Discretization–Filtering–Reconstruction: Railway Detection in Images for Navigation of Inspection UAV

Abstract

High-precision railway detection constitutes one of the primary challenges in railway inspection systems using unmanned aerial vehicle (UAV) imagery. In this article, a discretization-filtering-reconstruction (DFR) method is proposed for high-precision railway detection in the wild, by leveraging the representation capability of deep networks and the intrinsic characteristics of railway lines. As a preprocessing step, the DFR method segments railway images through a lightweight convolutional neural network, where a split-recursion-merge (SRM) module is designed to enhance the railway features among different rows or columns. Then, the segmentation mask is discretized into a series of trapezoidal connected components, which are formulated as nodes to construct a directed graph. The graph nodes are grouped into different clusters, and background clusters are filtered according to their shape confidence. Finally, the connected components of filtered clusters are reconstructed into a polynomial curve (the railway line).With the discretization, filtering, and reconstruction procedures of the DFR method, the problems of false segmentation, background inference and occlusions are greatly alleviated, producing continuous and smooth railway lines. Experiments on the dataset released in this article show that the DFR method significantly improves the accuracy of railway detection. Based on the DFR method and a UAV control strategy, a real-time visual UAV navigation system is built for railway inspection. The dataset and code for DFR are available at github.com/ksws0499733/DFR.