A Reference-Free Lens-Flare-Aware Detector for Autonomous Driving.

Area of Science:

• Computer Vision
• Autonomous Systems
• Optical Engineering

Background:

• Autonomous driving technology is advancing, increasing vehicle deployment in urban areas.
• Lens flare is an optical artifact that negatively impacts object detection accuracy in autonomous driving systems.
• Existing lens flare mitigation techniques are often not suitable for real-time applications.

Purpose of the Study:

• To propose a novel, lightweight lens flare perception network for autonomous driving.
• To address the limitations of current lens flare mitigation methods in real-time scenarios.
• To enhance the performance of object detection systems under challenging visual conditions.

Main Methods:

• Development of a reference-free lens flare perception model using a ResNet18 backbone and a Multi-Layer Perceptron (MLP).
• Utilizing a teacher-student framework for knowledge distillation from a reference-based model optimized with the Learned Perceptual Image Patch Similarity (LPIPS) metric.
• Integration of the lens flare perception network with a baseline object detection system without requiring additional hardware or complex pre-processing.

Main Results:

• The proposed lens flare perception network significantly improves the performance of the baseline object detection network.
• The method demonstrates superior performance compared to previous lens flare mitigation techniques.
• The lightweight nature of the network facilitates seamless integration into existing autonomous driving systems.

Conclusions:

• The developed lens flare perception network effectively alleviates the adverse effects of lens flare on object detection.
• The proposed solution is efficient, requires minimal computational resources, and is suitable for real-time deployment in autonomous driving.
• This approach offers a practical and effective way to enhance the safety and reliability of autonomous vehicles.