Graph Attention Layer Evolves Semantic Segmentation for Road Pothole Detection: A Benchmark and Algorithms.

Area of Science:

• Computer Vision and Machine Learning for Road Infrastructure Monitoring

Background:

• Existing road pothole detection methods rely on computer vision or machine learning, with challenges in data annotation for deep learning models.
• While computer vision dominated research, machine learning-based methods, particularly using convolutional neural networks (CNNs), were less explored due to data limitations.
• A recent stereo vision dataset and disparity transformation algorithm improved pothole distinguishability, but benchmarks for state-of-the-art (SoTA) CNNs were lacking.

Purpose of the Study:

• To evaluate the performance of SoTA CNNs for semantic segmentation in road pothole detection using diverse image modalities.
• To introduce a novel Graph Attention Layer (GAL) integrated into CNNs to enhance feature representations for improved pothole detection.
• To establish a benchmark for road pothole detection using CNNs trained on RGB, disparity, and transformed disparity images.

Main Methods:

• Evaluation of nine SoTA CNNs designed for semantic segmentation on road pothole detection tasks.
• Development and integration of a novel Graph Attention Layer (GAL) inspired by Graph Neural Networks (GNNs) into CNN architectures.
• Comparative analysis of the proposed GAL-DeepLabv3+ against other SoTA CNNs using RGB images, disparity images, and transformed disparity images for training.

Main Results:

• The proposed GAL-DeepLabv3+ demonstrated the highest overall road pothole detection accuracy across all tested training data modalities (RGB, disparity, transformed disparity).
• Extensive experiments confirmed the effectiveness of GAL in optimizing feature representations for semantic segmentation in the context of pothole detection.
• The study provides a comprehensive benchmark for comparing different CNN approaches on a newly released stereo vision pothole dataset.

Conclusions:

• The novel Graph Attention Layer (GAL) significantly enhances the performance of CNNs for road pothole detection, outperforming existing SoTA methods.
• GAL-DeepLabv3+ represents a promising advancement for accurate and robust road surface monitoring systems.
• The publicly available dataset, source code, and benchmark facilitate further research and development in automated road inspection.