Convolutional point transformer for semantic segmentation of sewer sonar point clouds

Abstract

The application of sonar technology in sewer inspections offers significant potential for improving inspection efficiency. However, the point cloud data obtained via sonar encounters challenges such as excessive noise, irregular spatial distribution, and imbalanced data distribution. This study introduces the Convolutional Point Transformer for Semantic Segmentation (CPTSS) approach, specifically tailored for the precise identification of sewer defects. The architecture of CPTSS features a streamlined encoder-decoder framework, where the encoder module effectively combines the strengths of point transformer and convolutional techniques. This integration optimizes the model's ability to extract both local and global features, capture remote contextual information, and improve overall learning performance. Additionally, an α-balanced focal loss is proposed to address the imbalanced data distribution during training. The CPTSS was validated through field testing. The resulting metrics, including macro precision, macro recall, macro F1 score, and mean Intersection over Union (MIoU), yielded impressive values of 0.9562, 0.9020, 0.9234, and 0.8662, respectively. Furthermore, the CPTSS outperforms state-of-the-art methods including Point Transformer, Randla-Net, and KPConv in terms of MIoU, and exhibits strong generalization capability across diverse sewer conditions. These findings highlight the CPTSS as a significant advancement in sonar-based sewer inspection method, with the potential to substantially reduce the time and resources required for accurate inspections.