Potholes and traffic signs detection by classifier with vision transformers

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Road accidents in India cause approximately 2500 annual fatalities, often linked to undetected potholes and traffic signs.
• Current detection methods fail to identify challenging conditions like water-filled, illuminated, or shaded potholes, and perspective or nighttime traffic signs.

Purpose of the Study:

• To develop a robust and accurate real-time detection system for potholes and traffic signs.
• To overcome limitations of existing methods in diverse and challenging environmental conditions.

Main Methods:

• A novel approach combining a cascade classifier for pattern identification and a vision transformer for detailed analysis and classification.
• Training and evaluation conducted on multiple datasets: ICTS, GTSRDB, KAGGLE, and CCSAD.

Main Results:

• The proposed method achieved high performance metrics, including a mean Average Precision (mAP) of 97.14% for traffic sign detection and 98.27% for pothole detection.
• Outperformed state-of-the-art techniques such as YOLOv3, YOLOv4, Faster RCNN, and SSD in recognition accuracy.

Conclusions:

• The integrated cascade classifier and vision transformer approach offers a significant advancement in detecting potholes and traffic signs.
• This technology holds great promise for enhancing the safety and reliability of driver assistance systems and autonomous vehicles.