An Automatic Concrete Crack-Detection Method Fusing Point Clouds and Images Based on Improved Otsu's Algorithm

Xiaolong Chen, Jian Li, Shuowen Huang, Hao Cui, Peirong Liu, Quan Sun2021 · DOI

TL;DR

This paper introduces a novel method for automatically detecting cracks in concrete structures by fusing 3D point cloud data with 2D images. By combining the geometric accuracy of 3D scans with the rich texture information of images and employing an improved Otsu's algorithm for segmentation, the method achieves significantly higher detection accuracy and robustness compared to using either data source alone.

What Problem Does This Paper Address?

Detecting cracks in concrete structures is crucial for assessing their structural integrity and planning maintenance. Traditional methods relying solely on 2D images are often hampered by environmental factors like shadows, stains, and varying illumination, leading to false positives and missed detections. While 3D laser scanning offers more robust geometric data, its accuracy can be limited by scanner resolution and the inherent difficulty in capturing fine crack details. This paper addresses the challenge of achieving highly accurate and reliable concrete crack detection by overcoming the limitations of individual 2D image-based and 3D point cloud-based approaches.

Methods

The proposed method integrates 3D point cloud data and 2D images through a four-step process:

  1. Data Preprocessing: Both 2D images and 3D point clouds undergo initial processing. Images are converted to grayscale and denoised using a median filter. 3D point clouds are filtered and denoised, and then transformed into a depth image by fitting a plane to the concrete surface and projecting the point cloud data onto it.
  2. Fusion of 3D Point Clouds and 2D Images: This is a key innovation.
  3. Image Registration: The depth image derived from the point clouds is precisely aligned with the original 2D grayscale image using circular targets as control points and a bilinear interpolation method for geometric transformation.
  4. Pixel-Level Fusion: A weighted average method is employed to fuse the registered depth and grayscale images. This combines the complementary information, enhancing crack features while reducing noise and background interference.
  5. Crack Detection Based on Improved Otsu's Method: The fused image is processed using a novel, block-based Otsu's method. Instead of a single global threshold, the image is divided into sub-blocks, and adaptive thresholds are calculated for each. Crucially, local and global constraints (pixel ratio, local/global gray averages, standard deviation) are introduced to distinguish crack sub-blocks from background sub-blocks, preventing misclassification of background noise.
  6. Denoising and Connection of Cracks: After the initial rough detection, connected component labeling is used to remove noise based on shape and area constraints. Morphological operations (open and close) are then applied to remove crack burrs and connect fragmented cracks, refining the final crack detection.

Datasets

The paper does not explicitly mention training datasets. The evaluation was performed on concrete crack data collected from various pavements and walls at a university, covering an area of 3 square kilometers. This dataset included 3D point clouds acquired by both a handheld 3D laser scanner (HandySCAN700) and a terrestrial 3D laser scanner (RIEGL VZ-2000), along with high-resolution 2D images captured by a separate camera. Three typical crack types (transverse, longitudinal, and inclined) were used for evaluation.

Key Insights and Ideas

  • Synergistic Fusion: The core insight is that combining the geometric depth information from 3D point clouds with the textural and visual information from 2D images provides a more comprehensive representation of cracks than either modality alone. This fusion effectively compensates for the weaknesses of each individual method.
  • Improved Otsu for Robustness: The traditional Otsu's method struggles with complex backgrounds and noise. The proposed block-based approach with local and global constraints significantly enhances its ability to differentiate cracks from artifacts, leading to fewer false positives.
  • Scanner Choice Matters: The study highlights that the accuracy of the 3D point cloud data (influenced by scanner type and resolution) directly impacts the performance of the fusion method. Higher precision point clouds generally lead to better detection results.
  • Adaptive Weighting: The effectiveness of the fusion relies on carefully chosen weighting factors for depth and grayscale information, which may need to be adapted based on crack characteristics (width, environmental interference) and scanner parameters.

Key Results

The proposed method demonstrated superior performance compared to single 2D image-based and 3D point cloud-based methods:

  • Overall Performance: Achieved an average precision of 89.0%, recall of 84.8%, and an F1 score of 86.7%.
  • Comparison with Single Methods: The average F1 score was 19.1% higher than single 2D-image methods (67.6%) and 10.7% higher than single 3D point-cloud methods (76.0%).
  • Robustness: Significantly reduced false detection rates and improved accuracy in the presence of stains and shadows.
  • Crack Type Handling: Effectively detected transverse, longitudinal, and inclined cracks.
  • Comparison with Deep Learning: Outperformed the DeepCrack deep learning method in terms of accuracy, recall, and F1 score.
  • Width Error: For cracks with higher precision point clouds, the maximum width error was generally low, with relative errors within 7.3% for 2cm cracks and less than 2% for wider cracks.

Limitations and Future Work

  • Weighting Factor Optimization: The paper acknowledges that the optimal weighting factors for fusion are dependent on crack characteristics and scanner parameters. While guidance is provided, a more universally applicable method for determining these weights is a future research direction.
  • Scanner Efficiency vs. Accuracy: The study notes the trade-off between the high accuracy of handheld scanners and the efficiency of terrestrial scanners. Future work could explore combining the strengths of different scanner types.
  • Image Segmentation Algorithm: The authors suggest further research to improve the image segmentation algorithm.
  • Deep Learning Integration: The paper mentions the growing trend of deep learning in crack detection and proposes future research to integrate deep learning with their proposed fusion method.
  • Data Availability: The data is available on request from the corresponding author, but not publicly available due to privacy concerns.

Summary updated

BibTeX
@misc{chen2021automatic,
  author = {Xiaolong Chen and Jian Li and Shuowen Huang and Hao Cui and Peirong Liu and Quan Sun},
  title = {{An Automatic Concrete Crack-Detection Method Fusing Point Clouds and Images Based on Improved Otsu's Algorithm}},
  year = {2021},
  doi = {10.3390/s21051581},
}
Figure