Project Details

Abstract

Pavement condition assessment is essential for roadway asset management, yet current methods are fragmented, manual, and resource-intensive. Traditional workflows require separate tools for distress detection, severity and depth estimation, and overall condition classification, leading to inefficiencies. While deep learning models have emerged for tasks like crack segmentation or pavement condition index (PCI) prediction, most remain task-specific and lack integration. This project proposes a unified, end-to-end multitask framework using multimodal data for automated pavement assessment. By fusing high-resolution RGB images with stereo-derived depth maps, the system jointly performs distress detection, severity estimation, depth prediction, and condition classification. It also includes a natural language processing (NLP) module to generate human-readable reports tailored to agency workflows. The architecture features a shared encoder with four task-specific decoders, leveraging cross-task correlations to enhance generalization and reduce the need for separate models. Training will use a regional dataset annotated for all four outputs, with performance evaluated using intersection-over-union (IoU), mean absolute error (MAE), and F1-score. Report quality will be assessed using text similarity metrics and expert feedback. The central hypothesis is that multitask learning with multimodal inputs will improve accuracy and efficiency while reducing manual labor. This builds on the PI’s prior work in multitask PCI estimation, multimodal segmentation, and explainable reporting for infrastructure.

Project Word Files