traces SLAM from feature detection and descriptor matching to learned front-ends, emphasizing the trade-offs between robustness, speed, and hardware cost. It also explains why low-light degradation and dynamic scenes undermine correspondence-based localization and mapping.

  • Feature extraction: Feature detection identifies interest points, while description represents them for matching across views, with robustness to rotation, scale, and illumination as key goals.Features include distinctive corners, edges, and textures; detection and description are treated as separate steps.
  • Feature extraction: FAST and ORB use fast local-intensity tests, whereas SIFT and SURF improve scale and luminance invariance at substantial computational cost.SIFT and SURF can produce higher-quality matches than ORB in some scenarios, but their overhead limits real-time robotic use.
  • Learned front-ends: Learned extractors and matchers model landmarks under difficult appearance changes and spatial relationships, but their combination increases computational overhead and hardware requirements.SuperPoint targets shifts in perspective, illumination, and color appearance, while LightGlue uses graph neural networks and self-attention.
  • Feature matching: Classical matching compares descriptors using Hamming or L2 distance, while approximate nearest-neighbor methods use structures such as KD-trees to accelerate search.These methods trade a negligible amount of precision for faster matching as landmark counts grow.
  • Low-light challenges: Low light reduces contrast, increases sensor noise and motion blur, and can distort color, causing feature mismatches that produce drift or complete tracking failure.The transition from 300–500 lux to 10–50 lux, or below 1 lux, progressively breaks assumptions used by traditional visual SLAM algorithms.
  • Robust estimation: Dynamic objects and non-static scene elements can invalidate feature correspondences, motivating rejection mechanisms and robust optimization to preserve localization and mapping.The proposed combination of learned dynamic-object detection and adaptive robust optimization is intended to maintain real-time performance in dynamic environments.

5 Experiments

Experiments benchmark six visual, visual-inertial, and learning-based systems on five LaMARia sequences spanning varied difficulty and visual degradation. Kimera-VIO achieves full coverage with low local error but growing global error, while DPVO and DPV-SLAM remain robust and classical or filter-based methods often fail under challenging conditions.

  • Dataset and evaluation: LaMARia was selected because it benchmarks visual-inertial SLAM under realistic egocentric conditions with broad environmental and illumination variation.The evaluation used sequence_4_10, sequence_4_11, R_03_easy, R_06_medium, and R_09_hard.
  • Systems evaluated: The evaluation spans feature-based ORB-SLAM3, direct DSO, optimization-based Kimera-VIO, filtering-based OpenVINS, and learning-based DPVO and DPV-SLAM.DPV-SLAM extends DPVO with loop closure and global optimization, enabling a direct comparison between the learning-based odometry and SLAM systems.
  • Results: Kimera-VIO provides full coverage on all five tracks, with competitive ATE and low RPE that remains stable as ATE grows with trajectory length.The growing ATE is attributed to accumulated error that cannot be canceled because the evaluated configuration lacks loop closure.
  • Results: OpenVINS fails severely on sequence_4_10 and R_09_hard, while DSO and ORB-SLAM3 are undermined by motion blur, illumination change, low texture, and insufficient parallax.OpenVINS shows signs consistent with filter divergence; ORB-SLAM3 recovery after tracking loss is unstable because a new map starts at arbitrary scale.
  • Results: DPVO and DPV-SLAM avoid tracking failure despite lacking inertial priors, but their robustness comes with lower accuracy than Kimera-VIO.DPV-SLAM retains DPVO’s real-time patch-based tracking while adding loop closure and global optimization.

6 Limitations

The benchmark exposes method-specific limitations rooted in core assumptions, from feature scarcity and brightness-constancy violations to filter divergence and unbounded global drift. Evaluation caveats—including subset selection, synthetic low-light degradation, and Sim(3) alignment—limit generalization to deployed UGVs.

  • Method-specific limitations: Feature scarcity under low illumination, motion blur, low texture, and near-pure rotation causes ORB-SLAM3 to lose tracking and reinitialize maps at arbitrary scale.Its indirect front-end requires repeatable, well-localized keypoints and sufficient parallax; monocular tracking nearly fails on all but the easy sequence.
  • Method-specific limitations: Brightness-constancy violations from auto-exposure changes, sensor noise, and motion blur undermine DSO’s photometric alignment, producing large ATE and failures on difficult sequences.The residual no longer reliably represents geometric misalignment when scene intensity changes across frames.
  • Method-specific limitations: Short-lived, outlier-ridden feature tracks under handheld motion, occlusions, and moving platforms cause OpenVINS’s MSCKF filter to diverge rather than merely drift.On the hard and one low-light sequence, OpenVINS produces no usable initial interval.
  • Method-specific limitations: Kimera-VIO’s low local error but growing global drift results from visual-inertial odometry without loop closure, motivating full-SLAM loop closure and pose-graph optimization.The passage links low RPE to IMU-supported local estimation and growing ATE to uncorrected accumulated error; Table 3 shows high CP@1m but collapsed R@5m.
  • Evaluation limitations: The evaluation uses only a subset of LaMARia, partly synthetic low-light degradation, and Sim(3)-aligned monocular comparisons, limiting generalization of rankings to deployed UGVs.These caveats affect how the reported ATE, RPE, and control-point recall should be interpreted.

7 Conclusions

The report surveys modern SLAM, motivates low-light UGV navigation, and benchmarks six representative methods on LaMARia. Future work targets proposed improvements and a UGV-specific reduced-illumination dataset for more focused evaluation.

  • Conclusions: The report connects six SLAM methods’ performance on LaMARia to the assumptions underlying their paradigms.It also surveys algorithmic additions intended to improve low-light robustness.
  • Conclusions: Future work includes implementing the proposed improvements within the evaluated SLAM approaches.These improvements are intended to address robustness in low-light conditions.
  • Conclusions: A planned UGV-specific dataset will target reduced illumination and enable more focused evaluation of modern systems.It may also reveal domain challenges absent from currently available datasets.

A SLAM Taxonomy

The section presents a taxonomy of standard modules used in SLAM pipeline implementations, following the framework outlined by Álvarez Tuñón et al. (2024).

  • Standard SLAM modules: Figure 1 outlines the standard modules used in SLAM pipeline implementations.The figure is based on the framework outlined in Álvarez Tuñón et al. (2024).
  • Standard SLAM modules: The taxonomy organizes SLAM implementations around their standard pipeline modules.The supplied passage identifies Figure 1 as presenting these modules.
  • Standard SLAM modules: Álvarez Tuñón et al. (2024) provide the framework used to outline the standard SLAM pipeline modules.This attribution is given in the Figure 1 caption.

B Additional evaluation · B.1 Metrics

The evaluation measures local accuracy at surveyed control points and global trajectory consistency using CP@1m and R@5m. Together, these metrics distinguish localized consistency from drift or uniform scale and orientation errors.

  • B.1 Metrics: LaMARia evaluates both local accuracy at known surveyed locations and global consistency across the estimated pose sequence.Control points assess local accuracy, while broader pose-sequence evaluation assesses global trajectory consistency.
  • B.1 Metrics: Each sequence provides K geo-referenced control points whose positions are precisely surveyed, including through GNSS-RTK measurements.Fiducial markers enable automatic detection of these control points in imagery.
  • B.1 Metrics: Estimated control-point positions are triangulated from keyframe observations by minimizing reprojection error.The estimates are then aligned with surveyed positions using a robustly optimized Sim(3) transformation.
  • B.1 Metrics: CP@1m reports the fraction of aligned control-point estimates within 1 metre of surveyed positions.Because alignment is estimated from the control points themselves, CP@1m measures local accuracy and is insensitive to drift elsewhere.
  • B.1 Metrics: Pseudo-ground-truth poses fuse visual, inertial, and control-point information through joint optimization to evaluate global trajectory consistency.These poses are matched to estimated keyframes within a temporal tolerance.
  • B.1 Metrics: R@5m reports the fraction of estimated poses whose locations align well with pseudo-ground-truth under the control-point-derived Sim(3) alignment.The metric uses the translational component of each transformation.

B.2 Results

Results for the surveyed-control-point sequences use CP@1m and R@5m, with association-based comparison preferred because strict 1 ms timing deflates coverage through timegrid mismatch. A looser ∼100 ms covariance tolerance restores coverage near 100% for systems tracking the full sequences.

  • Control-point evaluation: CP@1m and R@5m are reported for the two sequences with surveyed control points.Table 4 further examines the R@5m denominator and tolerance choices for systems producing usable estimates.
  • Tolerance sensitivity: Roughly 50–80% coverage occurs under the strict 1 ms full column for Kimera-VIO and DPVO because of timegrid mismatch alone.The passage identifies this deflation as an evaluation-timing artifact rather than tracking loss.
  • Tolerance sensitivity: Near 100% coverage returns with the looser ∼100 ms cov tolerance, where cov ≈assoc supports full-sequence tracking and makes assoc the fair comparison figure.This interpretation applies to the systems with usable estimates in the reported evaluation.

C Trajectories

Trajectory visualizations show that scale observability governs long-sequence performance: OpenVINS and DSO suffer catastrophic geometric failures, whereas Kimera-VIO remains topologically correct with gradual drift. These failures reflect the absence of mechanisms to correct incorrect visual scale without loop closure or IMU integration.

  • OpenVINS: OpenVINS tracks correctly for the first 150–200 meters on sequence_4_10 before crossing a discrete scale observability threshold and permanently entering a warped coordinate frame.The method abruptly veers off-map mid-sequence and has no recovery mechanism.
  • DSO: DSO develops oscillating scale estimates on sequence_4_10 and sequence_4_11, causing camera poses to bounce between over- and under-estimation and destroying topology.The failure originates in early depth-estimation bias propagated throughout each sequence.
  • OpenVINS: OpenVINS on sequence_4_11 maintains good local tracking but weaves perpendicular to the actual path while accumulating large global error because of anisotropic scale bias.Forward motion is estimated more accurately than lateral motion, producing systematic perpendicular weaving.
  • Kimera-VIO: Kimera-VIO maintains topological correctness on all sequences, staying on-path with minor jitter and gradual ATE growth rather than catastrophic failure.Its trajectories show graceful degradation without spirals.
  • Cross-method interpretation: Scale observability limits all methods on longer sequences, and incorrect visual scale cannot be corrected without loop closure or IMU integration.The passages identify scale as the fundamental constraint determining method performance.
Loading 2607.17699v1…