Application of Traffic Cone Target Detection Algorithm Based on Improved YOLOv5

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Highway maintenance operations require efficient and automated traffic cone deployment and retraction.
• Existing systems may lack the speed and accuracy needed for real-time operations.
• Embedded systems offer potential for on-site processing but require optimized models.

Purpose of the Study:

• To develop and deploy a lightweight neural network for automated traffic cone recognition and positioning.
• To improve the automation level of highway maintenance operations using embedded devices.
• To meet the speed and accuracy requirements for traffic cone placement and retraction.

Main Methods:

• Utilized the lightweight YOLOv5-Lite-s neural network with a ShuffleNet backbone for feature extraction.
• Reduced computational complexity by replacing convolutional layers with focus modules and minimizing C3 layer usage.
• Deployed the optimized network on embedded devices for real-time traffic cone recognition and positioning.

Main Results:

• The YOLOv5-Lite-s network achieved approximately 89% recognition accuracy and 9 frames per second (fps) under varied conditions (distance, lighting, occlusion).
• The system met technical requirements for deploying/retrieving 30 cones per minute at a vehicle speed of 20 km/h.
• Demonstrated accurate and stable operation of the automatic traffic cone placement and retraction system.

Conclusions:

• The lightweight YOLOv5-Lite-s network effectively enables machine vision applications in traffic cone retraction operations.
• The developed system enhances highway maintenance automation with acceptable model inference accuracy and speed.
• The optimized neural network is suitable for deployment on embedded devices for real-time traffic management tasks.