Real-time detection of rare roadside obstacles using YOLOv8-n in autonomous vehicles

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous vehicles (AVs) face significant safety risks from rare road obstacles like traffic cones, fallen trees, debris, barrels, and rocks.
• Accurate and efficient detection of these varied obstacles is critical for safe AV operation, especially on edge devices with limited computational power.

Purpose of the Study:

• To develop a lightweight, real-time object detection framework capable of accurately identifying rare road obstacles for autonomous driving systems.
• To ensure the framework is suitable for resource-constrained hardware, prioritizing low latency and computational efficiency.

Main Methods:

• Utilized the YOLOv8-n model for its lightweight architecture and real-time performance.
• Combined and curated multiple open-source datasets containing annotated images of rare road objects.
• Employed transfer learning for model refinement and data augmentation (brightness fluctuation, rotation, flipping, geometric distortion) to enhance resilience to varying illumination and partial occlusion.

Main Results:

• Achieved high detection performance with 95.4% precision, 93.9% recall, 94.6% F1-score, and 98.1% mean average precision (mAP@0.5).
• Maintained a fast inference speed of 68 frames per second on a mid-range NVIDIA P100 GPU.
• Demonstrated the framework's effectiveness on resource-constrained hardware without requiring expensive computational resources.

Conclusions:

• The developed YOLOv8-n framework provides precise, real-time detection of rare road obstacles, making it suitable for edge-based autonomous driving systems.
• The system's efficiency and accuracy address the critical need for robust obstacle detection in safety-sensitive AV applications.
• This research contributes to the advancement of safer and more reliable autonomous navigation systems.