Session: 03-04-03: Artificial Intelligent Applications in Manufacturing III

Paper Number: 165746

Automated Road Crack Detection and Quantification Using Detectron2: An Instance Segmentation Approach 

Automated Road Crack Detection and Quantification Using Detectron2: An Instance Segmentation Approach

Road surface deterioration, particularly cracking, poses significant challenges to transportation safety and infrastructure management. Early detection and precise quantification of these defects are crucial for proactive maintenance planning and optimal resource allocation. Traditional inspection methods rely heavily on manual visual assessments, which inevitably introduce inconsistency, inefficiency, and subjective judgment. This research presents an innovative application of deep learning for automated road crack detection and measurement using Detectron2, a powerful instance segmentation framework developed by Facebook AI Research.

The Detectron2 architecture employed in this study integrates a Feature Pyramid Network (FPN) with a ResNet backbone, creating a multi-scale feature representation that effectively captures cracks of varying sizes and patterns. This architectural design incorporates a Region Proposal Network (RPN) that identifies potential crack regions, followed by a mask branch that generates pixel-precise segmentation masks. This two-stage approach enables both localization and detailed morphological analysis of road surface defects.

For model development and evaluation, we compiled a comprehensive dataset comprising 305 high-resolution road surface images captured under diverse lighting conditions and surface textures. The dataset was strategically partitioned into 200 training images, 80 validation images, and 25 testing images to ensure robust model generalization. Precise annotation of crack regions was performed using the Makes Sense AI tool, establishing high-quality ground truth segmentation masks essential for effective supervised learning.

Performance evaluation using standard COCO metrics demonstrated exceptional results across multiple dimensions. The model achieved bounding box detection scores of AP: 48.72, AP50: 75.43, and AP75: 66.81, indicating strong performance in localizing crack regions. More importantly, segmentation performance reached AP: 49.05, AP50: 76.21, and AP75: 68.34, confirming the model's ability to precisely delineate crack boundaries at different intersection-over-union thresholds. With an overall confidence level of 99% in crack identification, the system substantially outperformed conventional computer vision approaches.

A distinctive contribution of this research was the implementation of quantitative crack assessment through area measurement. By incorporating a calibration reference object—a standard-sized square marker—in each image, we established a consistent scale for converting pixel measurements to physical dimensions. This methodology enabled precise calculation of crack areas, providing transportation authorities with actionable metrics for severity assessment and prioritization of maintenance interventions.

The system's ability to automatically detect, segment, and quantify road cracks represents a significant advancement in infrastructure inspection technology. By reducing reliance on subjective human assessment while maintaining high accuracy, this approach offers transportation departments a scalable solution for comprehensive road condition monitoring. The integration of such AI-driven inspection systems into maintenance workflows promises more efficient resource utilization, data-driven decision making, and ultimately, improved road infrastructure quality and safety.

 

Presenting Author: Sathish Kumar Gurupatham Kennesaw State University

Presenting Author Biography: N/A

Authors:

Moneesh Rajaram Kennesaw State University
Ryan Woodall Kennesaw State University
Diego Garcia Kennesaw State University
Sathish Kumar Gurupatham Kennesaw State University