Vehicular Road Crack Detection with Deep Learning: A New Online Benchmark for Comprehensive Evaluation of Existing Algorithms

TL;DR

The paper surveys state-of-the-art deep learning methods for road crack detection across supervised, unsupervised, semi-supervised, weakly supervised, and data-fusion paradigms, and introduces the UDTIRI-Crack benchmark as the first comprehensive online dataset for this task. It systematically compares a broad range of methods on UDTIRI-Crack and other public datasets, evaluating accuracy, efficiency, and generalizability, and also explores the feasibility of foundation models like SAM for this domain. Key findings show that crack-detection-specific models outperform general-purpose ones, yet generalization across diverse conditions remains challenging, highlighting the need for lightweight, label-efficient approaches and task-tuned foundation models. The work emphasizes the practical importance of scalable benchmarks and targeted methods for robust, real-time road condition assessment in intelligent road inspection systems. Overall, the study provides practical guidance for building next-generation road health assessment tools and identifies avenues where future research can improve generalizability and efficiency.

Abstract

In the emerging field of urban digital twins (UDTs), advancing intelligent road inspection (IRI) vehicles with automatic road crack detection systems is essential for maintaining civil infrastructure. Over the past decade, deep learning-based road crack detection methods have been developed to detect cracks more efficiently, accurately, and objectively, with the goal of replacing manual visual inspection. Nonetheless, there is a lack of systematic reviews on state-of-the-art (SoTA) deep learning techniques, especially data-fusion and label-efficient algorithms for this task. This paper thoroughly reviews the SoTA deep learning-based algorithms, including (1) supervised, (2) unsupervised, (3) semi-supervised, and (4) weakly-supervised methods developed for road crack detection. Also, we create a dataset called UDTIRI-Crack, comprising $2,500$ high-quality images from seven public annotated sources, as the first extensive online benchmark in this field. Comprehensive experiments are conducted to compare the detection performance, computational efficiency, and generalizability of public SoTA deep learning-based algorithms for road crack detection. In addition, the feasibility of foundation models and large language models (LLMs) for road crack detection is explored. Afterwards, the existing challenges and future development trends of deep learning-based road crack detection algorithms are discussed. We believe this review can serve as practical guidance for developing intelligent road detection vehicles with the next-generation road condition assessment systems. The released benchmark UDTIRI-Crack is available at https://udtiri.com/submission/.

Figures (6)

• Figure 1: Intelligent road inspection vehicles with automatic road crack detection systems.
• Figure 2: The overall outline of the reviewed computer deep learning-based road crack detection methods.
• Figure 3: Comparison results of eleven CNN-based and eight Transformer-based supervised semantic segmentation methods in terms of detection performance (IoU), resource consumption (model parameters) and computational complexity (FPS).
• Figure 4: Examples of experimental results of the compared methods on the proposed UDTIRI-Crack dataset: (a) ECSNet zhang2023ecsnet; (b) Deepcrack18 zou2018deepcrack; (c) Deepcrack19 liu2019deepcrack; (d) SCCDNet li2021sccdnet; (e) Crack-Att xu2023crack; (f) CDLN manjunatha2024crackdenselinknet; (g) SegDecNet++ tabernik2023automated; (h) LM-Net lu2024lm (i) SwinTransformer liu2021swin; (j) SegFormer xie2021segformer; (k) TransUnetchen2021transunet; (l) SCTNet xu2024sctnet; (m) LECSFormer chen2022refined; (n) CT-crackseg tao2023convolutional. The true-positive, false-positive, and false-negative pixels are shown in green, blue, and red, respectively.
• Figure 5: Examples of Grounded-SAM ren2024grounded on the AigleRN amhaz2016automatic, CrackNJ156 xu2022pavement, and proposed UDTIRI-Crack datasets: (a) Input road image; (b) Pixel-level label; (c) Results with text prompt "crack"; (d) Results with text prompt "damage"; (e) Results with text prompt "fissure"; (f) Results with text prompt "gap"; (g) Results with text prompt "road crack"; (h) Results with text prompt "split";