A Study on Surface Anomaly Detection for Metro Vehicle Underbody Parts Based on 3D Point Cloud

Abstract

The surface anomaly detection of train parts is one of the key technologies to ensure the safe operation of a metro system. Its core often lies in image processing. However, some anomalies do not exhibit significant changes in color, shape, and texture features, making detection difficult on 2D images. 3D point clouds add depth information on top of 2D images. This can help reflect the surface features of components more precisely and contribute to accurate anomaly detection. This paper addresses the common feature that most metro vehicle underbody components are rigid structures, combined with the 3D point cloud processing technology, and induces a general anomaly detection method. This method adopts point cloud preprocessing, point cloud registration, etc., and can detect changes in the depth information of the surface of subway underbody rigid structure components. However, it doesn't perform well when processing parts with high-ablaze reflective surfaces and mesh surfaces. Therefore, the paper uses the High Voltage Distribution Box (HVB) joint area and the hollow reactor oblique area as examples to improve this method. For the HVB joint area, regional segmentation is first used to obtain pure joint and wire harness regions; then, the statistical method is used to complete the detection of the former, and the general detection method is used to complete the detection of the latter. For the hollow reactor inclined surface area, plane filtering is used to adjust the process of the general detection method, eliminating the interference of noise points behind the grid plane and avoiding the problem of poor point cloud registration caused by plane filtering. Finally, using actual data from a metro company as a test set, it was verified that the new method, while ensuring a detection rate not lower than the original method, eliminates over 90% of false alarms, significantly reducing false detection rates. This offers theoretical guidance for the practical application of related anomaly detection systems.

Keywords

• metro vehicle
• surface defect
• anomaly detection
• 3d point cloud
• region segmentation
• plane filtering