cd ~/

SLAM in Low Light: Pushing the Sensor Limits

Techniques for maintaining visual SLAM in challenging lighting conditions - from sensor improvements to algorithmic adaptations.

Evyatar Bluzer
2 min read

Low light kills visual SLAM. Fewer photons means noisier images, fewer features, more tracking failures. V2 must handle rooms lit by candles.

The Low-Light Challenge

At 1 lux (moonlit night):

  • Standard camera: ~10 photons per pixel per frame
  • SNR: under 10dB (signal buried in noise)
  • Feature detection: mostly false positives

At 100 lux (dim indoor):

  • ~1000 photons per pixel per frame
  • SNR: ~30dB (usable but not great)
  • Feature detection: functional but degraded

Current system fails below ~30 lux. V2 target: reliable at 3 lux.

Sensor-Level Improvements

Larger Pixels

More photon collection area. Trade-off: fewer pixels.

  • V1: 3μm pixels, 640x480
  • V2 option: 4μm pixels, 1280x720

Net: 4x more light per pixel, 4x more pixels total = 16x more photons.

Higher Quantum Efficiency

Better conversion of photons to electrons.

  • Standard silicon: 50-60% QE
  • BSI (backside illuminated): 70-80% QE

+30% more signal.

Lower Read Noise

Modern sensors achieving under 1e- read noise vs V1's 2-3e-.

At very low light, read noise dominates. Lower noise = usable signal at lower photon counts.

Global Shutter Considerations

Global shutter sensors typically have higher noise than rolling shutter. But rolling shutter causes artifacts with fast motion.

We need global shutter for tracking. Accepting some noise penalty.

Algorithmic Improvements

Feature Detection Adaptation

  • Lower detection thresholds: Accept more features, let matching filter bad ones
  • Larger patches: More pixels in descriptor = more robust to noise
  • Multi-scale: Features visible at different scales

Tracking Strategy Changes

  • Tighter IMU coupling: Trust IMU more when vision is poor
  • Feature lifetime tracking: Reuse features across more frames
  • Degraded mode: Reduce update rate, increase integration time

Learning-Based Features

Trained feature detectors may outperform hand-crafted in low light:

  • SuperPoint, D2-Net style architectures
  • Trained on low-light data (real + synthetic)

Exploring this for V2.

IR Illumination Option

Could we add IR illumination for tracking?

  • Pro: Solves low light completely
  • Con: Power consumption, battery impact, outdoor interference

Current direction: passive cameras with active assist for extreme cases.

Testing Protocol

New low-light test suite:

  • 1 lux controlled environment
  • Calibrated neutral density filters for repeatability
  • Standardized motion profiles
  • Pass criteria: under 3 tracking losses per 10 minutes at 3 lux

V2 sensors arriving for evaluation next month.

Comments