Related Experiment Video
Updated: Aug 5, 2026
10:28
Compact Lens-less Digital Holographic Microscope for MEMS Inspection and Characterization
Published on: July 5, 2016
Lightweight Monocular Distance Estimation via Anisotropic Geometry Loss for Low-Light Driving Environments
Ricky Christanto1, Shaou-Gang Miaou1
1Department of Electronic Engineering, Chung Yuan Christian University, Taoyuan 320314, Taiwan.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a lightweight Anisotropic Geometry Loss (AGL) framework for robust monocular distance estimation in autonomous driving. The method enhances low-light performance and improves accuracy across various distances, crucial for safety systems.
Area of Science:
- Computer Vision
- Robotics
- Autonomous Systems
Background:
- Monocular distance estimation is vital for autonomous driving safety but struggles in low-light conditions.
- Current state-of-the-art models are computationally intensive and degrade performance in adverse lighting.
- Lightweight 2D detectors lack the geometric constraints for precise depth estimation.
Purpose of the Study:
- To develop a lightweight and robust monocular distance estimation framework for varying illumination conditions.
- To improve the accuracy and reliability of depth estimation in low-light environments for autonomous vehicles.
- To provide a computationally efficient solution for camera-only Autonomous Emergency Braking (AEB) systems.
Main Methods:
- Proposed the Anisotropic Geometry Loss (AGL) framework, enforcing ground-plane consistency via an anisotropic bottom-edge constraint.
- Integrated a Contrast Limited Adaptive Histogram Equalization (CLAHE) module for luminance-channel contrast enhancement at inference.
- Evaluated on the Dark-KITTI dataset, assessing performance across different distance ranges and illumination conditions.
Main Results:
- Achieved an RMSE of 10.91 ± 0.68 m, outperforming YOLOv10n and YOLOv26n in low-light conditions.
- Maintained real-time inference (>160 FPS) with a compact 2.71 M-parameter footprint.
- CLAHE further reduced RMSE to 10.55 ± 0.72 m and demonstrated significant improvements in the medium-distance range (15-30 m).
Conclusions:
- Explicit geometric constraints offer an effective and efficient solution for robust monocular distance estimation across varying illumination.
- The AGL framework demonstrates practical potential for camera-only AEB systems on edge-computing platforms, aligning with Euro NCAP safety protocols.
- The integration of CLAHE enhances low-light visibility and accuracy, making the system more reliable in challenging driving scenarios.
