2019 CSCE Annual Conference - Laval (Greater Montreal)

2019 CSCE Annual Conference - Laval (Greater Montreal) Conference

Concrete Surface Defect Detection Using Deep Neural Network Based on LiDAR Scanning

Return to Session

Mr. Majid Nasrollahi, Concordia University (Presenter)
Ms. Neshat Bolourian, Concordia University
Dr. Amin Hammad

Structural inspection of bridges is essential to improve the safety of the infrastructure systems. Visual inspection is the principal method of detecting surface defects of bridges. In order to automate the process of structural inspection, it is important to collect proper data sets and use an efficient approach to analyze them and find the defects. Light Detection and Ranging (LiDAR) scanners can collect high-quality 3D point cloud datasets. Furthermore, Deep Neural Networks (DNNs) have been recently used for detecting 3D objects within 3D point clouds. This paper aims to develop a method for detecting concrete surface defects using a DNN based on LiDAR scanning. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are two popular types of DNNs. PointNet is a CNN used extensively for analyzing 3D point sets. PointNet has three parts: classification, part segmentation, and semantic segmentation. The semantic segmentation part is originally designed to detect indoor building elements. In this paper, PointNet is adapted to detect surface defects using point cloud datasets from scanning bridge surfaces. The reason for selecting PointNet is that it is robust to missing and corrupted data.

Training and testing datasets are collected from concrete bridges and annotated manually. Five point cloud datasets are prepared in five areas, resulting in 3,572 annotated segments. These segments are classified into two classes, defect and non-defect. In order to use K-fold strategy for training and testing, about 20% of the segments are assigned as a testing dataset. Promising initial results have been obtained in spite of the small size of the training dataset. In order to increase the accuracy, our future work aims to prepare a bigger dataset to train the model.