Enhancing Road Safety Through Real-Time Unsupervised Anomaly Detection in Crash Barriers

Abstract

Real-time detection of damage in roadside crash barriers is vital for maintaining highway safety and enabling data-driven infrastructure management. Manual inspection methods are labor-intensive and impractical over long highway stretches. This paper presents a real-time unsupervised anomaly-detection framework based on deep convolutional autoencoders for identifying damaged crash barriers from roadway imagery. The proposed system integrates YOLOv10n for crash-barrier detection and the Segment Anything Model (SAM) for precise segmentation, isolating the barrier region before analysis. The autoencoder, trained exclusively on undamaged samples, classifies anomalies by comparing reconstruction errors against an adaptive threshold. A dataset comprising over 50,000 GoProcaptured images from Indian highways was used for training and validation. Experimental results demonstrate 97% accuracy and an AUC of 0.97, confirming the framework's ability to detect dents, cracks, and deformations under varying environmental conditions. The proposed approach eliminates the need for labeled data, offering a scalable, costeffective solution for autonomous road-safety monitoring and predictive maintenance.