Combining Low-Light Scene Enhancement for Fast and Accurate Lane Detection

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Artificial Intelligence

Background:

• Lane detection is vital for autonomous vehicle navigation.
• Challenges include low-light, occlusions, and blurred lane lines, complicating feature identification.
• Existing methods struggle with poor visibility conditions.

Purpose of the Study:

• To develop an improved lane detection method for autonomous driving, specifically addressing low-light conditions.
• To enhance the robustness and accuracy of lane detection systems.
• To integrate low-light enhancement with advanced feature refinement techniques.

Main Methods:

• Proposed Low-Light Enhancement Fast Lane Detection (LLFLD) method.
• Utilized Automatic Low-Light Scene Enhancement Network (ALLE) for image preprocessing.
• Incorporated Symmetric Feature Flipping Module (SFFM) and Channel Fusion Self-Attention Mechanism (CFSAT) for feature refinement.
• Devised a novel structural loss function incorporating lane geometric priors.

Main Results:

• LLFLD demonstrated superior performance compared to state-of-the-art methods on the CULane dataset.
• Significant improvements were observed in both daytime and nighttime, particularly in challenging low-light scenarios.
• The integrated approach effectively enhanced contrast, reduced noise, and improved feature distinction.

Conclusions:

• The proposed LLFLD method offers a robust solution for lane detection under adverse lighting.
• The combination of low-light enhancement and advanced feature processing modules is effective.
• This work advances the safety and reliability of autonomous driving systems in diverse conditions.