Turnout Feature Recognition Based on One-Dimensional Deep Learning Method

This paper addresses the challenges in railway switch detection by proposing a novel approach based on one-dimensional sequence modeling of point cloud cross-sections. Traditional methods, including 2D vision and 3D point cloud processing, face limitations in efficiency, robustness, and fine-grained recognition. The proposed method transforms 3D point cloud data into 1D sequences, capturing spatial variations and enabling efficient deep learning processing. Various 1D deep learning models, such as TextCNN, BiRNN, and ResNet-1D, are evaluated for switch component classification, aiming to improve accuracy and adaptability in complex environments. The approach is designed to overcome the computational and generalization challenges of existing methods, particularly in handling complex and noisy point cloud data from switch areas. By leveraging the sequential nature of the data, the method enhances the detection and classification of critical switch components.