EdgeCaseDNet: An enhanced detection architecture for edge case perception in autonomous driving

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Autonomous driving perception systems struggle with edge cases like adverse weather and multi-scale targets.
• Detection reliability is compromised in challenging environmental conditions.

Purpose of the Study:

• To develop an optimized object detection framework for autonomous driving edge cases.
• To introduce a novel edge case dataset with precise annotations for extreme road conditions.

Main Methods:

• Developed EdgeCaseDNet, extending YOLOv8 with Haar_HGNetv2 backbone, asymptotic feature pyramid network, hybrid partial depth-wise separable convolution, and Wise-IoU loss.
• Created a new edge case dataset encompassing fog, rain, snow, and nighttime conditions.

Main Results:

• EdgeCaseDNet demonstrated superiority over YOLOv8 in comprehensive evaluations.
• Achieved +10.6% mAP@50 and +8.4% mAP@[.5:.95] improvements.

Conclusions:

• EdgeCaseDNet significantly enhances object detection reliability in autonomous driving edge cases.
• The proposed framework and dataset contribute to more robust perception systems for autonomous vehicles.