Evaluating the Impact of Feature Extraction and Clustering Techniques in Highway Guardrail Classification

This paper presents an approach for the detection of highway guardrails using camera-based systems and advanced machine learning techniques. The proposed methodology combines feature extraction with Convolutional Neural Networks (CNN), specifically MobileNetV2, ResNet18, and VGG16, and clustering algorithms applied to these features. The effectiveness of the models is evaluated through clustering and classification metrics, with a particular emphasis on using the Gaussian Mixture Model (GMM) for forming more cohesive and well-separated clusters compared to K-means. The results indicate that the combination of ResNet18 with GMM provides high accuracy in distinguishing between concrete and metal guardrails, outperforming other tested combinations. This study contributes to the advancement of automatic guardrail detection on highways, providing insights for applications in road asset management.