Vehicle Target Detection of Autonomous Driving Vehicles in Foggy Environments Based on an Improved YOLOX Network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Vehicle-mounted visual sensors face challenges in detecting targets within foggy environments.
• Degraded image quality due to fog leads to feature loss and reduced detection accuracy.

Purpose of the Study:

• To propose an improved YOLOX network for robust vehicle target detection in foggy conditions.
• To enhance the feature extraction and detection performance of deep learning models in adverse weather.

Main Methods:

• An improved YOLOX network incorporating fog-specific image characteristics into training.
• Integration of attention mechanisms and an image enhancement module into the YOLOX architecture.
• Optimization of the loss function tailored for foggy environments and application of transfer learning.

Main Results:

• The improved YOLOX network demonstrated significantly enhanced performance in foggy environments.
• Achieved superior mean Average Precision (mAP) compared to YOLOv5, YOLOv7, and Faster R-CNN.
• Showcased improved robustness and detection accuracy in challenging weather conditions.

Conclusions:

• The proposed method effectively addresses the limitations of vehicle target detection in fog.
• The enhanced YOLOX network offers a robust solution for real-world applications requiring reliable detection in adverse weather.
• The integration of fog-specific training and network optimizations leads to substantial performance gains.