LOL-VO: Toward Robust Visual Perception and Ego-Motion Estimation in Low-Light Conditions

Abstract

Visual odometry is a fundamental technology for autonomous systems, including robotics and self-driving cars. However, its performance often degrades in low-light conditions due to inherently low contrast, significant noise amplified by high gains, and motion blur induced by long exposure times. To mitigate these issues, we introduce a robust monocular visual odometry pipeline, named LOL-VO, that integrates a novel coarse-to-fine low-light enhancement network with a low-light robust learning-based pose estimator. The enhancement network employs histogram equalization on down-scaled images and progressively restores details through learnable gamma correction, achieving not only high-speed enhancement but also effective noise suppression. The pose network first fuses illumination-robust global features from the DINOv3 foundation model with local convolutional features, then estimates low-resolution optical flow through iterative refinement, and finally regresses the camera ego-motion using a differentiable weighted epipolar optimization block. Experiments on synthetic/real-world and indoor/outdoor scenarios show that LOL-VO achieves competitive VO accuracy and efficient visual enhancement across challenging low-light conditions. The code is released on https://github.com/HITCSC/LOL-VO Note to Practitioners—Autonomous robots and vehicles operating in low-light conditions may face visual localization failures due to low-contrast, noise, and motion blur. This paper presents a practical visual odometry system designed specifically to improve robustness in low-light conditions. The proposed approach combines a lightweight image enhancement module with a learning-based ego-motion estimation pipeline that explicitly incorporates epipolar constraints. The key benefit of this work is improved estimation reliability of rotation and translation direction in scenarios where lighting conditions are limited, without requiring additional sensors such as LiDARs or infrared cameras. The system can be integrated as a drop-in front-end for mobile robots operating in poor illumination environments. Due to the frame-to-frame monocular implementation, current limitations include accumulated drift over long trajectories and lack of scale. Future works will consider incorporating loop closure and inertial sensing. Potential applications include low-light factory inspection, night-time navigation, and underground exploration.