Adaptive ORB feature extraction running on the EuRoC V1_01_easy sequence.

Brightness-adaptive ORB feature extraction for improved robustness in visual SLAM systems operating under varying illumination conditions.

This project extends the ORB-SLAM3 feature extraction pipeline by dynamically adjusting FAST detection thresholds and the target number of ORB features based on real-time scene brightness. The adaptive approach increases feature density in darker environments while reducing redundant computation in well-lit scenes, maintaining minimal runtime overhead.

The adaptive module is integrated directly inside the ORB extractor and operates before feature distribution. Downstream SLAM components — tracking, local mapping, and loop closing — remain unchanged.

Standard ORB-SLAM3 uses fixed feature extraction parameters across all frames. However, lighting conditions strongly influence keypoint detection quality and tracking stability.

This project introduces a lightweight adaptive strategy that:

• Computes mean frame brightness in real time
• Normalizes brightness to a 0–1 range
• Dynamically adjusts nFeature counts
• Scales the ORB feature budget between 85%–115% of baseline

All modifications were designed to preserve the original ORB-SLAM3 architecture and threading model.

Key modifications were implemented inside the ORB extractor:

• Brightness estimation using cv::mean(image)
• Dynamic adjustment of target feature count:

nfeatures = nOriginalFeatures * featureMultiplier

Where:

featureMultiplier ∈ [0.85 , 1.15]

Behavior:

• Dark scenes → increased feature density
• Bright scenes → reduced feature count for efficiency

Observations

• Comparable performance on easier sequences
• Improved robustness on challenging lighting conditions (e.g., MH04 difficult)
• Negligible runtime overhead (<1% variation)

• Modified a production-scale visual SLAM pipeline (ORB-SLAM3)
• Implemented real-time adaptive parameter tuning with minimal runtime overhead
• Preserved compatibility with existing tracking, mapping, and loop-closing modules
• Designed changes to integrate cleanly into an existing C++ codebase

This repository does not include the full ORB-SLAM3 source code.

Clone ORB-SLAM3 separately:

git clone https://github.com/UZ-SLAMLab/ORB_SLAM3.git

Apply the adaptive modifications:

./scripts/apply_changes.sh /path/to/ORB_SLAM3

Then rebuild ORB-SLAM3 following the upstream instructions.

Datasets:

• EuRoC Machine Hall (MH01, MH04, MH05)
• EuRoC Vicon Room (V101, V103)

Metrics:

• Absolute Trajectory Error (ATE)
• Runtime
• CPU utilization
• Memory usage

Full analysis available in:

report/Report.pdf

This work builds upon:

ORB-SLAM3
https://github.com/UZ-SLAMLab/ORB_SLAM3

Original authors:
Campos, Elvira, Gómez Rodríguez, Montiel, Tardós

Brayden Currier
Computer Engineering