PAIRS Lab

Journal Robotics and Autonomous Systems · 2026

SR-SLAM

Scene reliability-based RGB-D SLAM in diverse environments

Haolan Zhang1, Chenghao Li1, Thanh Nguyen Canh1, Lijun Wang1, Nak Young Chong1,2,*

1Japan Advanced Institute of Science and Technology   2Hanyang University   *Corresponding author

Paper arXiv Video
A real office scene: detected objects in green boxes, a moving person masked in red, features kept around the person, and the frame's scene-reliability score at the top left.
Real scene: the moving person is masked, nearby static features are kept, and the frame gets a reliability score.

01 · Abstract

How reliable is this scene for feature-based tracking?

In feature-based SLAM, the quantity and quality of extracted features drive performance, and both vary across environments. Current approaches face two challenges: limited adaptability in dynamic feature culling and pose estimation, and insufficient environmental awareness in assessment and optimization.

SR-SLAM is a scene reliability-based framework built on ORB-SLAM3. A unified scene reliability assessment combines several metrics with historical observations and guides the rest of the system: adaptive dynamic region selection with flexible geometric constraints, a depth-assisted self-adjusting clustering (adaptive DBSCAN) for dynamic feature removal, reliability-aware pose refinement that brings in a direct method when features are insufficient, and reliability-based keyframe selection with frame-weighted optimization.

On public datasets and in real-world scenarios, SR-SLAM improves accuracy by up to 90% over its ORB-SLAM3 baseline. It is not the most accurate method everywhere: the results explorer below shows each sequence, including those where a baseline wins.

02 · Video

SR-SLAM in four minutes

03 · Method

One reliability score guides the whole pipeline

SR-SLAM overview: RGB-D frames go to a YOLOv8 detection thread, the scene reliability assessment and ORB-SLAM3 tracking; reliability selects geometric constraints for potential dynamic regions, chooses depth-assisted adaptive DBSCAN or direct removal, triggers pose refinement, and drives new keyframe selection and frame-weighted optimization.
Fig. 1: overview of SR-SLAM.
  1. 1

    Detection thread

    SR-SLAM builds on ORB-SLAM3 and adds a YOLOv8 detection thread that runs alongside tracking.

  2. 2

    Scene reliability

    Each frame is scored from detection confidence, spatial distribution, feature quality and depth quality, weighted by its change from a reference frame.

  3. 3

    Adaptive dynamic regions

    Epipolar constraints are added only in GOOD (reliable) scenes; otherwise LK optical flow alone marks anomalous features.

  4. 4

    Dynamic feature removal

    Depth-assisted adaptive DBSCAN in low motion, direct removal in high motion.

  5. 5

    Pose refinement

    In BAD frames a direct method refines the pose, fused with the feature-based pose by a weight from the motion residual.

  6. 6

    Keyframes and weighting

    Reliability decides new keyframes and weights each frame in the optimization, instead of weighting single features.

04 · Dynamic culling

Remove the people, keep the features around them

Dynamic culling strategy: in low motion, points are clustered by depth and an adaptive DBSCAN picks the final dynamic regions; in high motion, potential dynamic points are removed directly.
Fig. 2: black points are static features, red points potential dynamic features; black circles are depth clusters, red circles the dynamic regions found by the depth-assisted adaptive DBSCAN.
Feature removal compared in a low dynamic scene (top) and a moderately dynamic scene (bottom): proportion-based, K-means, convex hull, standard DBSCAN and depth-assisted adaptive DBSCAN.
Fig. 3: from left to right, proportion-based, K-means, convex hull, standard DBSCAN and the depth-assisted adaptive DBSCAN. Rigid cluster shapes remove static features around the people; standard DBSCAN keeps them but leaves dynamic points on the people.

05 · Results

Accuracy on TUM, BONN and OpenLORIS

Pick a benchmark, a metric and a sequence. On TUM, SR-SLAM is compared with its ORB-SLAM3 baseline and three ORB-SLAM3-based dynamic SLAM methods; on BONN and OpenLORIS with the dynamic SLAM systems the paper reports.

ATE plots on five TUM sequences: ORB-SLAM3 (top) and SR-SLAM (bottom); ground truth black, estimate blue, error red.
Fig. 4: TUM, ORB-SLAM3 (top) vs SR-SLAM (bottom). Black: ground truth; blue: estimate; red: error.
2D trajectories of ORB-SLAM3, OVD-SLAM, SG-SLAM and SR-SLAM on eight OpenLORIS sequences.
Fig. 7: OpenLORIS trajectories. Black: ground truth; blue ORB-SLAM3, yellow OVD-SLAM, green SG-SLAM, red SR-SLAM. SG-SLAM fails on cafe1-2, home1-2 and corridor1-1.

06 · Ablation and runtime

What each module adds, and what it costs

m · lower is better · Table 10 (TUM) · SP = reliability assessment + pose refinement, D = dynamic culling, B = keyframe selection and weighted optimization

Overall runtime per frames / frame · average over six TUM sequences, same hardware (Table 12) · lower is faster

SR-SLAM runs at about 14 FPS, slower than ORB-SLAM3, OVD-SLAM and SG-SLAM; detection with geometric constraints is the main cost.

07 · Real world

Fixed camera and mobile robot, Intel RealSense D455

A fixed camera should not move, so any estimated motion is drift. Pick a scene and a drift measure (Table 13).

mm · lower is better

Real experimental settings: a fixed RealSense camera on a tripod and its scene (top), and the mobile robot with the camera and its scene (bottom).
Fig. 10: fixed-camera (top) and mobile-robot (bottom) setups. On the robot, only SR-SLAM produces a continuous straight-line trajectory (Fig. 12 of the paper).

08 · Cite

Citation

@article{zhang2026srslam,
  title   = {{SR-SLAM}: Scene reliability-based {RGB-D} {SLAM} in diverse environments},
  author  = {Zhang, Haolan and Li, Chenghao and Canh, Thanh Nguyen and Wang, Lijun and Chong, Nak Young},
  journal = {Robotics and Autonomous Systems},
  volume  = {197},
  pages   = {105306},
  year    = {2026},
  doi     = {10.1016/j.robot.2025.105306}
}