EMG-YOLO: road crack detection algorithm for edge computing devices.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Engineering

Background:

• Road cracks reduce infrastructure lifespan and manual detection is inefficient.
• Existing YOLOv5 models face challenges with noisy data and high computational demands on edge devices.
• Edge computing devices require optimized models for real-time sensor data processing.

Purpose of the Study:

• To propose a novel road crack detection algorithm, EMG-YOLO, specifically designed for edge computing environments.
• To address the limitations of current models in handling poor-quality sensor data and computational burdens.
• To enhance the accuracy and efficiency of road crack detection on resource-constrained devices.

Main Methods:

• Implemented an Efficient Decoupled Header in YOLOv5 to separate classification and localization tasks, reducing computational load.
• Upgraded the loss function to MPDIOU (Minimum Point Distance Intersection over Union) to minimize bounding box point discrepancies.
• Integrated the GCC3 module for global context modeling, replacing traditional convolution to improve feature representation.

Main Results:

• EMG-YOLO demonstrated superior performance over mainstream algorithms, improving YOLOv5 accuracy by 2.7%.
• Achieved significant gains in mean Average Precision (mAP), with a 2.9% increase at IoU threshold 0.5 and 0.9% at 0.9.
• The model showed enhanced detection capabilities in complex environments on edge computing devices.

Conclusions:

• EMG-YOLO effectively tackles poor data quality and computational burden issues in edge-based road crack detection.
• The algorithm offers improved accuracy and efficiency, particularly in challenging real-world scenarios.
• Further research can focus on developing more lightweight and efficient object detection models for edge AI applications.