The Railway Detection via Adaptive Multi-scale Fusion Processing

One of the main problems for safe autonomous driving vehicles that have not been solved completely is the high-precision and timely lane detection. In this work, we present a novel operator for railway detection to settle these tasks based on lane detection for the first time, called adaptive multi-scale fusion Sobel operators. The new operators can eliminate the noises generated by the environment in the railway image and derive more integrated edge feature information from the 0°, 45°, 90°, and 135° detection via 4 matrixes of 3 * 3 operators for permutation and summation. The image processing for railway detection includes the preprocess for images, railway edge detection, and track line polynomial fitting. Our experiment has validated that this improved detection method has realized the high accuracy and efficiency for rail detection. The dynamic rail detection and identification in the video of the railway track prove that this method has a significant effect on the left and right curved railway detection. It has good robustness and applicability.