A Novel End-to-End Low-Light Image Segmentation Method and Its Application to Robotic Perception

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Semantic segmentation in low-light conditions is difficult due to poor visibility and reduced object contrast.
• Existing methods struggle with accurate object identification and texture differentiation in nighttime scenes.

Purpose of the Study:

• To develop an end-to-end trainable semantic segmentation network for low-light and nighttime scenes.
• To improve the accuracy and robustness of semantic segmentation in challenging visual environments.

Main Methods:

• A dual closed-loop bipartite matching algorithm was proposed, combining unsupervised illumination enhancement loss and supervised intersection-over-union loss.
• The Hungarian algorithm was used for joint minimization of losses, enabling end-to-end training.
• A weakly supervised method with selective refinement was employed using reference daylight images to boost performance in underperforming categories.

Main Results:

• The proposed network demonstrated superior performance on benchmark datasets like Cityscapes, Dark Zurich, and Nighttime Driving.
• Experimental results showed significant improvements over existing state-of-the-art methods in low-light semantic segmentation.
• The model achieved superior segmentation capabilities when deployed on quadruped robots in outdoor low-light environments.

Conclusions:

• The developed semantic segmentation network effectively addresses the challenges of low-light and nighttime scene perception.
• The proposed approach offers a robust solution for real-time semantic segmentation in autonomous systems operating under adverse lighting conditions.
• The study highlights the potential of the model for real-world applications in robotic perception and autonomous navigation.