Robust Object Detection in Adverse Weather Conditions: ECL-YOLOv11 for Automotive Vision Systems

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Intelligent transportation systems and autonomous driving rely heavily on visual perception.
• Adverse weather conditions degrade image quality, compromising object detection accuracy and traffic safety.
• Traditional detection models struggle with reduced contrast, blurring, and noise in challenging environments.

Purpose of the Study:

• To propose an enhanced object detection framework, ECL-YOLOv11, for robust performance in adverse weather.
• To improve both detection accuracy and real-time processing for in-vehicle perception systems.
• To provide a reliable solution for autonomous driving safety under challenging conditions.

Main Methods:

• Developed ECL-YOLOv11 integrating three modules: Convolutional Edge-enhancement (CE), Context-guided Multi-scale Fusion Network (AENet), and Lightweight Shared Convolutional Detection Head (LDHead).
• CE module fuses edge and convolutional features to preserve boundary information in low visibility.
• AENet enhances small/distant object perception via multi-scale fusion and context modeling.
• LDHead optimizes efficiency with shared convolutions and GroupNorm for real-time inference.

Main Results:

• ECL-YOLOv11 achieved mAP@50 of 62.7% and mAP@50-95 of 40.5%, outperforming baseline YOLOv11 by 1.3% and 0.8% respectively.
• Precision reached 73.1% with a processing speed of 237.8 FPS, demonstrating a balance between accuracy and speed.
• Ablation studies confirmed the effectiveness of individual modules and their synergistic integration.
• Qualitative results showed high-confidence detections and reduced errors in various adverse conditions.

Conclusions:

• ECL-YOLOv11 offers a reliable and adaptable foundation for all-weather perception in autonomous driving.
• The framework ensures operational safety and real-time responsiveness by enhancing object detection robustness.
• The integrated modules effectively address feature degradation, multi-scale challenges, and computational efficiency.