Object detection model design for tiny road surface damage.

Area of Science:

• Computer Vision
• Machine Learning
• Road Infrastructure Monitoring

Background:

• Current road surface damage detection methods lack generalization, struggle with small defects, and face challenges in balancing accuracy and computational efficiency.
• Effective detection is vital for highway maintenance and ensuring traffic safety.

Purpose of the Study:

• To develop a novel road surface damage object detection model (RSDD) that overcomes the limitations of existing methods.
• To improve the detection of tiny road surface damages and enhance the overall accuracy-speed trade-off.

Main Methods:

• Designed a specialized backbone for road surface damage feature extraction, addressing feature loss and tiny damage detection.
• Implemented multi-attention mechanisms for efficient feature fusion and a bi-directional feature fusion path for inter-stage information exchange.
• Constructed an enhanced feature pyramid and a multi-scale decoupled detection head for accurate detection of various damage sizes.

Main Results:

• The proposed RSDD model demonstrates significant advantages in detecting tiny road damages.
• Achieved 70.8% and 61.2% mAP50 on two datasets with low inference latency (4.5 ms) and 16.5M parameters.
• Outperformed YOLOv8s by 5.5% and 3.3% in detection accuracy while improving inference speed by 0.6 ms.

Conclusions:

• RSDD offers a superior balance of accuracy, scale, and speed for road surface damage detection.
• The model shows excellent generalization performance, particularly for detecting small-scale road damages.
• This research contributes a robust solution for efficient and accurate highway maintenance and traffic safety.