Research on the Algorithm for Height Measurement of Sand Pipe Based on Deep Learning

Abstract

During the daily operation of locomotive, affected by the locomotive itself and external environmental factors, the height of the sand pipe may change, which will lead to wheel slippage failure in severe cases, resulting in safety problems. Therefore, a sand pipe height measurement algorithm based on deep learning is proposed in this paper. Firstly, the object detection of sand pipe in 2D image and the fitting segmentation of rail plane were realized through the deep-learning-based YOLOv3-tiny algorithm and DeepLabv3+ model. Secondly, for the noise point cloud in the measurement environment, this paper implements point cloud filtering through depth value filtering and clustering segmentation. For the foreign matters on the sand pipe that affect the height measurement, the edge extraction and false edge removal are designed to eliminate the influence of foreign matters on the measurement accuracy. Finally, the rail plane is fitted by RANSAC algorithm to achieve 3D high-precision measurement of sand pipe height. Using the proposed algorithm and a 10-division vernier caliper, a comparative experiment is conducted on the height measurement of sand pipes with different positions, shapes and light intensities. The result shows that using this algorithm, the height measurement error of sand pipes can be basically controlled within ± 1 mm, with an average measurement accuracy of 98.74%, indicating that the algorithm has a certain degree of robustness.

Keywords

• deep learning
• height measurement
• object detection
• point cloud filtering
• 3d high-precision measurement