Enhanced nighttime nail detection using improved YOLOv5 for precision road safety.

Area of Science:

• Computer Vision
• Robotics
• Road Safety Engineering

Background:

• Nighttime road maintenance presents challenges for detecting and removing hazards like nails.
• Existing detection systems lack the necessary accuracy and efficiency for effective nighttime operations.

Purpose of the Study:

• To develop an enhanced nighttime nail detection system for improved road hazard recognition and retrieval.
• To increase the accuracy and efficiency of road nail detection and removal systems.

Main Methods:

• An improved YOLOv5 model incorporating modified C3 modules, reparametrized feature pyramid networks (RepGFPN), and optimal transport assignment loss (OTALoss).
• Deployment on an NVIDIA Jetson Orin Nano device with a stereo matching algorithm for synchronized recognition and localization.
• Integration with a binocular vision-based electromagnetic retrieval system and a ring marker system for automated operation.

Main Results:

• The enhanced YOLOv5 model achieved an average recognition accuracy of 91.5%, an 11.3% improvement over the original YOLOv5.
• Localization errors were maintained within 2.0 cm over a 120° field of view.
• The complete robot control system demonstrated retrieval and marking accuracies exceeding 98%.

Conclusions:

• The developed system significantly enhances nighttime road nail detection and retrieval capabilities.
• The system offers a practical and valuable solution for improving road traffic safety through efficient hazard removal.
• The integration of advanced AI models and robotic systems shows promise for future road maintenance technologies.